EVA-RSI:通过视频评估器的递归自我改进来评判时序质量
摘要
EVA-RSI 提出了一种面向视频评估器的多智能体递归自我改进框架,采用通过构造定义的代理测试(时间反转、丢帧)和跨维度纳什议价,以 Qwen3-VVL-8B 为起点提升运动质量判断能力,在动态性基准测试中取得最高分,并可作为视频生成的有效强化学习奖励。
查看缓存全文
缓存时间: 2026/11/20 09:10
# Recursive Self-Improvement of Video Evaluatorsfor Judging Temporal Quality
Source: [https://arxiv.org/html/2611.23890](https://arxiv.org/html/2611.23890)
## EVA\-RSI: Recursive Self\-Improvement of Video Evaluators for Judging Temporal Quality
Yu Yuan††thanks:Work done during an internship at Roblox\.Ruizhou DingAffiliation:RobloxShaoheng FangAffiliation:RobloxAffiliation:University of Texas at AustinStanley H\. ChanAffiliation:Purdue UniversityZhuo ChenAffiliation:Roblox
###### Abstract
Video evaluators play an important role in video generation, yet they miss motion defects, limiting the quality of generated videos\. Recursive self\-improvement \(RSI\) has made strong progress in mathematics and coding, where explicit verifiers provide reliable feedback on a single goal\. Video evaluation poses two difficulties for RSI: it has no verifier, and it must judge many dimensions at once, such as physics, smoothness and dynamics, where a gain on one can cost another\. We introduceEVA\-RSI, a multi\-agent framework for exploring RSI under these two constraints\. We build a fixed proxy exam with construction\-defined targets by applying controlled changes to videos, such as temporal reversal and frame dropping\. These edits supply supervision without human annotation\. A small set of human\-labeled real videos anchors the exam to human judgment\. Each dimension has an LLM agent that proposes improvements\. A leader agent coordinates the proposals, and a Nash bargaining rule selects updates that balance progress across dimensions\. Humans design the exam, review candidate changes and revise the curriculum between stages\. Starting from Qwen3\-VL\-8B, EVA\-RSI improves motion evaluation, with the largest gain on smoothness, and the gains carry over to defect types never seen during the run\. On the dynamics subsets of public benchmarks, our evaluator achieves the highest observed scores among 21 tested models\. As an RL reward, it increases generated motion and earns human preference\. A data\-selection pipeline incorporating it also increases motion\.
## 1Introduction
Video evaluation shapes what video generators learn\. Generators learn from evaluators, used as RL rewards\[[6](https://arxiv.org/html/2611.23890#bib.bib6),[31](https://arxiv.org/html/2611.23890#bib.bib7)\]and as filters for training data\[[7](https://arxiv.org/html/2611.23890#bib.bib19),[41](https://arxiv.org/html/2611.23890#bib.bib20)\], making motion judgment central to generator training\. For example, VideoReward\[[32](https://arxiv.org/html/2611.23890#bib.bib2)\], when used as the sole RL reward for Wan2\.1\[[49](https://arxiv.org/html/2611.23890#bib.bib1)\], reduces video motion after 50 updates \(, bottom\)\. General vision\-language models also miss motion errors: Qwen3\-VL\-8B\[[2](https://arxiv.org/html/2611.23890#bib.bib5)\]assigns the same score to the original clip and an edited copy in which the person suddenly freezes \(, top\)\. An evaluator that misjudges motion passes its blind spot on to the generator\.
Human labels for motion are scarce, so this skill is hard to teach directly\. Recursive self\-improvement \(RSI\) offers another route: an agent system improves a model round after round, and each round starts from the best model so far\[[72](https://arxiv.org/html/2611.23890#bib.bib21),[11](https://arxiv.org/html/2611.23890#bib.bib73)\]\. In mathematics and code, models improve themselves over many rounds\[[69](https://arxiv.org/html/2611.23890#bib.bib9),[16](https://arxiv.org/html/2611.23890#bib.bib11),[43](https://arxiv.org/html/2611.23890#bib.bib25),[70](https://arxiv.org/html/2611.23890#bib.bib23),[72](https://arxiv.org/html/2611.23890#bib.bib21)\], because a unit test or a proof checker verifies each answer\. Video evaluation poses two difficulties for RSI\. First, it has no such verifier\. Second, it has no single goal: an evaluator must judge physics, smoothness, dynamics and more at once, and a change that helps one dimension can hurt another\. We ask:*can RSI improve a video evaluator without a native verifier and across many dimensions?*
EVA\-RSIaddresses both \([Fig\.2](https://arxiv.org/html/2611.23890#S3.F2)\)\. We build a fixed proxy exam with construction\-defined targets\. We reverse a clean clip, drop frames, or make an object disappear, and assign the original clip as the pairwise target\. Human\-labeled real videos connect this synthetic supervision to perceptual judgments\. Each dimension has a proposer agent that studies the evaluator’s errors and proposes fixes\. A leader agent coordinates the proposals, and engineer agents train new versions of the evaluator\. An acceptance rule based on Nash bargaining\[[42](https://arxiv.org/html/2611.23890#bib.bib13)\]unifies the dimensions: it accepts a version only if the main dimensions improve, or hold once they saturate, and no dimension falls below its floor\. Each accepted version is a*champion*\. Each round builds on the current champion, by training from it or merging it with other candidates, and its errors guide the agents’ next experiments\. Between stages, humans design a better exam where the current one is too coarse, so the system combines agent search with human curriculum design\. The evaluator increases motion as an RL reward and within a data\-selection pipeline; human raters prefer the RL results at update 50 \([Sec\.5](https://arxiv.org/html/2611.23890#S5)\)\.
We study EVA\-RSI as an exploratory system: can a constructed exam guide useful search, can the loop improve motion judgments while retaining other skills, and do these gains benefit generation? Our contributions are:
- •Multi\-agent RSI for video evaluation\.Two designs address the missing verifier and competing objectives: a constructed\-target exam with human curriculum design, and dimension\-specific agents coordinated by a leader and Nash\-based acceptance \([Sec\.3](https://arxiv.org/html/2611.23890#S3)\)\.
- •A stronger video evaluator\.The loop improves every main dimension over Qwen3\-VL\-8B, preserves the guarded skills, and transfers to unseen defects and public benchmarks \([Sec\.4](https://arxiv.org/html/2611.23890#S4)\)\. Training controls and champion histories reveal how improvements accumulate and where trade\-offs arise\. The resulting evaluator increases generated motion as an RL reward and within a data\-selection pipeline, with human preference gains at update 50 \([Sec\.5](https://arxiv.org/html/2611.23890#S5)\)\.
## 2Related Work
Video evaluators\.Benchmarks such as VBench\[[22](https://arxiv.org/html/2611.23890#bib.bib4)\], EvalCrafter\[[34](https://arxiv.org/html/2611.23890#bib.bib35)\]and T2V\-CompBench\[[46](https://arxiv.org/html/2611.23890#bib.bib36)\]use a separate tool for each quality dimension\. For example, they use optical flow to measure dynamic degree\. Evaluators trained on human scores or preferences include VideoScore\[[18](https://arxiv.org/html/2611.23890#bib.bib15)\], VideoPhy\[[4](https://arxiv.org/html/2611.23890#bib.bib37),[5](https://arxiv.org/html/2611.23890#bib.bib14)\], VisionReward\[[64](https://arxiv.org/html/2611.23890#bib.bib38)\], LiFT\[[55](https://arxiv.org/html/2611.23890#bib.bib57)\], T2VQA\[[24](https://arxiv.org/html/2611.23890#bib.bib56)\], UnifiedReward\[[56](https://arxiv.org/html/2611.23890#bib.bib39)\], VideoScore2\[[17](https://arxiv.org/html/2611.23890#bib.bib45)\]and VideoReward\[[32](https://arxiv.org/html/2611.23890#bib.bib2)\]\. VideoReward gives scores for visual quality \(VQ\), motion quality \(MQ\) and text alignment \(TA\)\. Vision\-language models are also zero\-shot or trained judges\[[74](https://arxiv.org/html/2611.23890#bib.bib40),[2](https://arxiv.org/html/2611.23890#bib.bib5),[63](https://arxiv.org/html/2611.23890#bib.bib71)\]\. Self\-Taught Evaluators\[[52](https://arxiv.org/html/2611.23890#bib.bib72)\]train a text judge over iterations on synthetic pairs with known answers; we construct proxy targets for video motion and let agents search the whole training recipe\. Video understanding benchmarks\[[28](https://arxiv.org/html/2611.23890#bib.bib16),[35](https://arxiv.org/html/2611.23890#bib.bib17),[14](https://arxiv.org/html/2611.23890#bib.bib41),[8](https://arxiv.org/html/2611.23890#bib.bib68),[71](https://arxiv.org/html/2611.23890#bib.bib69),[10](https://arxiv.org/html/2611.23890#bib.bib70)\]show that these models miss small temporal details, and motion and physics benchmarks\[[51](https://arxiv.org/html/2611.23890#bib.bib64),[30](https://arxiv.org/html/2611.23890#bib.bib65),[39](https://arxiv.org/html/2611.23890#bib.bib66),[40](https://arxiv.org/html/2611.23890#bib.bib67)\]find the same for generated videos\. Quality assessment models learn from human opinion scores\[[58](https://arxiv.org/html/2611.23890#bib.bib51),[59](https://arxiv.org/html/2611.23890#bib.bib52),[61](https://arxiv.org/html/2611.23890#bib.bib53)\]and from synthetic distortions\[[33](https://arxiv.org/html/2611.23890#bib.bib42),[29](https://arxiv.org/html/2611.23890#bib.bib43),[60](https://arxiv.org/html/2611.23890#bib.bib54),[66](https://arxiv.org/html/2611.23890#bib.bib55)\]\. Video\-language models also learn temporal order from reversed or shuffled negative examples\[[1](https://arxiv.org/html/2611.23890#bib.bib44)\]\. In all of these evaluators, people choose the training recipe\. In our work, agents propose the recipes, and edits and human labels decide which recipe survives\.
Recursive self\-improvement\.Language models improve from their own outputs with self\-training\[[69](https://arxiv.org/html/2611.23890#bib.bib9),[21](https://arxiv.org/html/2611.23890#bib.bib32),[15](https://arxiv.org/html/2611.23890#bib.bib34)\], self\-feedback\[[38](https://arxiv.org/html/2611.23890#bib.bib33),[67](https://arxiv.org/html/2611.23890#bib.bib10)\]and RL with verifiable rewards\[[25](https://arxiv.org/html/2611.23890#bib.bib12),[16](https://arxiv.org/html/2611.23890#bib.bib11)\]\. LLM agents also improve programs and agents\. They optimize prompts\[[65](https://arxiv.org/html/2611.23890#bib.bib31)\], evolve code\[[45](https://arxiv.org/html/2611.23890#bib.bib26),[43](https://arxiv.org/html/2611.23890#bib.bib25),[26](https://arxiv.org/html/2611.23890#bib.bib24)\], design or rewrite agents\[[68](https://arxiv.org/html/2611.23890#bib.bib28),[20](https://arxiv.org/html/2611.23890#bib.bib27),[70](https://arxiv.org/html/2611.23890#bib.bib23),[75](https://arxiv.org/html/2611.23890#bib.bib22)\]and conduct machine learning research\[[36](https://arxiv.org/html/2611.23890#bib.bib29),[9](https://arxiv.org/html/2611.23890#bib.bib30)\], including autonomous post\-training\[[11](https://arxiv.org/html/2611.23890#bib.bib73)\]and data engineering\[[12](https://arxiv.org/html/2611.23890#bib.bib74)\]\. Flex\-RSI\[[72](https://arxiv.org/html/2611.23890#bib.bib21)\]tests such loops on vision and robotics tasks, and it keeps a change only if the score on held\-out data increases\. These loops need a task with a reliable score, such as a unit test, a proof checker or a held\-out metric\. Self\-rewarding models\[[67](https://arxiv.org/html/2611.23890#bib.bib10)\]judge their own outputs\. The judge and model then evolve together without a fixed reference\. We apply RSI to a video evaluator, where the score itself is not reliable, and keep the verifier outside the loop\.
## 3Method
Figure 2:The EVA\-RSI system\.① A team of LLM agents proposes, implements and trains candidate evaluators from the current champion\. ② A frozen verifier scores each candidate on an exam with fixed targets, and a Nash bargaining rule accepts or rejects it\. An accepted candidate becomes the champion of the next round; every outcome enters a shared memory that the proposers read\. Humans design the verifier and redesign the exam between stages\.EVA\-RSI improves a video evaluator in rounds \([Fig\.2](https://arxiv.org/html/2611.23890#S3.F2)\)\. In each round, agents propose and train new candidate evaluators \([Sec\.3\.2](https://arxiv.org/html/2611.23890#S3.SS2)\)\. A frozen checker scores each candidate on questions with constructed answers \([Sec\.3\.1](https://arxiv.org/html/2611.23890#S3.SS1)\)\. Then an acceptance rule, frozen within each stage, decides whether the candidate replaces the current*champion*\([Sec\.3\.3](https://arxiv.org/html/2611.23890#S3.SS3)\)\. Each new champion anchors training in the next round\.
Three choices shape the design\. First, the verifier comes before the agents: targets and scoring are fixed within each stage, making candidate scores comparable\. Second, agents control the evaluator’s data, recipe, architecture and inference; the verifier remains frozen\. Third, a candidate wins only through balanced progress, so no step can buy a gain on one dimension with a large loss on another\. Humans design the curriculum and review candidate changes \([Sec\.3\.4](https://arxiv.org/html/2611.23890#S3.SS4)\)\.
### 3\.1Constructed answers
The checker uses fixed construction\-defined targets and human labels\. The questions cover nine dimensions in four groups\.
Injected defects\(physics, consistency, smoothness, quality\)\. We pair each clean clip with a copy containing one defect\. For physics, the clip plays backward, or an object freezes, disappears or is duplicated\. For consistency, we swap the background, insert a scene cut or change the colors suddenly\. For smoothness, we freeze, drop, shake or flicker frames\. For quality, we blur, add noise or compress the clip heavily\. Each defect also has subtle versions, for example a freeze of 2–3 frames\. The clean clip is the assigned pairwise target\. We ask the question as a pair \(“which video is better?”\) and as two separate 1–5 ratings with construction\-assigned targets\. These targets encode the edit intervention; human labels assess perceived quality on real videos\.
Motion from optical flow\(dynamics\)\. We pair clips that have clearly different optical flow, and assign the higher\-flow clip as the dynamics target\.
Real generated videos\(real; stage 2\)\. We pair generated videos that have clearly different human scores in VideoPhy\-2\[[5](https://arxiv.org/html/2611.23890#bib.bib14)\]and VideoFeedback\[[18](https://arxiv.org/html/2611.23890#bib.bib15)\]\. These questions check that gains on injected defects transfer to real generated videos\.
Guards against forgetting\(semantics, human preference, general; general in stage 2\)\. These questions use caption matching, expert preference pairs from VideoGen\-RewardBench\[[32](https://arxiv.org/html/2611.23890#bib.bib2)\], and video question answering from MVBench\[[28](https://arxiv.org/html/2611.23890#bib.bib16)\]and TempCompass\[[35](https://arxiv.org/html/2611.23890#bib.bib17)\]\.
Scores combine pairwise correctness \(ties:0\.50\.5\), normalized rating agreement1−\|𝔼\[r\]−t\|/41\-\|\\mathbb\{E\}\[r\]\-t\|/4with targettt, and multiple\-choice accuracy\. We average within families, then dimensions; the main mean averages five dimensions\.
Two development banks, each with about 2\.9k questions, drive the loop\. A hidden test bank uses other clips and prompts and adds defect types that are not in the development banks\. It shares no questions, source clips or prompts with the development banks, and we use it to score champions, not to accept candidates\. Training uses a separate split from the exam and public evaluation data\.
### 3\.2A team of agents
Coding agents based on language models run the loop \([Fig\.2](https://arxiv.org/html/2611.23890#S3.F2)\) in four roles: proposers, a leader, engineers and a reviewer\. Humans design the fifth part, the verifier \([Sec\.3\.4](https://arxiv.org/html/2611.23890#S3.SS4)\)\.
Proposers\.Each dimension has a proposer, which speaks for the player of that dimension in the bargaining game \([Sec\.3\.3](https://arxiv.org/html/2611.23890#S3.SS3)\): five for the main dimensions and four guardians for semantics, quality, human preference and general understanding \(four and three in stage 1\)\. Each proposer reviews the champion’s scores, its weakest question families and past candidates\. It proposes up to two experiments and flags risks in other proposals\.
Leader\.The leader picks, merges or changes proposals, one for each free GPU slot, to maximize the expected Nash product\. Each round covers multiple dimensions, including a bold idea and one protecting the guardian closest to its floor\. The leader specifies each experiment\.
Engineers\.An engineer turns one specification into code and a training recipe from the current champion, with at most 1\.5 hours of training on four GPUs\. If the job fails, it reads the log and repairs the candidate once\.
Shared memory\.A ledger records every candidate: its idea, the proposing role, its development score on each dimension, the decision and the reason for a rejection\. Proposers and the leader read it every round, so they combine ideas that worked and drop ideas that failed\. This memory, together with training from the champion, makes the search recursive: each round starts from what earlier rounds found\.
A fixed program trains and scores each candidate and applies the acceptance rule\.
Reviewer\.The reviewer reads the code changes of each candidate that the rule accepts and can veto it for rule violations, such as training on bank questions or benchmark data\. A candidate becomes champion only if it passes both the rule and the review\.
### 3\.3Nash bargaining acceptance
We treat each dimensioniias a player with utilityui=si−diu\_\{i\}=s\_\{i\}\-d\_\{i\}\. Here,sis\_\{i\}is the score of the candidate\. The disagreement pointdid\_\{i\}is the score of the start model minus a small tolerance\. A candidate replaces the champion when it meets three conditions\. First, each dimension stays above its floor \(ui\>0u\_\{i\}\>0\)\. Second, the main dimensions improve on average, with no large drop on any of them; once they saturate, they stay within0\.0050\.005\([Sec\.A\.3](https://arxiv.org/html/2611.23890#A1.SS3)\)\. Third, the weighted log Nash product increases:
∑iwilog\(si−di\)\>∑iwilog\(sichamp−di\)\.\\textstyle\\sum\_\{i\}w\_\{i\}\\log\(s\_\{i\}\-d\_\{i\}\)\>\\sum\_\{i\}w\_\{i\}\\log\(s\_\{i\}^\{\\text\{champ\}\}\-d\_\{i\}\)\.\(1\)The main dimensions are physics, consistency, smoothness, dynamics and real, and each has weight 1\. The guard dimensions have smaller weights\. The rule suits competing dimensions for two reasons\. First, it penalizes proximity to a dimension’s floor, becauselogui\\log u\_\{i\}falls quickly asui→0u\_\{i\}\\to 0\. Second, the Nash comparison is invariant to positive rescaling of each utility\. Floors and drop limits separately specify score tolerances\. Family guards additionally bound score drops within each defect family \([Sec\.A\.3](https://arxiv.org/html/2611.23890#A1.SS3)\)\.
### 3\.4Human supervision and search controls
SafeguardTargeted failureTwo banks, 64 question types, both orders, ratingsa shortcut on one bank or formatFloors and family guardstrading one skill for the meanHidden test bankoverfitting the development banks12 public benchmarksoverfitting the exam as a wholeFrozen scorer and rule, diff filterediting the judge, training on banksReviewer agent, human inspectionviolations the program misses
Table 1:Controls on evaluator search\.Humans construct the exam, score the start model on each family, and replace saturated or noisy families\. During a run, they review decisions and reasons for rejection\. Between stages, they revise the exam\. When a stage\-1 candidate raised smoothness by giving up the time\-reversal family, they added guards for single families; when the stage\-1 gains did not reach real generated videos, they added real\-video questions \([Sec\.4\.3](https://arxiv.org/html/2611.23890#S4.SS3)\)\. At the end, they judge the generated videos in a blind study \([Sec\.5](https://arxiv.org/html/2611.23890#S5)\)\.
The search uses four layers of controls \([Tab\.1](https://arxiv.org/html/2611.23890#S3.T1)\)\.A large and diverse exam: 64 question types over nine dimensions, in two development banks from disjoint sources\. Each pair is asked in both orders and as separate ratings to measure order sensitivity and single\-video judgments\. Floors and family guards constrain score drops\.A hidden test bankwith other clips, prompts and defect types scores only champions and never decides acceptance\. Agents have no access to its questions\. One stage\-2 prompt names a weakness observed on the stage\-1 test bank\. Excluding the 36 questions on that weakness raises the score from0\.9060\.906to0\.9160\.916\.Public benchmarks: 12 benchmarks assess transfer beyond the exam\. MVBench supplies separate videos for the general guard and public evaluation \([AppendixD](https://arxiv.org/html/2611.23890#A4)\)\.Review of the verifier boundary: the scorer and the rule run from a frozen copy that agents cannot edit, the program rejects any code change that refers to development or test assets, the reviewer agent can veto every accepted change, and humans inspect the recipe of every champion\.
#### Two stages\.
The loop ran in two stages\. Stage 1 uses injected defects, motion from optical flow and the semantics, quality and human\-preference guards\. Stage 2 starts from the last stage\-1 champion\. It adds the real dimension, the general guard, more subtle defect versions and family guards \([AppendixA](https://arxiv.org/html/2611.23890#A1)\)\. The champions carry over, and the rule keeps its form, with guards for the new families\. Humans designed this second exam after seeing where the first was too coarse; this curriculum step is the human part of the system\. Public\-benchmark gains concentrate in this second stage \([Sec\.4\.3](https://arxiv.org/html/2611.23890#S4.SS3)\)\.
## 4Evaluating the EVA\-RSI design
### 4\.1Setup
We start from Qwen3\-VL\-8B\-Instruct\[[2](https://arxiv.org/html/2611.23890#bib.bib5)\]and run EVA\-RSI in two stages, with 170 candidates over 88 rounds\. Stage 2 continues from the last stage\-1 champion with the larger question bank \([Sec\.3](https://arxiv.org/html/2611.23890#S3)\)\. Each candidate trains for at most 1\.5 hours\. The agents run on Claude Opus 5\.5, with fallbacks when it is unavailable \([AppendixA](https://arxiv.org/html/2611.23890#A1)\)\. The loop used 944 B200 GPU hours\. We score all eight champions on a hidden test bank of 3,450 questions\.
### 4\.2The exam guides evaluator improvement
Figure 3:RSI improves the evaluator over rounds\.\(a\) Normalized score of each champion on the hidden test bank\. \(b\) Scored candidates \(grey\) and champions \(blue\) on the development banks, which change at the start of stage 2\. \(c\) Mean rank score on 12 public benchmarks, against all evaluators except VideoReward \([AppendixD](https://arxiv.org/html/2611.23890#A4)\)\. The development banks determine the final champion\.Figure 4:Test\-bank pairs before and after RSI\.B is the correct answer in each pair\. P is the probability that each model gives to B\. Left: close\-up of four consecutive frames; dashed lines mark a fixed position\. Middle: objects suddenly change size\. Right: people and boxes vanish and reappear\.DimensionQwen3\-VL\-8BAfter stage 1EVA\-RSIGainPhysics0\.7390\.8810\.858\+0\.119\+0\.119Consistency0\.8770\.9770\.997\+0\.120\+0\.120Smoothness0\.4610\.8410\.888\+0\.427\+0\.427Dynamics0\.7690\.9040\.923\+0\.154\+0\.154Real videos0\.6840\.7550\.863\+0\.179\+0\.179Semantics0\.9880\.9920\.992\+0\.004\+0\.004Quality0\.9680\.9930\.996\+0\.028\+0\.028Human pref\.0\.6240\.6400\.660\+0\.036\+0\.036General0\.6850\.7260\.726\+0\.041\+0\.041
Table 2:Normalized scores on the hidden test bank\.Top: main dimensions\. Bottom: guard dimensions\. Chance is0\.50\.5for pairwise questions\.[Figure3](https://arxiv.org/html/2611.23890#S4.F3)a shows the test score of each champion\. The mean over the main dimensions rises from0\.7060\.706to0\.9060\.906\.
The largest gain is where Qwen3\-VL\-8B was weakest \([Tab\.2](https://arxiv.org/html/2611.23890#S4.T2)\): smoothness rises from0\.4610\.461to0\.8880\.888\. At the same 54\-frame input length, Qwen3\-VL\-8B stays near chance\. Dynamics and physics also rise, so the evaluator judges both how much a video moves and whether its motion is plausible; physics peaks in stage 1 and gives back part of its gain in stage 2\. Scores also rise on real videos labeled by human raters\. The gains hold for single\-video ratings, and pairwise answers depend less on video order\. No guard dimension drops from the starting model on the test bank\.[Figure4](https://arxiv.org/html/2611.23890#S4.F4)shows test pairs before and after RSI\.
The gains reach unseen defects\.The test bank holds ten defect families of physics, consistency and smoothness that never appear in the development banks or the training data, such as teleporting objects, texture swaps and local frame shuffles\. On these families, the score rises from0\.7540\.754to0\.9130\.913\(gain:0\.1590\.159, 95% interval\[0\.116,0\.200\]\[0\.116,0\.200\]\), with the largest gain on smoothness\. The gain also holds without the questions that the stage\-1 test bank shares \([Sec\.A\.4](https://arxiv.org/html/2611.23890#A1.SS4)\)\.
StageChampionRoleRoundRecipe in one sentenceTest gain1c007dynamics5A weight average \(“soup”\) of three earlier candidates, which keeps their gains with less drift in human preference\.\+0\.068\+0\.0681c009physics5A weight soup plus a short LoRA polish with graded time\-reversal labels and new physics and consistency defect families; proposed in the same round as c007 and judged against it after c007 was accepted\.\+0\.142\+0\.1421c022physics14A direction\-aware vision tower: a time\-antisymmetric change to the patch embedding, trained on motion and direction labels, language model frozen\.\+0\.153\+0\.1531c029semantics17Composes c022’s vision tower with the language model of c021, at a higher resolution\.\+0\.158\+0\.1581c046human26The best 32\-frame vision tower plus a language\-model LoRA, with a repair step for human preference and semantics\.\+0\.166\+0\.1662c001human1Fine\-tuning on real generated videos with human labels, with base\-model anchors for text alignment and replay of earlier defect families\.\+0\.190\+0\.1902c033general16A polish that adds position\-debiased base\-model answers on multiple\-choice questions, which keeps general understanding\.\+0\.196\+0\.1962c071 \(EVA\-RSI\)physics30An arrow\-of\-time teacher \(is this clip played forward or backward?\), built from about 2,400 forward and reversed training clips at 54 frames and230,400230\{,\}400pixels, with a dynamics re\-fit\.\+0\.200\+0\.200Table 3:The eight champions found by RSI\.Role: the agent that proposed the candidate\. Test gain: main mean minus the start model on the hidden test bank\.
### 4\.3Dynamics on public benchmarks
ModelVideoScore\-BenchVBenchEVA\-RSI \(ours\)0\.6110\.906After stage 1 \(ours\)0\.5640\.781Qwen3\-VL\-8B \(start\)0\.5380\.812Qwen3\-VL\-32B0\.4950\.875Qwen3\.8\-27B \(thinking\)0\.4530\.750InternVL3\.5\-8B\[[53](https://arxiv.org/html/2611.23890#bib.bib46)\]0\.4330\.750VideoScore2\[[17](https://arxiv.org/html/2611.23890#bib.bib45)\]0\.5340\.406UnifiedReward\-7B\[[56](https://arxiv.org/html/2611.23890#bib.bib39)\]0\.4860\.500VideoReward−\-0\.3110\.312Claude Opus 5\.50\.5490\.781GPT\-5\.6 Sol0\.5750\.844Muse Spark 1\.30\.5120\.844Grok 4\.60\.5490\.781Gemini 4 Argon0\.4800\.719Table 4:Video dynamics on public benchmarks\.Spearman correlation with human dynamics scores \(VideoScore\-Bench\) and preference accuracy on dynamics pairs \(VBench\)\. 80 videos and 16 pairs from 6 prompts\. Bold: highest observed score among 21 models \([Tab\.S7](https://arxiv.org/html/2611.23890#A4.T7)\)\.[Table4](https://arxiv.org/html/2611.23890#S4.T4)compares EVA\-RSI with our stage\-1 champion and 19 open models, trained video evaluators and proprietary models on the dynamics parts of two public benchmarks\. In VideoScore\-Bench\[[18](https://arxiv.org/html/2611.23890#bib.bib15)\], models rate the dynamic degree of generated videos, and we report the Spearman correlation with human ratings\. In VBench\[[22](https://arxiv.org/html/2611.23890#bib.bib4)\], models pick the more dynamic of two generated videos, as human annotators did\. All models use their default inference settings \([AppendixD](https://arxiv.org/html/2611.23890#A4)\)\.
EVA\-RSI leads on the evaluated dynamics subsets\.It has the highest point estimate among 21 models on both metrics \([Tab\.4](https://arxiv.org/html/2611.23890#S4.T4)\)\. All three stage\-2 champions exceed the proprietary models on these two metrics\.[TableS7](https://arxiv.org/html/2611.23890#A4.T7)in[AppendixD](https://arxiv.org/html/2611.23890#A4)gives the full comparison on 12 public benchmarks, where proprietary models lead on the overall mean and EVA\-RSI leads all open models\.[Figure3](https://arxiv.org/html/2611.23890#S4.F3)c scores each champion on all 15 evaluation metrics of these benchmarks; the public score rises from0\.390\.39to0\.710\.71, almost all in stage 2\.
Stage 2 improves public\-benchmark performance\.Without real\-video questions, stage 1 raises the hidden\-test score from0\.7060\.706to0\.8720\.872, with the public score remaining near its starting level \([Fig\.3](https://arxiv.org/html/2611.23890#S4.F3)a,c\)\. Stage 2 combines human\-labeled real videos with subtler defects and new guards; its first champion raises the public score\.
### 4\.4Coordinating improvement across dimensions
OutcomeStage 1Stage 2AllAccepted538Scored and rejected413879main gain too small313263dimension below floor10515a family guard–11Not scored344983training job failed232952implementation rejected11920driver restart066unfinished055Total8090170Table 5:Outcomes of the 170 candidates\.Each rejected candidate counts under the first rule that it failed\.Of the 170 candidates, 8 became champions \([Tabs\.3](https://arxiv.org/html/2611.23890#S4.T3)and[5](https://arxiv.org/html/2611.23890#S4.T5)\)\. The hidden\-test point estimate rises at every change of champion; three of eight steps remain significant after Holm correction \([Sec\.A\.2](https://arxiv.org/html/2611.23890#A1.SS2.SSS0.Px4)\)\.
In stage 1, the agents averaged the weights of earlier candidates, made the vision encoder sensitive to the direction of time, and added defect pairs in which both clips move at the same speed\. In stage 2, they fine\-tuned on real generated videos, used the base model’s answers as anchors, and trained on forward and reversed clips at the same 54\-frame length used for evaluation\. EVA\-RSI adds an arrow\-of\-time teacher to the previous champion \([AppendixB](https://arxiv.org/html/2611.23890#A2)\)\. The dynamics score rises most when the agents add graded time\-reversal labels in round 5\.
Improvement with skill retention\.One LoRA trained once on all the loop’s training data, including the stage\-2 pools \(Union SFT\), reaches0\.8890\.889on test and0\.5580\.558on human preference, versus0\.9060\.906and0\.6600\.660for EVA\-RSI; its public score is0\.530\.53against0\.710\.71\. Adding anchors satisfies the rule but reaches0\.8480\.848and0\.8440\.844across two seeds \([Sec\.A\.6](https://arxiv.org/html/2611.23890#A1.SS6)\)\. These controls expose the trade\-off addressed by coordinated search: fitting the pooled data improves motion scores, while retaining other skills requires further recipe choices\.
In a non\-recursive control, the same agents trained each candidate from Qwen3\-VL\-8B using the same data and acceptance rule\. They saw earlier proposals’ titles and roles, but no scores or outcomes\. One of ten proposals failed a budget check; the other nine missed human\-preference or general\-understanding floors\. EVA\-RSI outperformed the development\-selected candidate and all four tested candidates on the hidden test \([Tab\.S6](https://arxiv.org/html/2611.23890#A1.T6)\)\.
#### The final recipe uses inherited capabilities\.
Replaying the c033 recipe directly from Qwen3\-VL\-8B yields test means of0\.7570\.757and0\.7560\.756across two seeds; c033 trained along the champion chain reaches0\.9020\.902\([Tabs\.S6](https://arxiv.org/html/2611.23890#A1.T6)and[S3](https://arxiv.org/html/2611.23890#A1.T3)\)\. The difference is pronounced on smoothness:0\.5140\.514and0\.5050\.505from direct training versus0\.8500\.850along the chain\. The recipe thus uses accumulated vision and motion capabilities\. Compute differs: 944 GPU hours for EVA\-RSI, 384 for Union SFT, 352 per Union \+ anchors seed and 212 for the non\-recursive search, stopped after observed stagnation\. The controls assess the complete search design at these budgets\.
### 4\.5Lessons from running RSI
Exam coverage shapes the gains\.Real\-video scores stayed between0\.7000\.700and0\.7550\.755through stage 1; the first champion under the expanded stage\-2 exam raised the score to0\.8520\.852\.Returns diminish on a fixed exam\.All stage\-2 candidates scored after EVA\-RSI landed within0\.0040\.004of it, motivating a broader exam for future search \(more in[AppendixC](https://arxiv.org/html/2611.23890#A3)\)\.
## 5EVA\-RSI as a reward and a data filter
We use EVA\-RSI as an RL reward and as a filter for training data\. Both pipelines increase motion in generated videos; human raters prefer the RL results at update 50\.
RewardUpd\.Dyn\.↑\\uparrowFlow↑\\uparrowObj\.↑\\uparrowAdh\.↑\\uparrowElem\.↑\\uparrowNone \(start\)00\.57913\.771\.433\.6680\.745Qwen3\-VL\-8B, seed 1500\.79015\.641\.673\.7410\.762Qwen3\-VL\-8B, seed 2500\.74914\.971\.493\.6840\.751EVA\-RSI, seed 1500\.84817\.641\.763\.7300\.761EVA\-RSI, seed 2500\.89219\.812\.033\.7140\.757Qwen3\-VL\-8B, seed 11000\.77014\.931\.373\.7540\.761Qwen3\-VL\-8B, seed 21000\.70013\.481\.343\.6710\.748EVA\-RSI, seed 11000\.91320\.531\.883\.6750\.750EVA\-RSI, seed 21000\.89217\.591\.623\.5190\.712VideoReward \(fair\), seed 1500\.3517\.940\.823\.0410\.617VideoReward \(fair\), seed 2500\.2717\.330\.692\.9560\.604VideoReward500\.2896\.730\.702\.5750\.522VideoReward1200\.0954\.060\.521\.6900\.307*Trained to the full budget \(500 GPU hours\)*Qwen3\-VL\-8B3460\.77821\.082\.001\.8410\.265EVA\-RSI3420\.98726\.342\.793\.1140\.616
Table 6:RL with different rewards\(630 test videos\)\. Dyn\.: dynamic degree\. Obj\.: object motion\. Adh\. and Elem\.: judge scores\. VideoReward \(fair\): same penalties and settings as EVA\-RSI\. VideoReward: stress test without penalties\.Figure 5:Same prompt and seed; only the reward changes\(update 50\)\. EVA\-RSI keeps the scene and gives the most motion\. VideoReward: visual and motion quality, without penalties\. Corner: optical flow\.Questionvs\. Qwen3\-VL\-8Bvs\. VideoReward \(fair\)More motion0\.77\[0\.68,0\.85\]\[0\.68,0\.85\]0\.85\[0\.75,0\.94\]\[0\.75,0\.94\]Better overall0\.73\[0\.65,0\.80\]\[0\.65,0\.80\]0\.69\[0\.55,0\.82\]\[0\.55,0\.82\]Visual quality0\.71\[0\.63,0\.79\]\[0\.63,0\.79\]0\.66\[0\.51,0\.79\]\[0\.51,0\.79\]Prompt adherence0\.73\[0\.65,0\.80\]\[0\.65,0\.80\]0\.71\[0\.58,0\.82\]\[0\.58,0\.82\]Table 7:Human study at update 50 \(training seed 1\), share of answers for EVA\-RSI \(0\.5: no preference\)\. Six raters; 48 and 24 pairs\.### 5\.1RL reward
All runs fine\-tune Wan2\.1 T2V\-1\.3B\[[49](https://arxiv.org/html/2611.23890#bib.bib1)\]with DiffusionART\[[50](https://arxiv.org/html/2611.23890#bib.bib18)\], a group\-based policy\-gradient method; all runs share the settings that we chose in pilots with EVA\-RSI, and only the reward model changes\. The EVA\-RSI reward sums its 1–5 ratings for dynamics, physics, smoothness and quality, with fixed penalties for flicker and uneven global motion \([AppendixE](https://arxiv.org/html/2611.23890#A5)\)\. We compare it with the evaluator before RSI, Qwen3\-VL\-8B, which reads the same 54 frames, so the two rewards differ only in their weights\. We also run VideoReward\[[32](https://arxiv.org/html/2611.23890#bib.bib2)\]with the same penalties and settings \(fair\) and without penalties \(stress test\)\. We measure motion with VBench\[[22](https://arxiv.org/html/2611.23890#bib.bib4)\]*dynamic degree*\(RAFT\[[47](https://arxiv.org/html/2611.23890#bib.bib49)\]\),*optical flow*\[[13](https://arxiv.org/html/2611.23890#bib.bib75)\]and*object motion*, the flow that remains after we remove camera motion\. A blind Qwen3\.8\-27B judge rates prompt*adherence*\. We test on 630 videos from VBench and VidProM\[[54](https://arxiv.org/html/2611.23890#bib.bib8)\]prompts, and confirm each result on 300 MovieGenBench\[[44](https://arxiv.org/html/2611.23890#bib.bib50)\]prompts\.
#### More motion\.
All 32 motion tests pass after Holm correction: for both seeds, updates 50 and 100 and both prompt sets, EVA\-RSI gives more motion than the start model and than Qwen3\-VL\-8B \([Tab\.6](https://arxiv.org/html/2611.23890#S5.T6),[Sec\.E\.1](https://arxiv.org/html/2611.23890#A5.SS1)\)\. Object motion rises in every comparison after camera motion is removed\. The gain tracks the loop: the stage\-1 champion increases motion over Qwen3\-VL\-8B at update 50; EVA\-RSI sustains a significant lead at update 100 \([AppendixE](https://arxiv.org/html/2611.23890#A5)\)\. At update 50, adherence matches Qwen3\-VL\-8B; it falls with longer training \([AppendixE](https://arxiv.org/html/2611.23890#A5)\)\.
#### Better motion\.
Six raters compared pairs of videos from training seed 1 at update 50 with the same prompt and sampling seed\. We randomized their left–right order and counted ties as half a vote\. Raters judged EVA\-RSI videos to move more than Qwen3\-VL\-8B videos and to be better overall, in visual quality and in adherence \([Tab\.7](https://arxiv.org/html/2611.23890#S5.T7)\)\. When they see more motion in the EVA\-RSI video, they also prefer it overall in nine of ten answers\.
#### Motion quality beyond flow magnitude\.
A raw optical\-flow reward with the same settings moves more than EVA\-RSI at update 50, but its videos jitter more, and by update 100 EVA\-RSI leads it in both motion and adherence \([AppendixE](https://arxiv.org/html/2611.23890#A5)\)\.
#### VideoReward reduces motion under these settings\.
VideoReward, which is anti\-correlated with human dynamics ratings \([Tab\.4](https://arxiv.org/html/2611.23890#S4.T4)\), does the opposite\. Its matched\-setting run reduces dynamic degree to0\.3510\.351and0\.2710\.271from the start model’s0\.5790\.579\([Tab\.6](https://arxiv.org/html/2611.23890#S5.T6)\), and raters prefer EVA\-RSI over it on motion and overall\. Without penalties, it collapses to still close\-ups of faces \(and[5](https://arxiv.org/html/2611.23890#S5.F5)\)\.
### 5\.2Data filter
Picked−\-randomStep 1,500Step 3,000Dynamic degree↑\\uparrow\+0\.143∗\+0\.143^\{\*\}\+0\.277∗\+0\.277^\{\*\}Optical flow↑\\uparrow\+3\.47∗\+3\.47^\{\*\}\+7\.38∗\+7\.38^\{\*\}Adherence↑\\uparrow\+0\.10\+0\.10\+0\.03\+0\.03Elements↑\\uparrow\+0\.022\+0\.022−0\.001\-0\.001Dyn\. degree, picked / random0\.867 / 0\.7230\.873 / 0\.597Table 8:Selection pipeline incorporating EVA\-RSI versus random data\(300 prompts\)\.∗: 95% interval excludes zero\.We fine\-tune Wan2\.1 on 50k clips picked by a pipeline that ranks clips with EVA\-RSI after flow and caption filters, and on 50k random clips from a different pool, with the same recipe to evaluate the selection pipeline \([AppendixF](https://arxiv.org/html/2611.23890#A6)\)\. The flow threshold and caption filter select moving clips without scene cuts\. The two training sets differ in source composition and share about one fifth of their clips\. We evaluate both models on the same 300 prompts\. After 3,000 steps, the model trained on selected data produces motion in87%87\\%of its videos, compared with60%60\\%for the model trained on random data, with an adherence difference of\+0\.03\+0\.03whose 95% interval includes zero \([Tab\.8](https://arxiv.org/html/2611.23890#S5.T8)\)\.
### 5\.3Limitations and future work
The main limitation of EVA\-RSI is its verifier exam\. Controlled edits simplify motion failures and differ from the distribution of real generated videos, where defects interact with scene content and intent\. Construction\-defined targets describe individual edits\. Human\-labeled real videos broaden exam coverage; complex, interacting failures remain a target for expansion\. Future work should build richer exams from generator failures, validate targets with human judgments, and test transfer across generators\.
## 6Conclusion
EVA\-RSI explores how recursive improvement can extend to video evaluation through two designs: a constructed\-target exam supplies feedback, and dimension\-specific agents with Nash\-based acceptance coordinate competing objectives\. The resulting search improves temporal judgments while retaining guarded skills, transfers to unseen defects, and produces an evaluator that benefits video generation through RL rewards and data selection\. This study establishes the practical promise of evaluator self\-improvement and identifies richer, human\-validated exams as a central direction for further progress\.
## Acknowledgments
We thank Tinghui Zhou, Jiapeng Tang, Haiyue Yu, Charles Shang, and Kris Kitani for helpful discussions\.
## LLM Usage Statement
Generative AI tools \(Codex, Claude Code\) assisted with language editing and figure preparation\. The authors designed these protocols, audited their outputs, verified the reported results, and take full responsibility for the paper\.
## References
- \[1\]P\. Bagad, M\. Tapaswi, and C\. G\.M\. Snoek\(2023\)Test of Time: Instilling Video\-Language Models with a Sense of Time\.InCVPR,pp\. 2503–2516\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[2\]S\. Bai, Y\. Cai, R\. Chen, K\. Chen, X\. Chen, Z\. Cheng, L\. Deng, W\. Ding, C\. Gao, C\. Ge, W\. Ge, Z\. Guo, Q\. Huang, J\. Huang, F\. Huang, B\. Hui, S\. Jiang, Z\. Li, M\. Li, M\. Li, K\. Li, Z\. Lin, J\. Lin, X\. Liu, J\. Liu, C\. Liu, Y\. Liu, D\. Liu, S\. Liu, D\. Lu, R\. Luo, C\. Lv, R\. Men, L\. Meng, X\. Ren, X\. Ren, S\. Song, Y\. Sun, J\. Tang, J\. Tu, J\. Wan, P\. Wang, P\. Wang, Q\. Wang, Y\. Wang, T\. Xie, Y\. Xu, H\. Xu, J\. Xu, Z\. Yang, M\. Yang, J\. Yang, A\. Yang, B\. Yu, F\. Zhang, H\. Zhang, X\. Zhang, B\. Zheng, H\. Zhong, J\. Zhou, F\. Zhou, J\. Zhou, Y\. Zhu, and K\. Zhu\(2025\)Qwen3\-VL Technical Report\.arXiv preprint arXiv:2511\.21631\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px3.p2.1),[§1](https://arxiv.org/html/2611.23890#S1.p1.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[§4\.1](https://arxiv.org/html/2611.23890#S4.SS1.p1.1)\.
- \[3\]S\. Bai, K\. Chen, X\. Liu, J\. Wang, W\. Ge, S\. Song, K\. Dang, P\. Wang, S\. Wang, J\. Tang, H\. Zhong, Y\. Zhu, M\. Yang, Z\. Li, J\. Wan, P\. Wang, W\. Ding, Z\. Fu, Y\. Xu, J\. Ye, X\. Zhang, T\. Xie, Z\. Cheng, H\. Zhang, Z\. Yang, H\. Xu, and J\. Lin\(2025\)Qwen2\.5\-VL Technical Report\.arXiv preprint arXiv:2502\.13923\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px3.p2.1)\.
- \[4\]H\. Bansal, Z\. Lin, T\. Xie, Z\. Zong, M\. Yarom, Y\. Bitton, C\. Jiang, Y\. Sun, K\. Chang, and A\. Grover\(2025\)VideoPhy: Evaluating Physical Commonsense for Video Generation\.InICLR,pp\. 102075–102121\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[5\]H\. Bansal, C\. Peng, Y\. Bitton, R\. Goldenberg, A\. Grover, and K\. Chang\(2026\)VideoPhy\-2: A Challenging Action\-Centric Physical Commonsense Evaluation in Video Generation\.InICLR,pp\. 118456–118470\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[§3\.1](https://arxiv.org/html/2611.23890#S3.SS1.p4.1)\.
- \[6\]K\. Black, M\. Janner, Y\. Du, I\. Kostrikov, and S\. Levine\(2024\)Training Diffusion Models with Reinforcement Learning\.InICLR,pp\. 4965–4987\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p1.1)\.
- \[7\]A\. Blattmann, T\. Dockhorn, S\. Kulal, D\. Mendelevitch, M\. Kilian, D\. Lorenz, Y\. Levi, Z\. English, V\. Voleti, A\. Letts, V\. Jampani, and R\. Rombach\(2023\)Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets\.arXiv preprint arXiv:2311\.15127\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p1.1)\.
- \[8\]M\. Cai, R\. Tan, J\. Zhang, B\. Zou, K\. Zhang, F\. Yao, F\. Zhu, J\. Gu, Y\. Zhong, Y\. Shang, Y\. Dou, J\. Park, J\. Gao, Y\. J\. Lee, and J\. Yang\(2024\)TemporalBench: Benchmarking Fine\-grained Temporal Understanding for Multimodal Video Models\.arXiv preprint arXiv:2410\.10818\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[9\]J\. S\. Chan, N\. Chowdhury, O\. Jaffe, J\. Aung, D\. Sherburn, E\. Mays, G\. Starace, K\. Liu, L\. Maksin, T\. Patwardhan, A\. Madry, and L\. Weng\(2025\)MLE\-bench: Evaluating Machine Learning Agents on Machine Learning Engineering\.InICLR,pp\. 50466–50494\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[10\]D\. Cores, M\. Dorkenwald, M\. Mucientes, C\. G\. M\. Snoek, and Y\. M\. Asano\(2025\)Lost in Time: A New Temporal Benchmark for VideoLLMs\.InBMVC,Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[11\]Y\. Du, X\. Yang, Z\. Zhou, Y\. Ge, C\. Wang, J\. Wang, S\. Chen, Z\. Liu, Y\. Zhang, W\. Gu, J\. Zhang, Z\. Lei, and S\. Chen\(2026\)RSI\-Master: Structuring Experiments to Guide Autonomous Model Improvement\.arXiv preprint arXiv:2609\.35561\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p2.1),[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[12\]Y\. Du, X\. Yang, Z\. Zhou, W\. Liu, Z\. Lei, Z\. Chen, F\. Liu, H\. Wu, Y\. Cai, Z\. Liu, X\. Zhu, W\. Wang, L\. Zhang, C\. Qian, and S\. Chen\(2026\)DataMaster: Data\-Centric Autonomous AI Research\.arXiv preprint arXiv:2605\.10906\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[13\]G\. Farnebäck\(2003\)Two\-frame motion estimation based on polynomial expansion\.InProceedings of the 13th Scandinavian Conference on Image Analysis,Lecture Notes in Computer Science, Vol\.2749,pp\. 363–370\.Cited by:[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[14\]C\. Fu, Y\. Dai, Y\. Luo, L\. Li, S\. Ren, R\. Zhang, Z\. Wang, C\. Zhou, Y\. Shen, M\. Zhang, P\. Chen, Y\. Li, S\. Lin, S\. Zhao, K\. Li, T\. Xu, X\. Zheng, E\. Chen, C\. Shan, R\. He, and X\. Sun\(2025\)Video\-MME: The First\-Ever Comprehensive Evaluation Benchmark of Multi\-modal LLMs in Video Analysis\.InCVPR,pp\. 24108–24118\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[15\]C\. Gulcehre, T\. L\. Paine, S\. Srinivasan, K\. Konyushkova, L\. Weerts, A\. Sharma, A\. Siddhant, A\. Ahern, M\. Wang, C\. Gu, W\. Macherey, A\. Doucet, O\. Firat, and N\. de Freitas\(2023\)Reinforced Self\-Training \(ReST\) for Language Modeling\.arXiv preprint arXiv:2308\.08998\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[16\]D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi, X\. Zhang, X\. Yu, Y\. Wu, Z\. F\. Wu, Z\. Gou, Z\. Shao, Z\. Li, Z\. Gao, A\. Liu, B\. Xue, B\. Wang, B\. Wu, B\. Feng, C\. Lu, C\. Zhao, C\. Deng, C\. Ruan, D\. Dai, D\. Chen, D\. Ji, E\. Li, F\. Lin, F\. Dai, F\. Luo, G\. Hao, G\. Chen, G\. Li, H\. Zhang, H\. Xu, H\. Ding, H\. Gao, H\. Qu, H\. Li, J\. Guo, J\. Li, J\. Chen, J\. Yuan, J\. Tu, J\. Qiu, J\. Li, J\. L\. Cai, J\. Ni, J\. Liang, J\. Chen, K\. Dong, K\. Hu, K\. You, K\. Gao, K\. Guan, K\. Huang, K\. Yu, L\. Wang, L\. Zhang, L\. Zhao, L\. Wang, L\. Zhang, L\. Xu, L\. Xia, M\. Zhang, M\. Zhang, M\. Tang, M\. Zhou, M\. Li, M\. Wang, M\. Li, N\. Tian, P\. Huang, P\. Zhang, Q\. Wang, Q\. Chen, Q\. Du, R\. Ge, R\. Zhang, R\. Pan, R\. Wang, R\. J\. Chen, R\. L\. Jin, R\. Chen, S\. Lu, S\. Zhou, S\. Chen, S\. Ye, S\. Wang, S\. Yu, S\. Zhou, S\. Pan, S\. S\. Li, S\. Zhou, S\. Wu, T\. Yun, T\. Pei, T\. Sun, T\. Wang, W\. Zeng, W\. Liu, W\. Liang, W\. Gao, W\. Yu, W\. Zhang, W\. L\. Xiao, W\. An, X\. Liu, X\. Wang, X\. Chen, X\. Nie, X\. Cheng, X\. Liu, X\. Xie, X\. Liu, X\. Yang, X\. Li, X\. Su, X\. Lin, X\. Q\. Li, X\. Jin, X\. Shen, X\. Chen, X\. Sun, X\. Wang, X\. Song, X\. Zhou, X\. Wang, X\. Shan, Y\. K\. Li, Y\. Q\. Wang, Y\. X\. Wei, Y\. Zhang, Y\. Xu, Y\. Li, Y\. Zhao, Y\. Sun, Y\. Wang, Y\. Yu, Y\. Zhang, Y\. Shi, Y\. Xiong, Y\. He, Y\. Piao, Y\. Wang, Y\. Tan, Y\. Ma, Y\. Liu, Y\. Guo, Y\. Ou, Y\. Wang, Y\. Gong, Y\. Zou, Y\. He, Y\. Xiong, Y\. Luo, Y\. You, Y\. Liu, Y\. Zhou, Y\. X\. Zhu, Y\. Huang, Y\. Li, Y\. Zheng, Y\. Zhu, Y\. Ma, Y\. Tang, Y\. Zha, Y\. Yan, Z\. Z\. Ren, Z\. Ren, Z\. Sha, Z\. Fu, Z\. Xu, Z\. Xie, Z\. Zhang, Z\. Hao, Z\. Ma, Z\. Yan, Z\. Wu, Z\. Gu, Z\. Zhu, Z\. Liu, Z\. Li, Z\. Xie, Z\. Song, Z\. Pan, Z\. Huang, Z\. Xu, Z\. Zhang, and Z\. Zhang\(2025\)DeepSeek\-R1 incentivizes reasoning in LLMs through reinforcement learning\.Nature645\(8081\),pp\. 633–638\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p2.1),[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[17\]X\. He, D\. Jiang, P\. Nie, M\. Liu, Z\. Jiang, M\. Su, W\. Ma, J\. Lin, C\. Ye, Y\. Lu, K\. Wu, B\. Schneider, Q\. D\. Do, Z\. Li, Y\. Jia, Y\. Zhang, G\. Cheng, H\. Wang, W\. Zhou, Q\. Lin, Y\. Zhang, G\. Zhang, W\. Huang, and W\. Chen\(2026\)VideoScore2: Think Before You Score In Generated Video Evaluation\.Transactions on Machine Learning Research\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1),[Table S7](https://arxiv.org/html/2611.23890#A4.T7.3.1.9.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[Table 4](https://arxiv.org/html/2611.23890#S4.T4.3.8.1.1)\.
- \[18\]X\. He, D\. Jiang, G\. Zhang, M\. Ku, A\. Soni, S\. Siu, H\. Chen, A\. Chandra, Z\. Jiang, A\. Arulraj, K\. Wang, Q\. D\. Do, Y\. Ni, B\. Lyu, Y\. Narsupalli, R\. Fan, Z\. Lyu, B\. Y\. Lin, and W\. Chen\(2024\)VideoScore: Building Automatic Metrics to Simulate Fine\-grained Human Feedback for Video Generation\.InEMNLP,pp\. 2105–2123\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[§3\.1](https://arxiv.org/html/2611.23890#S3.SS1.p4.1),[§4\.3](https://arxiv.org/html/2611.23890#S4.SS3.p1.1)\.
- \[19\]V\. Hosu, F\. Hahn, M\. Jenadeleh, H\. Lin, H\. Men, T\. Sziranyi, S\. Li, and D\. Saupe\(2017\)The Konstanz natural video database \(KoNViD\-1k\)\.InQoMEX,pp\. 1–6\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1)\.
- \[20\]S\. Hu, C\. Lu, and J\. Clune\(2025\)Automated Design of Agentic Systems\.InICLR,pp\. 21344–21377\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[21\]J\. Huang, S\. Gu, L\. Hou, Y\. Wu, X\. Wang, H\. Yu, and J\. Han\(2023\)Large Language Models Can Self\-Improve\.InEMNLP,pp\. 1051–1068\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[22\]Z\. Huang, Y\. He, J\. Yu, F\. Zhang, C\. Si, Y\. Jiang, Y\. Zhang, T\. Wu, Q\. Jin, N\. Chanpaisit, Y\. Wang, X\. Chen, L\. Wang, D\. Lin, Y\. Qiao, and Z\. Liu\(2024\)VBench: Comprehensive Benchmark Suite for Video Generative Models\.InCVPR,pp\. 21807–21818\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[§4\.3](https://arxiv.org/html/2611.23890#S4.SS3.p1.1),[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[23\]D\. Jiang, M\. Ku, T\. Li, Y\. Ni, S\. Sun, R\. Fan, and W\. Chen\(2024\)GenAI Arena: An Open Evaluation Platform for Generative Models\.InNeurIPS,pp\. 79889–79908\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1)\.
- \[24\]T\. Kou, X\. Liu, Z\. Zhang, C\. Li, H\. Wu, X\. Min, G\. Zhai, and N\. Liu\(2024\)Subjective\-Aligned Dataset and Metric for Text\-to\-Video Quality Assessment\.InACM Multimedia,pp\. 7793–7802\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[25\]N\. Lambert, J\. Morrison, V\. Pyatkin, S\. Huang, H\. Ivison, F\. Brahman, L\. J\. Miranda, A\. Liu, N\. Dziri, X\. Lyu, Y\. Gu, S\. Malik, V\. Graf, J\. Hwang, J\. Yang, R\. Le Bras, O\. Tafjord, C\. Wilhelm, L\. Soldaini, N\. Smith, Y\. Wang, P\. Dasigi, and H\. Hajishirzi\(2025\)Tulu 3: Pushing Frontiers in Open Language Model Post\-Training\.InCOLM,Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[26\]R\. Lange, Y\. Imajuku, and E\. Cetin\(2026\)ShinkaEvolve: Towards Open\-Ended and Sample\-Efficient Program Evolution\.InICLR,pp\. 74026–74078\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[27\]B\. Li, Y\. Zhang, D\. Guo, R\. Zhang, F\. Li, H\. Zhang, K\. Zhang, P\. Zhang, Y\. Li, Z\. Liu, and C\. Li\(2025\)LLaVA\-OneVision: Easy Visual Task Transfer\.Transactions on Machine Learning Research\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px3.p2.1)\.
- \[28\]K\. Li, Y\. Wang, Y\. He, Y\. Li, Y\. Wang, Y\. Liu, Z\. Wang, J\. Xu, G\. Chen, P\. Luo, L\. Wang, and Y\. Qiao\(2024\)MVBench: A Comprehensive Multi\-modal Video Understanding Benchmark\.InCVPR,pp\. 22195–22206\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[§3\.1](https://arxiv.org/html/2611.23890#S3.SS1.p5.1)\.
- \[29\]H\. Lin, V\. Hosu, and D\. Saupe\(2019\)KADID\-10k: A Large\-scale Artificially Distorted IQA Database\.InQoMEX,pp\. 1–3\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[30\]X\. Ling, C\. Zhu, M\. Wu, H\. Li, X\. Feng, C\. Yang, A\. Hao, J\. Zhu, J\. Wu, and X\. Chu\(2025\)VMBench: A Benchmark for Perception\-Aligned Video Motion Generation\.InICCV,pp\. 13087–13098\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[31\]J\. Liu, G\. Liu, J\. Liang, Y\. Li, J\. Liu, X\. Wang, P\. Wan, D\. ZHANG, and W\. Ouyang\(2025\)Flow\-GRPO: Training Flow Matching Models via Online RL\.InNeurIPS,pp\. 45650–45685\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p1.1)\.
- \[32\]J\. Liu, G\. Liu, J\. Liang, Z\. Yuan, X\. Liu, M\. Zheng, X\. Wu, Q\. Wang, M\. Xia, X\. Wang, X\. Liu, F\. Yang, P\. Wan, D\. ZHANG, K\. Gai, Y\. Yang, and W\. Ouyang\(2025\)Improving Video Generation with Human Feedback\.InNeurIPS,pp\. 91326–91363\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p1.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[§3\.1](https://arxiv.org/html/2611.23890#S3.SS1.p5.1),[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[33\]X\. Liu, J\. Van De Weijer, and A\. D\. Bagdanov\(2017\)RankIQA: Learning from Rankings for No\-Reference Image Quality Assessment\.InICCV,pp\. 1040–1049\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[34\]Y\. Liu, X\. Cun, X\. Liu, X\. Wang, Y\. Zhang, H\. Chen, Y\. Liu, T\. Zeng, R\. Chan, and Y\. Shan\(2024\)EvalCrafter: Benchmarking and Evaluating Large Video Generation Models\.InCVPR,pp\. 22139–22149\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[35\]Y\. Liu, S\. Li, Y\. Liu, Y\. Wang, S\. Ren, L\. Li, S\. Chen, X\. Sun, and L\. Hou\(2024\)TempCompass: Do Video LLMs Really Understand Videos?\.InFindings of ACL,pp\. 8731–8772\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[§3\.1](https://arxiv.org/html/2611.23890#S3.SS1.p5.1)\.
- \[36\]C\. Lu, C\. Lu, R\. T\. Lange, J\. Foerster, J\. Clune, and D\. Ha\(2024\)The AI Scientist: Towards Fully Automated Open\-Ended Scientific Discovery\.arXiv preprint arXiv:2408\.06292\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[37\]Y\. Ma, X\. Wu, K\. Sun, and H\. Li\(2025\)HPSv3: Towards Wide\-Spectrum Human Preference Score\.InICCV,pp\. 15086–15095\.Cited by:[Appendix G](https://arxiv.org/html/2611.23890#A7.SS0.SSS0.Px3.p1.1)\.
- \[38\]A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao, S\. Wiegreffe, U\. Alon, N\. Dziri, S\. Prabhumoye, Y\. Yang, S\. Gupta, B\. P\. Majumder, K\. Hermann, S\. Welleck, A\. Yazdanbakhsh, and P\. Clark\(2023\)Self\-Refine: Iterative Refinement with Self\-Feedback\.InNeurIPS,pp\. 46534–46594\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[39\]F\. Meng, J\. Liao, X\. Tan, Q\. Lu, W\. Shao, K\. Zhang, Y\. Cheng, D\. Li, and P\. Luo\(2025\)Towards World Simulator: Crafting Physical Commonsense\-Based Benchmark for Video Generation\.InICML,pp\. 43781–43806\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[40\]S\. Motamed, L\. Culp, K\. Swersky, P\. Jaini, and R\. Geirhos\(2026\)Do Generative Video Models Understand Physical Principles?\.InWACV,pp\. 948–958\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[41\]K\. Nan, R\. Xie, P\. Zhou, T\. Fan, Z\. Yang, Z\. Chen, X\. Li, J\. Yang, and Y\. Tai\(2025\)OpenVid\-1M: A Large\-Scale High\-Quality Dataset for Text\-to\-video Generation\.InICLR,pp\. 1045–1064\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p1.1)\.
- \[42\]J\. F\. Nash\(1950\)The Bargaining Problem\.Econometrica18\(2\),pp\. 155–162\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p3.1)\.
- \[43\]A\. Novikov, N\. Vũ, M\. Eisenberger, E\. Dupont, P\. Huang, A\. Z\. Wagner, S\. Shirobokov, B\. Kozlovskii, F\. J\. R\. Ruiz, A\. Mehrabian, M\. P\. Kumar, A\. See, S\. Chaudhuri, G\. Holland, A\. Davies, S\. Nowozin, P\. Kohli, and M\. Balog\(2025\)AlphaEvolve: A coding agent for scientific and algorithmic discovery\.arXiv preprint arXiv:2506\.13131\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p2.1),[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[44\]A\. Polyak, A\. Zohar, A\. Brown, A\. Tjandra, A\. Sinha, A\. Lee, A\. Vyas, B\. Shi, C\. Ma, C\. Chuang, D\. Yan, D\. Choudhary, D\. Wang, G\. Sethi, G\. Pang, H\. Ma, I\. Misra, J\. Hou, J\. Wang, K\. Jagadeesh, K\. Li, L\. Zhang, M\. Singh, M\. Williamson, M\. Le, M\. Yu, M\. K\. Singh, P\. Zhang, P\. Vajda, Q\. Duval, R\. Girdhar, R\. Sumbaly, S\. S\. Rambhatla, S\. Tsai, S\. Azadi, S\. Datta, S\. Chen, S\. Bell, S\. Ramaswamy, S\. Sheynin, S\. Bhattacharya, S\. Motwani, T\. Xu, T\. Li, T\. Hou, W\. Hsu, X\. Yin, X\. Dai, Y\. Taigman, Y\. Luo, Y\. Liu, Y\. Wu, Y\. Zhao, Y\. Kirstain, Z\. He, Z\. He, A\. Pumarola, A\. Thabet, A\. Sanakoyeu, A\. Mallya, B\. Guo, B\. Araya, B\. Kerr, C\. Wood, C\. Liu, C\. Peng, D\. Vengertsev, E\. Schonfeld, E\. Blanchard, F\. Juefei\-Xu, F\. Nord, J\. Liang, J\. Hoffman, J\. Kohler, K\. Fire, K\. Sivakumar, L\. Chen, L\. Yu, L\. Gao, M\. Georgopoulos, R\. Moritz, S\. K\. Sampson, S\. Li, S\. Parmeggiani, S\. Fine, T\. Fowler, V\. Petrovic, and Y\. Du\(2024\)Movie Gen: A Cast of Media Foundation Models\.arXiv preprint arXiv:2410\.13720\.Cited by:[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[45\]B\. Romera\-Paredes, M\. Barekatain, A\. Novikov, M\. Balog, M\. P\. Kumar, E\. Dupont, F\. J\. R\. Ruiz, J\. S\. Ellenberg, P\. Wang, O\. Fawzi, P\. Kohli, and A\. Fawzi\(2024\)Mathematical discoveries from program search with large language models\.Nature625\(7995\),pp\. 468–475\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[46\]K\. Sun, K\. Huang, X\. Liu, Y\. Wu, Z\. Xu, Z\. Li, and X\. Liu\(2025\)T2V\-CompBench: A Comprehensive Benchmark for Compositional Text\-to\-video Generation\.InCVPR,pp\. 8406–8416\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[47\]Z\. Teed and J\. Deng\(2020\)RAFT: Recurrent All\-Pairs Field Transforms for Optical Flow\.InECCV,pp\. 402–419\.Cited by:[Appendix E](https://arxiv.org/html/2611.23890#A5.SS0.SSS0.Px4.p1.1),[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[48\]H\. Tong, Z\. Wang, Z\. Chen, H\. Ji, S\. Qiu, S\. Han, K\. Geng, Z\. Xue, Y\. Zhou, P\. Xia, M\. Ding, R\. Rafailov, C\. Finn, and H\. Yao\(2025\)MJ\-Video: Benchmarking and Rewarding Video Generation with Fine\-Grained Video Preference\.InNeurIPS,pp\. 79142–79173\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1)\.
- \[49\]Wan Team, A\. Wang, B\. Ai, B\. Wen, C\. Mao, C\. Xie, D\. Chen, F\. Yu, H\. Zhao, J\. Yang, J\. Zeng, J\. Wang, J\. Zhang, J\. Zhou, J\. Wang, J\. Chen, K\. Zhu, K\. Zhao, K\. Yan, L\. Huang, M\. Feng, N\. Zhang, P\. Li, P\. Wu, R\. Chu, R\. Feng, S\. Zhang, S\. Sun, T\. Fang, T\. Wang, T\. Gui, T\. Weng, T\. Shen, W\. Lin, W\. Wang, W\. Wang, W\. Zhou, W\. Wang, W\. Shen, W\. Yu, X\. Shi, X\. Huang, X\. Xu, Y\. Kou, Y\. Lv, Y\. Li, Y\. Liu, Y\. Wang, Y\. Zhang, Y\. Huang, Y\. Li, Y\. Wu, Y\. Liu, Y\. Pan, Y\. Zheng, Y\. Hong, Y\. Shi, Y\. Feng, Z\. Jiang, Z\. Han, Z\. Wu, and Z\. Liu\(2025\)Wan: Open and Advanced Large\-Scale Video Generative Models\.arXiv preprint arXiv:2503\.20314\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p1.1),[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[50\]Z\. Wan, J\. Chen, Z\. Zhang, P\. Zhou, W\. Zhao, F\. Kong, P\. Wan, K\. Gai, and Y\. You\(2026\)DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models\.arXiv preprint arXiv:2610\.12572\.Cited by:[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[51\]H\. Wang, W\. Zhu, L\. Miao, Y\. Xu, F\. Gao, Q\. Tian, and Y\. Wang\(2025\)Aligning Human Motion Generation with Human Perceptions\.InICLR,pp\. 77626–77656\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[52\]T\. Wang, I\. Kulikov, O\. Golovneva, P\. Yu, W\. Yuan, J\. Dwivedi\-Yu, R\. Y\. Pang, M\. Fazel\-Zarandi, J\. Weston, and X\. Li\(2024\)Self\-Taught Evaluators\.arXiv preprint arXiv:2408\.02666\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[53\]W\. Wang, Z\. Gao, L\. Gu, H\. Pu, L\. Cui, X\. Wei, Z\. Liu, L\. Jing, S\. Ye, J\. Shao, Z\. Wang, Z\. Chen, H\. Zhang, G\. Yang, H\. Wang, Q\. Wei, J\. Yin, W\. Li, E\. Cui, G\. Chen, Z\. Ding, C\. Tian, Z\. Wu, J\. Xie, Z\. Li, B\. Yang, Y\. Duan, X\. Wang, Z\. Hou, H\. Hao, T\. Zhang, S\. Li, X\. Zhao, H\. Duan, N\. Deng, B\. Fu, Y\. He, Y\. Wang, C\. He, B\. Shi, J\. He, Y\. Xiong, H\. Lv, L\. Wu, W\. Shao, K\. Zhang, H\. Deng, B\. Qi, J\. Ge, Q\. Guo, W\. Zhang, S\. Zhang, M\. Cao, J\. Lin, K\. Tang, J\. Gao, H\. Huang, Y\. Gu, C\. Lyu, H\. Tang, R\. Wang, H\. Lv, W\. Ouyang, L\. Wang, M\. Dou, X\. Zhu, T\. Lu, D\. Lin, J\. Dai, W\. Su, B\. Zhou, K\. Chen, Y\. Qiao, W\. Wang, and G\. Luo\(2025\)InternVL3\.5: Advancing Open\-Source Multimodal Models in Versatility, Reasoning, and Efficiency\.arXiv preprint arXiv:2508\.18265\.Cited by:[Table S7](https://arxiv.org/html/2611.23890#A4.T7.3.1.8.1),[Table 4](https://arxiv.org/html/2611.23890#S4.T4.3.7.1.1)\.
- \[54\]W\. Wang and Y\. Yang\(2024\)VidProM: A Million\-scale Real Prompt\-Gallery Dataset for Text\-to\-Video Diffusion Models\.InNeurIPS,pp\. 65618–65642\.Cited by:[§5\.1](https://arxiv.org/html/2611.23890#S5.SS1.p1.1)\.
- \[55\]Y\. Wang, Z\. Tan, J\. Wang, X\. Yang, C\. Jin, and H\. Li\(2024\)LiFT: Leveraging Human Feedback for Text\-to\-Video Model Alignment\.arXiv preprint arXiv:2412\.04814\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[56\]Y\. Wang, Y\. Zang, H\. Li, C\. Jin, and J\. Wang\(2025\)Unified Reward Model for Multimodal Understanding and Generation\.arXiv preprint arXiv:2503\.05236\.Cited by:[Table S7](https://arxiv.org/html/2611.23890#A4.T7.3.1.10.1),[§2](https://arxiv.org/html/2611.23890#S2.p1.1),[Table 4](https://arxiv.org/html/2611.23890#S4.T4.3.9.1.1)\.
- \[57\]Y\. Wang, Y\. Wang, D\. Zhao, C\. Xie, and Z\. Zheng\(2024\)VideoHallucer: Evaluating Intrinsic and Extrinsic Hallucinations in Large Video\-Language Models\.arXiv preprint arXiv:2406\.16338\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1)\.
- \[58\]H\. Wu, C\. Chen, J\. Hou, L\. Liao, A\. Wang, W\. Sun, Q\. Yan, and W\. Lin\(2022\)FAST\-VQA: Efficient End\-to\-End Video Quality Assessment with Fragment Sampling\.InECCV,pp\. 538–554\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[59\]H\. Wu, E\. Zhang, L\. Liao, C\. Chen, J\. Hou, A\. Wang, W\. Sun, Q\. Yan, and W\. Lin\(2023\)Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives\.InICCV,pp\. 20087–20097\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[60\]H\. Wu, Z\. Zhang, E\. Zhang, C\. Chen, L\. Liao, A\. Wang, K\. Xu, C\. Li, J\. Hou, G\. Zhai, G\. Xue, W\. Sun, Q\. Yan, and W\. Lin\(2024\)Q\-Instruct: Improving Low\-level Visual Abilities for Multi\-modality Foundation Models\.InCVPR,pp\. 25490–25500\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[61\]H\. Wu, Z\. Zhang, W\. Zhang, C\. Chen, L\. Liao, C\. Li, Y\. Gao, A\. Wang, E\. Zhang, W\. Sun, Q\. Yan, X\. Min, G\. Zhai, and W\. Lin\(2024\)Q\-Align: Teaching LMMs for Visual Scoring via Discrete Text\-Defined Levels\.InICML,pp\. 54015–54029\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[62\]X\. Xiang, X\. Liu, Z\. Li, Z\. Liu, and J\. Zhang\(2026\)AIGVE\-Tool: AI\-Generated Video Evaluation Toolkit with Multifaceted Benchmark\.InVISAPP,pp\. 346–357\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1)\.
- \[63\]T\. Xiong, X\. Wang, D\. Guo, Q\. Ye, H\. Fan, Q\. Gu, H\. Huang, and C\. Li\(2025\)LLaVA\-Critic: Learning to Evaluate Multimodal Models\.InCVPR,pp\. 13618–13628\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[64\]J\. Xu, Y\. Huang, J\. Cheng, Y\. Yang, J\. Xu, Y\. Wang, W\. Duan, S\. Yang, Q\. Jin, S\. Li, J\. Teng, Z\. Yang, W\. Zheng, X\. Liu, D\. Zhang, M\. Ding, X\. Zhang, S\. Huang, X\. Gu, M\. Huang, J\. Tang, and Y\. Dong\(2026\)VisionReward: Fine\-Grained Multi\-Dimensional Human Preference Learning for Image and Video Generation\.InAAAI,pp\. 11269–11277\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[65\]C\. Yang, X\. Wang, Y\. Lu, H\. Liu, Q\. V\. Le, D\. Zhou, and X\. Chen\(2024\)Large Language Models as Optimizers\.InICLR,pp\. 12028–12068\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[66\]Z\. You, Z\. Li, J\. Gu, Z\. Yin, T\. Xue, and C\. Dong\(2024\)Depicting Beyond Scores: Advancing Image Quality Assessment Through Multi\-modal Language Models\.InECCV,pp\. 259–276\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[67\]W\. Yuan, R\. Y\. Pang, K\. Cho, X\. Li, S\. Sukhbaatar, J\. Xu, and J\. E\. Weston\(2024\)Self\-Rewarding Language Models\.InICML,pp\. 57905–57923\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[68\]E\. Zelikman, E\. Lorch, L\. Mackey, and A\. T\. Kalai\(2024\)Self\-Taught Optimizer \(STOP\): Recursively Self\-Improving Code Generation\.InCOLM,Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[69\]E\. Zelikman, Y\. Wu, J\. Mu, and N\. Goodman\(2022\)STaR: Bootstrapping Reasoning With Reasoning\.InNeurIPS,pp\. 15476–15488\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p2.1),[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[70\]J\. Zhang, S\. Hu, C\. Lu, R\. Lange, and J\. Clune\(2026\)Darwin Gödel Machine: Open\-Ended Evolution of Self\-Improving Agents\.InICLR,Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p2.1),[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[71\]J\. Zhang, M\. Cai, and Y\. J\. Lee\(2024\)Vinoground: Scrutinizing LMMs over Dense Temporal Reasoning with Short Videos\.arXiv preprint arXiv:2410\.02763\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[72\]J\. Zhang, J\. Wang, Z\. Yu, Y\. Cai, W\. Chow, C\. Choy, Z\. Dou, H\. Hua, J\. Li, Z\. Liu, C\. Wang, S\. Wei, S\. Yang, Y\. Yuan, X\. Zhan, Y\. Zhang, S\. Zheng, X\. Li, L\. Liu, W\. Matusik, Z\. Tu, and W\. Wang\(2026\)Flex\-RSI: Benchmarking Recursive Self\-Improvement of Frontier Agents\.arXiv preprint arXiv:2611\.08946\.Cited by:[§1](https://arxiv.org/html/2611.23890#S1.p2.1),[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
- \[73\]Z\. Zhang, Z\. Jia, H\. Wu, C\. Li, Z\. Chen, Y\. Zhou, W\. Sun, X\. Liu, X\. Min, W\. Lin, and G\. Zhai\(2025\)Q\-Bench\-Video: Benchmark the Video Quality Understanding of LMMs\.InCVPR,pp\. 3229–3239\.Cited by:[Appendix D](https://arxiv.org/html/2611.23890#A4.SS0.SSS0.Px1.p1.1)\.
- \[74\]L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. Xing, H\. Zhang, J\. Gonzalez, and I\. Stoica\(2023\)Judging LLM\-as\-a\-Judge with MT\-Bench and Chatbot Arena\.InNeurIPS,pp\. 46595–46623\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p1.1)\.
- \[75\]T\. Zheng, X\. Wu, Z\. Zhang, Z\. He, C\. Zhang, B\. Coleman, R\. Wei, D\. Bai, H\. Liu, R\. Liu, X\. Wang, Y\. Zhuan, W\. Kang, R\. Xiang, H\. Huang, X\. Cheng, and Y\. Guo\(2026\)Dream\-RSI: Recursive Self\-Improvement through Evolving Worlds\.arXiv preprint arXiv:2609\.14858\.Cited by:[§2](https://arxiv.org/html/2611.23890#S2.p2.1)\.
Appendix
## Appendix ARSI details
### A\.1The evaluator
The evaluator is a vision\-language model that receives one or two videos and a question and answers with one token:AorBfor a pairwise question, a rating from1to5, or an option letter\. We read each answer from the next\-token probabilities, so a rating is the expected value over1…5\. The loop can change its weights, in practice with LoRA fine\-tuning, and its inference settings, such as the system prompt and the number and size of frames\.
### A\.2The two stages
EVA\-RSI ran in two stages \([Tab\.S1](https://arxiv.org/html/2611.23890#A1.T1)\)\. Stage 2 starts from the last stage\-1 champion and uses a larger question bank\. Its last champion is EVA\-RSI, which we use in all downstream experiments\.
StageStartRoundsCandidatesChampions1Qwen3\-VL\-8B498052stage\-1 champion39903Total881708Table S1:The two stages of EVA\-RSI\.#### Stage 1\.
Stage 1 started from Qwen3\-VL\-8B\-Instruct and ran for about 39 hours on 16 B200 GPUs\. Stage 1 raised real\-video test score from0\.6840\.684to0\.7550\.755\. The base model already scored near ceiling on several development families\.
#### Stage 2\.
Stage 2 added the real\-video and general dimensions, subtler defects, harder motion and semantic pairs, and guards for single families \([Sec\.A\.3](https://arxiv.org/html/2611.23890#A1.SS3)\)\. It ran for about 47 hours on 16 and then 32 B200 GPUs\. EVA\-RSI reads 54 frames per video\. Its last recipe teaches it to tell forward from reversed clips, while keeping the other dimensions close to the previous champion\.
#### A third stage\.
We also ran a third stage from EVA\-RSI with the same verifier and a larger test bank\. None of its 100 candidates met the rule\. Midway, we added a looser rule, and two candidates passed it; the better one is level with EVA\-RSI on the hidden test bank, so we keep EVA\-RSI\. When the questions are nearly solved, more rounds on them no longer help\.
#### Step sizes\.
On the stage\-2 hidden test bank, the main mean rises at every change of champion\.[TableS2](https://arxiv.org/html/2611.23890#A1.T2)tests each of the eight steps with a paired bootstrap\. Three steps stay significant after a Holm correction: the first two champions \(c007 and c009\) and the first stage\-2 champion \(c001\)\. Physics rises significantly at the first step \(\+0\.118\+0\.118, interval\[\+0\.085,\+0\.151\]\[\+0\.085,\+0\.151\]\)\. The step from c022 to c029 also raises the resolution from 147k to 230k pixels per frame\. Stage\-1 champions did not see the real\-video and general dimensions during acceptance\.
StepGain in main mean95% intervalHolmStart→\\toc007\+0\.068\+0\.068\[\+0\.054,\+0\.082\]\[\+0\.054,\+0\.082\]✓c007→\\toc009\+0\.074\+0\.074\[\+0\.056,\+0\.091\]\[\+0\.056,\+0\.091\]✓c009→\\toc022\+0\.011\+0\.011\[\+0\.001,\+0\.022\]\[\+0\.001,\+0\.022\]c022→\\toc029\+0\.005\+0\.005\[−0\.008,\+0\.019\]\[\-0\.008,\+0\.019\]c029→\\toc046\+0\.007\+0\.007\[−0\.002,\+0\.017\]\[\-0\.002,\+0\.017\]c046→\\toc001\+0\.024\+0\.024\[\+0\.012,\+0\.037\]\[\+0\.012,\+0\.037\]✓c001→\\toc033\+0\.006\+0\.006\[−0\.002,\+0\.014\]\[\-0\.002,\+0\.014\]c033→\\toc071\+0\.004\+0\.004\[−0\.008,\+0\.016\]\[\-0\.008,\+0\.016\]Table S2:Steps between champions\(stage\-2 hidden test bank\)\. Holm: significant after correction over the eight steps\.Stage 1Stage 2DimensionBasec007c009c022c029c046c001c033c071 \(EVA\-RSI\)Proposed in round–5514172611630Frames32†3232323232323254Pixels per frame \(k\)147†147147147230230230230230Physics0\.7390\.8570\.8740\.8680\.8930\.8810\.8860\.8680\.858Consistency0\.8770\.9430\.9600\.9680\.9770\.9770\.9950\.9980\.997Smoothness0\.4610\.5690\.7700\.8050\.7940\.8410\.8440\.8500\.888Dynamics0\.7690\.8020\.8960\.9030\.9060\.9040\.9020\.9340\.923Real videos0\.6840\.7000\.7420\.7530\.7520\.7550\.8520\.8600\.863Semantics0\.9880\.9880\.9920\.9960\.9960\.9920\.9880\.9880\.992Quality0\.9680\.9770\.9940\.9950\.9920\.9930\.9940\.9970\.996Human pref\.0\.6240\.6220\.6210\.6230\.6310\.6400\.6560\.6560\.660General0\.6850\.7040\.7230\.7300\.7150\.7260\.7110\.7190\.726Main mean0\.7060\.7740\.8480\.8600\.8650\.8720\.8960\.9020\.906Table S3:All eight champions on the hidden test bank\.Main: mean of the five main dimensions\.†Default settings\.
#### Test bank\.
The stage\-2 prompts name one weak point seen on the stage\-1 test bank, reversal of slow clips \(without these 36 test questions the final main mean is0\.9160\.916, not0\.9060\.906\), and the two stages’ test banks share1,2541\{,\}254questions outside the real and general dimensions\.
### A\.3Full acceptance rule
Each dimensioniihas utilityui=si−diu\_\{i\}=s\_\{i\}\-d\_\{i\}\. Heredid\_\{i\}is the score of the untouched Qwen3\-VL\-8B\-Instruct minus a tolerance of0\.020\.02\(0\.0150\.015for human preference\)\. A candidate replaces the champion only if all of the following hold:
1. 1\.every dimension stays above its floor,ui\>0u\_\{i\}\>0;
2. 2\.the mean of the main dimensions rises by at least0\.010\.01, with a 90% lower bound above 0 under a paired cluster bootstrap;
3. 3\.no main dimension drops by more than0\.030\.03;
4. 4\.the weighted log Nash product rises \([Eq\.1](https://arxiv.org/html/2611.23890#S3.E1)\)\.
The main dimensions are physics, consistency, smoothness, dynamics and real videos, with weight 1\. Semantics, quality and general have weight0\.50\.5, and human preference has weight0\.750\.75\. A second acceptance rule applies when the main dimensions saturate\. It accepts a candidate whose main scores stay within0\.0050\.005of the champion and whose weighted gain on the other dimensions is at least0\.0050\.005, with a bootstrap lower bound above 0\. Rules 1 and 4 still apply\. The last stage\-1 champion \(c046\) was accepted through this path\. In stage 2, no family with at least 40 questions may drop by more than0\.120\.12, and no VideoGen\-RewardBench aspect by more than0\.050\.05\. A reviewer agent can still veto an accepted candidate\.
#### Scoring\.
A pairwise question scores 1, 0\.5 \(tie\) or 0\. A rating question with targetttscores1−\|𝔼\[r\]−t\|/41\-\|\\mathbb\{E\}\[r\]\-t\|/4, and a multiple\-choice question scores 1 or 0\. A dimension score is the mean over its families of the mean question score\. Motion pairs use clips whose optical flow differs by at least2\.5×2\.5\\times\(1\.51\.5–2\.5×2\.5\\timesfor hard pairs\)\. Real\-video pairs use generated videos whose human scores differ by at least 2 points\.
### A\.4Unseen defects and shared questions
Ten test families of the main dimensions never appear in the development banks or the training data: crossfade drift \(two strengths\) and object texture swap for consistency; object scale pop, teleport \(two sizes\) and upside\-down clips for physics; and local frame shuffles \(two densities\) and speed bursts for smoothness\.[TableS4](https://arxiv.org/html/2611.23890#A1.T4)shows that EVA\-RSI gains0\.1590\.159on these 440 questions \(95% interval\[0\.116,0\.200\]\[0\.116,0\.200\]\), compared with0\.2470\.247on seen families\. Smoothness has the largest unseen\-family gain\. The two stages’ test banks share 1,254 questions; without them, the main\-mean gain is0\.1940\.194, compared with0\.2000\.200on the full bank\.
QuestionsStartEVA\-RSIGain \[95% int\.\]Seen families0\.6620\.909\+0\.247\+0\.247Unseen families0\.7540\.913\+0\.159\+0\.159\[0\.116, 0\.200\]physics0\.8490\.943\+0\.094\+0\.094consistency0\.9851\.000\+0\.015\+0\.015smoothness0\.4280\.795\+0\.367\+0\.367\[0\.252, 0\.479\]Main mean, all0\.7060\.906\+0\.200\+0\.200\[0\.181, 0\.219\]Main mean, not shared0\.7090\.902\+0\.194\+0\.194\[0\.172, 0\.216\]Table S4:Unseen defects and shared questionson the hidden test bank\. Families: mean over physics, consistency and smoothness\. Not shared: without the questions that the stage\-1 test bank also contains\. Intervals: bootstrap over clip clusters\.
### A\.5Replaying the acceptance rule on the hidden test
During the loop, only champions were scored on the hidden test bank\. Afterwards, we also scored the stage\-2 candidates that were rejected despite a clear development gain\. These candidates failed only on the Nash product or a family drop\.[TableS5](https://arxiv.org/html/2611.23890#A1.T5)compares each of them with its parent\. Their test means stay close to their parents’ scores\. All three candidates rejected by the Nash criterion lose score on physics\.
#### Simpler acceptance rules\.
For these recorded candidates, removing the Nash condition changes three decisions to acceptance\. Their test\-mean changes range from−0\.002\-0\.002to\+0\.003\+0\.003, with lower physics scores in each case\. This replay diagnoses decisions on a fixed trajectory; a new search under another rule would generate a different trajectory\. A rule of mean gain alone would also have accepted the best of the first ten stage\-1 candidates \(c006\), which does not transfer to public benchmarks \([Tab\.S6](https://arxiv.org/html/2611.23890#A1.T6)\)\.
Cand\.ParentDecisionDev gainTest mainΔ\\DeltamainΔ\\Deltahumanc010c001Nash\+0\.011\+0\.0110\.896\+0\.000\+0\.000−0\.007\-0\.007c035c001Nash\+0\.011\+0\.0110\.894−0\.002\-0\.002−0\.003\-0\.003c056c033Nash\+0\.013\+0\.0130\.905\+0\.003\+0\.003−0\.006\-0\.006c033 \(54 frames\)c033family drop\+0\.011\+0\.0110\.904\+0\.002\+0\.002\+0\.002\+0\.002c001stage\-1 champ\.accepted\+0\.021\+0\.0210\.896\+0\.024\+0\.024\+0\.016\+0\.016c033c001accepted\+0\.010\+0\.0100\.902\+0\.006\+0\.006\+0\.000\+0\.000c071c033accepted\+0\.014\+0\.0140\.906\+0\.004\+0\.004\+0\.004\+0\.004
Table S5:Acceptance replay on the stage\-2 hidden test bank\.Top: rejected candidates with a development gain of at least0\.010\.01\. Bottom: champions\.Δ\\Delta: test score minus the parent\.
### A\.6Baselines without recursion
ModelFramesPhys\.Cons\.Smooth\.Dyn\.RealSem\.Qual\.HumanGen\.MainPublicGPU hRuleQwen3\-VL\-8B320\.7390\.8770\.4610\.7690\.6840\.9880\.9680\.6240\.6850\.7060\.39––Union SFT540\.8590\.9770\.8440\.9040\.8590\.9920\.9840\.5580\.7260\.8890\.53384human floorReal\-only SFT540\.8440\.9070\.6080\.7510\.8620\.9920\.9600\.5980\.7150\.7940\.576\.2human floorUnion \+ anchors, seed 1540\.8670\.9800\.6830\.8530\.8560\.9920\.9780\.6470\.7150\.8480\.66352acceptUnion \+ anchors, seed 2540\.8730\.9690\.6790\.8470\.8540\.9920\.9640\.6540\.7230\.8440\.60352acceptc033 recipe, seed 1320\.8500\.9150\.5140\.7720\.7330\.9920\.9650\.6320\.7040\.7570\.467\.2general floorc033 recipe, seed 2320\.8510\.9150\.5050\.7750\.7370\.9920\.9690\.6080\.7000\.7560\.457\.3acceptBest of 10 \(c006\)320\.8450\.9570\.8540\.9100\.7170\.9960\.9950\.6110\.7040\.8570\.40≤\\leq60human floorNon\-recursive search \(c004\)320\.8420\.9750\.7260\.9160\.8620\.9960\.9950\.6100\.7040\.864–212human floorEVA\-RSI540\.8580\.9970\.8880\.9230\.8630\.9920\.9960\.6600\.7260\.9060\.71944accepted
Table S6:Baselines without the loop on the hidden test bank\.Main: mean of the five main dimensions\. Public: mean rank score as in[Fig\.3](https://arxiv.org/html/2611.23890#S4.F3)c\. Rule: decision on the development banks \([Sec\.A\.6](https://arxiv.org/html/2611.23890#A1.SS6)\)\. Bold: best\.[TableS6](https://arxiv.org/html/2611.23890#A1.T6)lists every model that we trained from Qwen3\-VL\-8B without the loop\. Each model is scored on the stage\-2 test bank with its own frame count\.
- •Union SFTandreal\-only SFTtrain one LoRA on all the loop’s training data, or on its real videos only\.
- •Union \+ anchorsadds the anchor answers that c033 uses to keep human preference and general understanding\.
- •c033 reciperepeats the full recipe of the c033 step, but starts from Qwen3\-VL\-8B and not from the champion before it\.
- •Best of 10is the best, by development gain, of the first ten stage\-1 candidates \(c001–c010, rounds 1–6\); candidate IDs follow proposal order\.
The last column applies the acceptance rule with Qwen3\-VL\-8B as the champion\. Three of these models would be accepted\. Union SFT used 384 GPU hours\. Each Union \+ anchors seed used 352 GPU hours on 32 GPUs \(11 wall\-clock hours\)\. The full EVA\-RSI search used 944 GPU hours; these comparisons evaluate resulting models at different compute budgets\. Union SFT uses all training pools, including the stage\-2 pools that humans built after stage 1\. It reaches0\.8890\.889on test, fails the human\-preference floor and transfers less to public benchmarks; EVA\-RSI is significantly better on test and on human preference \(95% intervals exclude zero\)\. Real\-only SFT reaches0\.570\.57on the public score and0\.7940\.794on test\. Union \+ anchors passes the rule and trails EVA\-RSI by0\.0580\.058and0\.0620\.062on test across two seeds, with the largest gap on smoothness\.
#### Non\-recursive search\.
The non\-recursive control used the same 16 GPUs and development banks, judging candidates against Qwen3\-VL\-8B\. Its context omitted leader notes, scores and outcomes\. We stopped after observed stagnation at 212 GPU hours, short of the full loop’s 944 GPU hours\. Since no candidate passes, we score on the test bank the three best by development main mean and the one closest to passing \(c008\)\. EVA\-RSI is significantly above both the pick by development gain \(c004,[Tab\.S6](https://arxiv.org/html/2611.23890#A1.T6)\) and the best of the four on test \(c005\)\.
## Appendix BHow the evaluator evolved
During the search, test evaluation covered champions\. Post\-search analyses also evaluated rejected candidates and controls \([Secs\.A\.5](https://arxiv.org/html/2611.23890#A1.SS5)and[A\.6](https://arxiv.org/html/2611.23890#A1.SS6)\)\.
### B\.1Champion lineage
[Table3](https://arxiv.org/html/2611.23890#S4.T3)in the main paper lists the eight champions with their recipes\. Most champions build on the previous one, and several combine parts of earlier candidates\.
### B\.2How the candidates ended
[Table5](https://arxiv.org/html/2611.23890#S4.T5)and[Fig\.S1](https://arxiv.org/html/2611.23890#A2.F1)\(a, b\) sort the 170 candidates of stages 1 and 2 by how they ended\. Eight became champions\. Most scored candidates were rejected because their gain was too small\. About half of all candidates never reached scoring, mostly because their training job failed or their code was rejected\.[FigureS1](https://arxiv.org/html/2611.23890#A2.F1)\(c\) adds the 100 candidates of stage 3\.
Figure S1:Outcomes of the candidates\.\(a–c\) Outcome per stage\. In stage 3 \([AppendixA](https://arxiv.org/html/2611.23890#A1)\), no candidate met the frozen rule, two passed only a looser rule \(teal\), and three training jobs failed \(unlabelled slices\)\. \(d\) Why scored candidates were rejected, stages 1 and 2\. \(e\) Candidates by proposing agent, stages 1 and 2; the dark part of each bar counts the accepted ones; the semantics and quality roles of stage 1 are counted as general\.#### Roles\.
The agents did not propose equally often\. In stage 1, the human\-preference agent proposed about a third of the candidates\. In stage 2, the general and physics agents proposed the most\.
### B\.3What worked
#### Smoothness in stage 1\.
The second champion \(c009\) gave the largest single gain in smoothness\. It added graded time\-reversal labels and defect pairs in which both clips move at the same speed\. Matching the speed forces the model to distinguish the clips by their motion defects\.
#### Real generated videos in stage 2\.
The first stage\-2 champion \(c001\) trained on real generated videos with human labels\. It gave the largest gain on real videos, on both the development and the test bank\.
#### Keeping general understanding\.
In stage 2, general understanding was the dimension closest to its floor\. c033 added the base model’s own answers to multiple\-choice questions, with shuffled answer order, as anchors\. This kept general understanding above its floor while the main dimensions still rose\.
#### Telling forward from backward\.
EVA\-RSI \(c071\) learned from a teacher whether a clip plays forward or backward\. The teacher is built from forward and reversed clips of the training split and sees the same 54 frames that EVA\-RSI is scored on\. This step raised smoothness from0\.8500\.850to0\.8880\.888and the main mean from0\.9020\.902to0\.9060\.906\.
### B\.4What failed
#### Trading one family for the mean\.
In stage 1, the rule saw only the means of the dimensions\. So it accepted c009 although one family \(time reversal\) fell sharply, because the main mean rose\. Later the same family fell again on the test bank\. Stage 2 therefore added guards for single families \([Sec\.A\.3](https://arxiv.org/html/2611.23890#A1.SS3)\)\.
#### Physics drifting in stage 2\.
Across the last two champions, test physics changed from0\.8860\.886to0\.8580\.858and the main mean from0\.8960\.896to0\.9060\.906\([Tab\.S3](https://arxiv.org/html/2611.23890#A1.T3)\)\.
#### Floors and failed jobs\.
The human\-preference floor rejected many candidates, mostly in stage 1\. Many training jobs failed, and some implementations were rejected before training \([Tab\.5](https://arxiv.org/html/2611.23890#S4.T5)\)\.
#### Text alignment\.
No candidate that targeted text alignment was accepted\. This matches the public benchmarks: on text alignment, EVA\-RSI scores above its starting model on AIGVE and below it on VideoScore\-Bench\-v2 and below VideoScore2 \([Tab\.S7](https://arxiv.org/html/2611.23890#A4.T7)\)\.
### B\.5Agents and cost
The agents are headless coding\-agent sessions, one per role and call\. Each has a fixed role prompt, can read the champion’s development scores and past candidates, and can write only the candidate’s code\. Stage 1 used Claude Opus 5\.5, with GPT\-5\.3\-Codex as a fallback on usage limits\. Stage 2 used Claude Opus 5\.5, with Claude Sonnet 5\.5 and Claude Opus 4\.6 as fallbacks\. Training and scoring of candidates used about 940 B200 GPU hours \(358 in stage 1 and 586 in stage 2\); measured over active candidate jobs\. The stage\-2 agent calls cost about $1,700\. A candidate trains for at most 1\.5 hours and, with scoring, takes at most about two hours on 4 GPUs\.
#### Reuse of the development banks\.
Every candidate is scored on the same development banks, so we report gains on the hidden test bank\.
#### Release\.
We will release the edit code, the development and test banks, the frozen scorer and rule, the agent prompts and stage\-2 transcripts, the champion checkpoints, and the RL code with configs\. Stage\-1 champions are kept; the stage\-1 state file was lost\.
## Appendix CLessons from running RSI
We ran the loop in two stages and a third stage that found nothing new \([AppendixA](https://arxiv.org/html/2611.23890#A1)\)\. Six lessons held across them\.
#### Agent choice affects the search\.
With Claude Opus 5\.5 as the agent, the loop produced eight champions and raised the hidden\-test score from0\.7060\.706to0\.9060\.906\. With Grok as the agent, from the same start and with the same exam and rule, it produced no new champion during the run\.
#### The verifier bounds the result\.
The first stage\-2 champion produces the largest step in real\-video performance after the exam adds human\-labeled real videos\. Through stage 1, scores on real generated videos stayed near0\.750\.75\(0\.7000\.700to0\.7550\.755over five champions\), with the public score remaining near its starting level\. Once the exam had real videos with human labels, the first stage\-2 champion raised it to0\.8520\.852in one step \([Tab\.S3](https://arxiv.org/html/2611.23890#A1.T3)\)\. A good exam is close to real failures, fine\-grained enough to separate strong candidates, and truly hidden\. Families that the start model already solves give no signal and should be replaced by harder versions between stages\.
#### Human revisions to the exam\.
Adding real videos was a human decision; public\-benchmark gains followed in stage 2\. Humans also catch exam hacks that the rule allows: in stage 1, a candidate raised the smoothness mean by giving up the time\-reversal family, which led to guards on single families\. The development and test banks can disagree on a family, such as time reversal or upside\-down clips, so the loop should not tune to one family on the development banks alone\. Finally, humans judge the generated videos by eye\.
#### Training failures account for the most unscored candidates\.
Of the 170 candidates, 52 never got a score because their training job failed, and 20 because their code was rejected \([Tab\.5](https://arxiv.org/html/2611.23890#S4.T5)\)\. Smoke checks before a job starts, a hard time limit and a repair agent that reads the log reduce this waste\.
#### Returns diminish\.
After a few champions, most candidates land within±0\.005\\pm 0\.005of the champion\. The 14 stage\-2 candidates scored after EVA\-RSI landed between−0\.004\-0\.004and\+0\.001\+0\.001, and none of the 100 candidates of the third stage met the rule\. These runs motivate expanding the exam and training data\.
#### Cost of rereading the search history\.
In stage 2, half of the agent cost went to engineer sessions, because every call re\-read a growing ledger \(10k to 28k tokens\) and the stacked champion code\. In the third stage, a rolling lessons file of at most 700 words, a short context, a code map and pruning of dead code kept the context flat for 100 candidates\.
## Appendix DComparison protocol
Mean rank scoreDynamicsSelected metricsModelTypeEval\. \(15\)Underst\. \(2\)VSB dyn\.VBench dyn\.VideoPhy\-2 phys\.VideoScore\-BenchVBench pref\.T2VQA\-DBAIGVE align\.VSB\-v2 textEVA\-RSI \(ours\)open, 8B0\.7210\.4210\.6110\.9060\.3170\.5980\.7170\.5540\.6560\.548After stage 1 \(ours\)open, 8B0\.4340\.4470\.5640\.7810\.0920\.2950\.7070\.3090\.5180\.556Qwen3\-VL\-8B \(start\)open, 8B0\.4070\.4470\.5380\.8120\.1080\.3270\.6410\.2630\.5990\.570Qwen3\-VL\-32Bopen0\.5420\.7110\.4950\.8750\.1560\.3470\.6850\.3340\.6300\.674Qwen3\.8\-27B \(thinking\)open0\.6280\.7110\.4530\.7500\.2210\.3890\.6010\.4440\.6960\.663InternVL3\.5\-8B\[[53](https://arxiv.org/html/2611.23890#bib.bib46)\]open0\.3400\.4210\.4330\.7500\.1240\.3610\.6610\.3210\.6560\.643VideoScore2\[[17](https://arxiv.org/html/2611.23890#bib.bib45)\]evaluator0\.5630\.0260\.5340\.4060\.2990\.5080\.6000\.3960\.7070\.677UnifiedReward\-7B\[[56](https://arxiv.org/html/2611.23890#bib.bib39)\]evaluator0\.5930\.1320\.4860\.5000\.2330\.3750\.6200\.4960\.7470\.627VideoRewardevaluator0\.425–−\-0\.3110\.3120\.1720\.1680\.5760\.1730\.7230\.552Claude Opus 5\.5proprietary0\.8260\.9470\.5490\.7810\.3160\.5310\.6620\.5320\.7570\.587GPT\-5\.6 Solproprietary0\.7750\.8950\.5750\.8440\.2060\.5210\.7000\.5300\.6910\.626Muse Spark 1\.3proprietary0\.7300\.8950\.5120\.8440\.1870\.5010\.7140\.4300\.7090\.610Grok 4\.6proprietary0\.6310\.8950\.5490\.7810\.1620\.4930\.6850\.4570\.6920\.586Gemini 4 Argonproprietary0\.5840\.7370\.4800\.7190\.2710\.4100\.6840\.4880\.7320\.592
Table S7:Comparison on 12 public benchmarks\(14 of 21 models; the other 7 are listed below\)\. Mean rank score: rank on each metric mapped to\[0,1\]\[0,1\], then averaged\. Bold: best per column\.#### Benchmarks\.
We compare 21 models on 12 public benchmarks: VideoPhy\-2\[[5](https://arxiv.org/html/2611.23890#bib.bib14)\], VideoScore\-Bench\[[18](https://arxiv.org/html/2611.23890#bib.bib15)\], VideoScore\-Bench\-v2\[[17](https://arxiv.org/html/2611.23890#bib.bib45)\], VBench preference pairs\[[22](https://arxiv.org/html/2611.23890#bib.bib4)\], AIGVE\[[62](https://arxiv.org/html/2611.23890#bib.bib63)\], T2VQA\-DB\[[24](https://arxiv.org/html/2611.23890#bib.bib56)\], GenAI\-Bench video\[[23](https://arxiv.org/html/2611.23890#bib.bib61)\], MJ\-Bench video\[[48](https://arxiv.org/html/2611.23890#bib.bib62)\], KoNViD\-1k\[[19](https://arxiv.org/html/2611.23890#bib.bib58)\], Q\-Bench\-Video\[[73](https://arxiv.org/html/2611.23890#bib.bib59)\], MVBench\[[28](https://arxiv.org/html/2611.23890#bib.bib16)\]and VideoHallucer\[[57](https://arxiv.org/html/2611.23890#bib.bib60)\]\. They give 15 evaluation metrics and 2 understanding metrics\. Each benchmark is a fixed sample of 67 to 385 items\. The two dynamics columns of[Tab\.S7](https://arxiv.org/html/2611.23890#A4.T7)use the dynamics subsets: 80 VideoScore\-Bench videos, and 24 VBench preference videos from 6 prompts, which give 16 pairs scored with half credit for ties\.[Figure3](https://arxiv.org/html/2611.23890#S4.F3)c ranks each champion alone against the other evaluators of[Tab\.S7](https://arxiv.org/html/2611.23890#A4.T7)except VideoReward, which changes the ranking pool relative to[Tab\.S7](https://arxiv.org/html/2611.23890#A4.T7)\. MVBench supplies the general guard;[Tab\.S7](https://arxiv.org/html/2611.23890#A4.T7)evaluates separate MVBench videos\.
#### How models answer\.
Each model uses its default inference\. Open models, including EVA\-RSI and its earlier champions, give their most likely answer; for ratings, this is the most likely integer from 1 to 5\. Proprietary models give a text answer\. VideoReward gives one reward per video, so we use it only on rating and preference benchmarks\.
#### Uncertainty\.
Per\-metric tests use a 90% paired bootstrap\. Against Qwen3\-VL\-8B, EVA\-RSI is significantly better on 6 metrics and worse on none\. For the mean rank score, we resample the items of each benchmark, re\-rank all 21 models and average \([Tab\.S8](https://arxiv.org/html/2611.23890#A4.T8)\)\. EVA\-RSI is above Qwen3\-VL\-8B in every resample and above every other open model in almost all of them; it is level with Muse Spark 1\.3 and mostly below GPT\-5\.6 Sol and Claude Opus 5\.5\.
StatisticEVA\-RSIClaude Opus 5\.5GPT\-5\.6 SolMuse Spark 1\.3Main result \(21 models, 15 metrics\)0\.7210\.8260\.7750\.73090% interval\[0\.61,0\.76\]\[0\.61,0\.76\]\[0\.71,0\.84\]\[0\.71,0\.84\]\[0\.67,0\.80\]\[0\.67,0\.80\]\[0\.61,0\.75\]\[0\.61,0\.75\]
Table S8:Mean rank scores and bootstrap intervals\.The seven models not shown in[Tab\.S7](https://arxiv.org/html/2611.23890#A4.T7)have these mean rank scores: Qwen3\-VL\-2B0\.1380\.138, Qwen3\-VL\-4B0\.3230\.323, Qwen3\-VL\-30B\-A3B0\.4190\.419\[[2](https://arxiv.org/html/2611.23890#bib.bib5)\], Qwen2\.5\-VL\-7B0\.1890\.189, Qwen2\.5\-VL\-32B0\.4920\.492\[[3](https://arxiv.org/html/2611.23890#bib.bib47)\], LLaVA\-OneVision\-7B0\.1610\.161\[[27](https://arxiv.org/html/2611.23890#bib.bib48)\], and Qwen3\.8\-27B without reasoning0\.5720\.572\.
## Appendix ERL details
#### Reward\.
The evaluator reward is a weighted sum of the dynamics, physics, smoothness and quality ratings \(weights 1, 1, 1 and 0\.5\)\. It has no text\-alignment term\. An anchor loss keeps the generator close to the start model; there is no KL term\. Two penalties act on each video: one for flicker and one for global judder\. Both thresholds are above the start model’s means, so the penalties act only on clear defects\. The EVA\-RSI, Qwen3\-VL\-8B and stage\-1 champion runs use the same reward code, penalties and settings\. All runs use the same settings, which we chose in pilot runs with EVA\-RSI\.
#### VideoReward runs\.
The fair VideoReward run sums its three scores \(visual quality, motion quality and text alignment\) with weight 1 each and uses the same learning rate, anchor and penalties as EVA\-RSI\. We rescale its reward once, on the first rollout, to a within\-group spread of 0\.75, so that the penalties have a similar scale\. The earlier VideoReward run uses only visual and motion quality and no penalties; the figures of still videos come from this run\.
#### A raw optical\-flow reward\.
With one seed and the same settings, a reward of optical flow alone gives significantly more motion than EVA\-RSI at update 50 \(dynamic degree0\.9320\.932against0\.8480\.848\) at similar adherence\. Its videos jitter more: EVA\-RSI is lower in local judder on both prompt sets, a measure of irregular changes of local flow that is not a penalty in either reward\. By update 100 it falls significantly behind EVA\-RSI in both dynamic degree and adherence\. Flow rewards the amount of motion, so the generator learns jittery motion that drifts from the prompt\.
#### Metrics\.
The blind judge, Qwen3\.8\-27B, lists the elements of each prompt and checks whether each video shows them, without knowing the source of the video\. It is from the same model family as EVA\-RSI\. Optical flow \(Farnebäck, on re\-decoded frames\) and VBench dynamic degree \(RAFT\[[47](https://arxiv.org/html/2611.23890#bib.bib49)\]\) share no code or input\. Global judder, one of the two penalties, measures how irregularly the median motion of the frame changes\.
#### Compute\.
[TableS9](https://arxiv.org/html/2611.23890#A5.T9)lists the runs\. At update 50, the runs used about 70 to 80 GPU hours each\. We compare runs at matched update counts\.
RewardUpdatesGPU hoursEVA\-RSI, seed 1342498\.6EVA\-RSI, seed 2100145\.0Qwen3\-VL\-8B, seed 1346498\.0Qwen3\-VL\-8B, seed 2100145\.6Stage\-1 champion100144\.4VideoReward130 \([Tab\.6](https://arxiv.org/html/2611.23890#S5.T6): 120\)188\.7VideoReward \(fair\), seed 15070\.6VideoReward \(fair\), seed 25070\.4Table S9:RL runsfrom Wan2\.1 T2V\-1\.3B\. The two long runs have a budget of 500 GPU hours\.
#### Choice of updates\.
Update 50 is the main comparison point, and update 100 the second; we fixed update 100 before evaluating EVA\-RSI there\. A stopping rule based on cross\-prompt similarity and a flow floor passes EVA\-RSI and Qwen3\-VL\-8B at update 50, and stops VideoReward and the stage\-1 champion\. We report all runs at the same updates\. The full\-budget rows of[Tab\.6](https://arxiv.org/html/2611.23890#S5.T6)are a stress test past the stopping points\.
#### Intermediate champion\.
[TableS10](https://arxiv.org/html/2611.23890#A5.T10)adds the stage\-1 champion as a reward and the update\-100 checkpoints\. At update 50, the stage\-1 champion already gives more motion than Qwen3\-VL\-8B\. With seed 1, EVA\-RSI gives about the same dynamic degree and less optical flow than the stage\-1 champion; with seed 2, it gives more of both\. At update 100 on the general prompts, the gain of the stage\-1 champion over Qwen3\-VL\-8B is no longer significant\.
RewardUpd\.Dyn\.↑\\uparrowFlow↑\\uparrowAdh\.↑\\uparrowElem\.↑\\uparrowStage\-1 champion500\.85718\.773\.6590\.748Qwen3\-VL\-8B, seed 11000\.77014\.933\.7540\.761Qwen3\-VL\-8B, seed 21000\.70013\.483\.6710\.748Stage\-1 champion1000\.80315\.293\.6480\.743EVA\-RSI, seed 11000\.91320\.533\.6750\.750EVA\-RSI, seed 21000\.89217\.593\.5190\.712
Table S10:More RL checkpoints\(630 test videos\)\.
#### Evaluator ratings\.
Evaluator ratings, flicker and global judder enter the training reward\. We compare reward models with independent motion metrics, a separate adherence judge and human ratings\.
### E\.1Full statistics of the RL comparison
#### Testing protocol\.
All differences are paired over prompts and seeds\. Intervals are 95% bootstrap intervals from 4,000 resamples, andpp\-values are two\-sided\. Videos of one prompt share content, so we also resample whole prompts; all confirmatory results hold under both schemes\. A difference is significant whenp<0\.05p<0\.05and its interval excludes zero\. Intervals quantify variation across test prompts within a training run\. We report results per seed and the gap between two Qwen3\-VL\-8B seeds\.
#### Confirmatory tests\.
We fixed one family of 32 tests before the confirmatory analysis\. It covers VBench dynamic degree and optical flow, on both prompt sets, both seeds and updates 50 and 100, each against the start model and against Qwen3\-VL\-8B of the same seed and update\. All 32 tests pass after Holm correction\. In 31 of them, no resample out of 4,000 crosses zero\. These tests havep<0\.0005p<0\.0005, giving an upper bound of0\.0160\.016on all adjustedpp\-values\. Some bootstrap summaries of seed 1 at update 50 were already visible when we fixed this family, and we widened it from 16 to 32 tests at that point\. All other metrics are described, not tested as a family\.
#### Per\-seed results\.
[TableS11](https://arxiv.org/html/2611.23890#A5.T11)gives every paired difference against Qwen3\-VL\-8B of the same seed and update\.
630 test videos300 general promptsSeed12121212Update50501001005050100100Motion↑\\uparrowDynamic degree\+0\.06∗\+0\.06^\{\*\}\+0\.14∗\+0\.14^\{\*\}\+0\.14∗\+0\.14^\{\*\}\+0\.19∗\+0\.19^\{\*\}\+0\.07∗\+0\.07^\{\*\}\+0\.10∗\+0\.10^\{\*\}\+0\.13∗\+0\.13^\{\*\}\+0\.14∗\+0\.14^\{\*\}Optical flow\+2\.0∗\+2\.0^\{\*\}\+4\.8∗\+4\.8^\{\*\}\+5\.6∗\+5\.6^\{\*\}\+4\.1∗\+4\.1^\{\*\}\+2\.7∗\+2\.7^\{\*\}\+6\.0∗\+6\.0^\{\*\}\+5\.4∗\+5\.4^\{\*\}\+4\.3∗\+4\.3^\{\*\}Object motion\+0\.09∗\+0\.09^\{\*\}\+0\.54∗\+0\.54^\{\*\}\+0\.51∗\+0\.51^\{\*\}\+0\.27∗\+0\.27^\{\*\}\+0\.17∗\+0\.17^\{\*\}\+0\.74∗\+0\.74^\{\*\}\+0\.54∗\+0\.54^\{\*\}\+0\.37∗\+0\.37^\{\*\}Judge↑\\uparrowAdherence−0\.01\-0\.01\+0\.03\+0\.03−0\.08\-0\.08−0\.15∗\-0\.15^\{\*\}0\.000\.00−0\.08\-0\.08−0\.04\-0\.04−0\.10\-0\.10Elements0\.000\.00\+0\.01\+0\.01−0\.01\-0\.01−0\.04∗\-0\.04^\{\*\}0\.000\.00−0\.02\-0\.02−0\.02\-0\.02−0\.03∗\-0\.03^\{\*\}Table S11:EVA\-RSI minus Qwen3\-VL\-8B as the RL reward, same seed and update\.∗: 95% interval excludes zero \(whole prompts resampled\)\. Dynamic degree and optical flow form the Holm family\.EVA\-RSI gives more motion in all eight cells, by all three motion measures, including object motion after camera motion is removed\. Prompt adherence does not fall significantly at update 50; it falls for seed 2 at update 100, on both prompt sets\. In the same cell, seed 2 is also below the start model in adherence on the 630 test videos, while seed 1 is level with the start model\. On the 630 videos, two Qwen3\-VL\-8B seeds differ in dynamic degree about as much as the EVA\-RSI lead for seed 1 at update 50\. In the other cells, the lead is at least twice this seed gap\. The automatic metrics measure the amount of motion; the human study below judges whether the videos are better\.
#### Adherence over training\.
From update 50 to update 100, adherence falls in both seeds of EVA\-RSI, more in seed 2\. Over the same updates, Qwen3\-VL\-8B stays roughly level\. At the full budget, EVA\-RSI still generates motion and scores below the base model on adherence\. The Qwen3\-VL\-8B run degrades much more by then, so at the full budget EVA\-RSI is ahead on adherence \(one seed each\)\.
#### Cross\-prompt similarity\.
Videos for different prompts become more alike under several rewards\. At update 100, both EVA\-RSI seeds are more alike across prompts than the Qwen3\-VL\-8B seed\.
#### Human study\.
We ran a blind study on the RL videos at update 50 \(training seed 1\)\. Each pair shows two videos for the same prompt and sampling seed side by side, in random order and without labels\. For each pair, a rater answers four questions: which video has more motion, better visual quality and better prompt adherence, and which is better overall\. Each question also allows “same”\. There are 48 pairs against Qwen3\-VL\-8B and 24 against fair VideoReward, drawn evenly from the four prompt subsets of the 630 test videos \(three VBench prompt lists and VidProM\)\. Six raters answered all pairs\. We exclude one more answer file, which came from an earlier version of the study with data problems; including it changes no score by more than0\.020\.02\. The score is the share of answers for EVA\-RSI, with “same” counted as half, and the 95% intervals resample raters and pairs\.
Raters prefer EVA\-RSI on motion and overall against both references\. The motion answers follow the videos: on the 27 pairs whose optical flow differs by at least2×2\\times, raters pick the video with more flow in89%89\\%of their answers that are not “same”\. Raters also prefer EVA\-RSI on visual quality and prompt adherence \([Tab\.7](https://arxiv.org/html/2611.23890#S5.T7)\)\. Against Qwen3\-VL\-8B, raters see more motion in the EVA\-RSI video in 171 of 288 answers\. In these answers, they prefer EVA\-RSI overall, in visual quality and in adherence about nine times in ten, and judge it worse overall in3%3\\%\. Against fair VideoReward, they see more motion in the EVA\-RSI video in 112 of 144 answers and prefer it overall in most of them\.
## Appendix FThe RSI evaluator as a data filter
We also tried EVA\-RSI as a filter for fine\-tuning data\. Both arms fine\-tune Wan2\.1 T2V\-1\.3B on 50k clips with the same recipe\. Data from our selection pipeline gives more motion at both steps \([Tab\.8](https://arxiv.org/html/2611.23890#S5.T8)in the main paper\)\. The random arm first gains motion and then falls back to the start model, while the picked arm keeps its motion\. Adherence does not differ\.
#### Choice of checkpoints\.
We chose the two evaluated steps after training\. We ran three picked\-versus\-random pairs: this pair, a smaller pair with about 20k clips per arm, whose gain in dynamic degree did not hold at the later step, and an earlier 50k pair with another recipe, which also favors selected data\.
## Appendix GMore visual results
#### Single test pairs\.
[Figure4](https://arxiv.org/html/2611.23890#S4.F4)shows three test pairs\. Qwen3\-VL\-8B rates the flawed video about the same as the good one, or even higher\. EVA\-RSI assigns the flawed video a lower score in all three\. We selected the examples after fixing EVA\-RSI\.[Table2](https://arxiv.org/html/2611.23890#S4.T2)reports normalized scores on the full test bank\.
#### How the evaluator changes over RSI\.
[FigureS2](https://arxiv.org/html/2611.23890#A7.F2)follows single questions through the chain of champions, from Qwen3\-VL\-8B to EVA\-RSI\. In every row, EVA\-RSI gives the correct answer the highest probability, or ties with an earlier champion\. On the real\-video questions, the jump comes with stage 2, which added generated videos with human labels\. On the edited questions, the last champion gives the highest probability, with its largest gains on smoothness\. None of these questions appears in the main paper\. We selected these questions after fixing EVA\-RSI\.
Figure S2:Single questions through the chain of champions\.Bars: probability of the correct answer\. ✓: correct video\.Figure S3:More prompts where EVA\-RSI moves most\(update 50\)\. VideoReward: stress\-test run\.
#### Downstream videos\.
[Figure5](https://arxiv.org/html/2611.23890#S5.F5)and[Figs\.S4](https://arxiv.org/html/2611.23890#A7.F4)and[S3](https://arxiv.org/html/2611.23890#A7.F3)use the same layout: RL of Wan2\.1 T2V\-1\.3B with three rewards, all at update 50, on the same prompt and seed\. Each cell shows three frames, and the corner number is the optical flow of the video\. We picked the examples in[Fig\.S4](https://arxiv.org/html/2611.23890#A7.F4)from test prompts where EVA\-RSI keeps adherence and moves more than Qwen3\-VL\-8B, and VideoReward is nearly still\. We used a minimum HPSv3\[[37](https://arxiv.org/html/2611.23890#bib.bib3)\]image\-preference score solely to select these examples\. For[Fig\.S3](https://arxiv.org/html/2611.23890#A7.F3)we used the same rule with imaging quality in place of HPSv3: EVA\-RSI moves at least1\.8×1\.8\\timesas much as Qwen3\-VL\-8B, VideoReward is nearly still, and EVA\-RSI is not worse on adherence or imaging quality\. None of these prompts appears elsewhere in the paper\. In both figures, VideoReward often turns the prompt into a still close\-up of a face, and in the garden prompt it drops the truck\. Over all 630 test videos \(seed 1, update 50\), EVA\-RSI has more optical flow than Qwen3\-VL\-8B in57%57\\%, more than the fair VideoReward run in92%92\\%and more than the stress\-test run in94%94\\%\.[FigureS6](https://arxiv.org/html/2611.23890#A7.F6)follows two prompts over training, from the eight fixed videos that we log at each evaluation; its last update is past the stopping point of the run\.[FigureS5](https://arxiv.org/html/2611.23890#A7.F5)shows the four test prompts where SFT on clips picked by EVA\-RSI is most dynamic against SFT on random clips \([AppendixF](https://arxiv.org/html/2611.23890#A6)\)\.
Figure S4:More examples\(update 50\)\.Figure S5:Largest motion gains of SFT on data picked by EVA\-RSI\.Figure S6:One prompt and seed over RL training\.相似文章
Video-RSI:通过 Harness 进化实现视频理解智能体的递归自我改进
Video-RSI 提出一种通过 harness 进化实现视频理解智能体递归自我改进的框架,无需更新模型权重,通过重新审视训练视频、收集证据并诊断失败来迭代改进提示、工具和记忆,从而在更少帧处理量下提升准确率。
递归自我改进已经到来。Sakana AI 展示了可运行的 RSI 循环。
Sakana AI 提出了多智能体自我监督(MASS),这是一种同构递归自我改进(RSI)循环,其中同一个模型同时扮演执行器、优化器和评估器的角色,将多智能体群体智能蒸馏为自我监督信号,并在研究基准上提升单位 token 得分表现。
RewardVerse:评分标准指导的策略优化用于视频奖励建模
RewardVerse 提出了一种基于评分标准的视频奖励建模框架,以缓解标量漂移并提供稳定的评估标准,从而增强视频生成中的强化学习。
RSI Arena:迈向基于人类反馈的递归自我改进
RSI Arena 推出了一项分为两个阶段的研究:十个前沿 AI 智能体在固定的 GPU 和 API 预算下,以人类偏好反馈为指导,从同一个基础模型出发,自主开展语言模型后训练。表现最强的智能体(GPT-6 Astra)弥合了基础模型与指令模型之间差距的 76%,且大多数模型在第二阶段比第一阶段更受人类青睐,展示了在结合人类反馈的递归自我改进(RSIHF)方向上迈出的一步。
Video-Arena:一个面向时空能力的可扩展智能体视频基准
Video-Arena 引入了一个可扩展的基准,将竞技类游戏转化为对视频中时空世界状态理解与跟踪能力的可验证测试,在原生与工具辅助的智能体协议下对模型进行评估,结果发现两种协议下的模型都难以随时间推移正确更新状态。