ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

# Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment
Source: [https://arxiv.org/html/2608.05102](https://arxiv.org/html/2608.05102)
\\newtcolorbox

promptbox\[1\] enhanced, breakable, colback=promptbg, colframe=promptred, coltitle=white, colbacktitle=promptred, title=\#1, fonttitle=, boxrule=1pt, arc=2mm, left=8pt, right=8pt, top=8pt, bottom=8pt, titlerule=0pt

Yijun Lu1,\*,Rui Ye1,\*,†,Jiajun Wang1,Yuwen Du1,Tian Jin1,Songhua Liu1,†,Siheng Chen1,† 1Shanghai Jiao Tong University,\*Equal Core Contributions †Corresponding Authors: \{yr991129, liusonghua, sihengc\}@sjtu\.edu\.cn

###### Abstract

Long\-horizon search agents must make multiple sequential actions \(steps\) to search, retrieve, verify, and integrate evidence to reach a final answer\. However, existing methods for training these agents typically treat all steps within a trajectory uniformly during both supervised fine\-tuning \(SFT\) and reinforcement learning \(RL\), failing to distinguish useful actions from erroneous or redundant ones\. In this paper, we proposeAnswer\-Backtracked Credit Assignment \(ABC\), a fine\-grained credit assignment framework for training long\-horizon search agents by converting sparse trajectory\-level outcomes into dense step\-level supervision that rewards useful actions \(even in failed trajectories\) while suppressing erroneous or redundant actions\. Specifically, given a potentially obscure query and its corresponding ground\-truth answer, ABC first performsAnswer\-Backtracked Clue Recovery, which traces back from the answer to recover intermediate clues required to solve the question\. It then appliesClue\-Anchored Step Scoringto evaluate each search step against these clues, converting sparse binary outcome supervision into dense step\-level rewards\. Based on these rewards, we developABC\-SFT, which reweights the loss of each turn, andABC\-GRPO, which uses the step\-level scores as rewards in GRPO\. Building on this framework, we trainABSeekerbased on Qwen3\.5\-4B with only 8\.5k examples\. ABSeeker achieves 37\.3% on BrowseComp and 39\.1% on BrowseComp\-ZH\. With context management, the scores further improve to 55\.3% and 52\.9%, respectively, significantly outperforming same\-scale \(4B\) agents and even matching the performance of larger ones \(∼\\sim30B\)\. These results demonstrate the effectiveness of answer\-backtracked step\-level credit assignment for training long\-horizon search agents\.

![Refer to caption](https://arxiv.org/html/2608.05102v1/x1.png)Figure 1:ABSeeker achieves the best performance among 4B models and remains competitive with several larger search agents\. Striped regions indicate results with context management enabled\.## 1Introduction

Search agents have emerged as a powerful approach for solving complex information\-seeking tasks\. Representative systems such as OpenAI Deep Research\(OpenAI,[2025](https://arxiv.org/html/2608.05102#bib.bib31)\), Tongyi DeepResearch\(Teamet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib2)\), and MiroThinker\(Teamet al\.,[2026a](https://arxiv.org/html/2608.05102#bib.bib3)\)move beyond single\-turn retrieval by conducting multi\-step investigations, iteratively formulating queries, inspecting evidence, revising hypotheses, and adapting subsequent actions based on newly acquired information\.

However, training search agents over long interaction horizons introduces a fundamental credit\-assignment challenge\. Existing methods typically treat all steps within a trajectory uniformly during both supervised fine\-tuning \(SFT\) and reinforcement learning \(RL\), without distinguishing their individual contributions to the final outcome\(Jinet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib4); Songet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib5); Chenet al\.,[2025b](https://arxiv.org/html/2608.05102#bib.bib6); Zhenget al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib7); Gaoet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib8); Xieet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib15)\)\. This uniform treatment is particularly problematic for long\-horizon search, where a single trajectory often contains heterogeneous actions: even a successful trajectory may include erroneous or redundant steps, while a failed trajectory may still contain useful actions that retrieve decisive evidence or refine the solution direction\. Treating all steps equally therefore fails to capture individual action quality and limits targeted optimization of critical search and reasoning behaviors\.

To address this challenge, the key is to identify an effective feedback signal that can evaluate a trajectory at a fine\-grained level, distinguishing the contributions of individual actions during training\. Our core idea is motivated by a distinctive property of search tasks: once the ground\-truth answer is available, the task becomes naturally backtrackable\. Starting from the answer, one can recover the key entities, facts, relations, and constraints that should have been discovered during the search process\. These answer\-backtracked clues thus provide a principled basis for assigning differentiated credit to different parts of the trajectory\.

Following this idea, we proposeAnswer\-Backtracked Credit Assignment \(ABC\), a fine\-grained credit assignment framework for training long\-horizon search agents\. Specifically, it consists of two stages: \(1\)Answer\-Backtracked Clue Recovery, which traces back from the verified answer to recover a set of intermediate clues—entities, facts, and relations that collectively characterize the evidence required to solve the question\. \(2\)Clue\-Anchored Step Scoring, which evaluates every search step according to how it discovers, verifies, refines, or incorrectly reasons about the recovered clues, transforming sparse binary outcome supervision into dense turn\-level rewards\. Based on these fine\-grained step rewards, we develop \(1\)ABC\-SFT, built upon standard SFT, which reweights the loss of each turn according to its assigned reward; and \(2\)ABC\-GRPO, built upon standard GRPO, which uses the step\-level scores as rewards\. Together, these methods provide step\-level supervision that rewards useful actions even in failed trajectories, suppresses erroneous or redundant behaviors in successful ones, and enables fine\-grained credit assignment through fixed answer\-backtracked evaluation criteria\.

Building on this framework, we trainABSeekerbased on Qwen3\.5\-4B\(Yanget al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib21)\)\. ABSeeker achieves 37\.3% on BrowseComp\(Weiet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib20)\), 39\.1% on BrowseComp\-ZH\(Zhouet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib23)\), 77\.0% on xbench\-2505\(Chenet al\.,[2025a](https://arxiv.org/html/2608.05102#bib.bib24)\), 46\.0% on xbench\-2510\(Chenet al\.,[2025a](https://arxiv.org/html/2608.05102#bib.bib24)\), and 81\.6% on GAIA\-text\(Mialonet al\.,[2024](https://arxiv.org/html/2608.05102#bib.bib25)\)\. With context management, its performance further improves to 55\.3% and 52\.9% on BrowseComp and BrowseComp\-ZH, respectively, outperforming recent same\-scale 4B baselines, such as QUEST\-4B\(Xieet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib15)\)and Dr\. Venus\(Teamet al\.,[2026b](https://arxiv.org/html/2608.05102#bib.bib16)\), as well as larger∼\\sim30B search agents such as Tongyi DeepResearch\(Teamet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib2)\)and OpenSeeker\(Duet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib18)\)\.

Our main contributions are summarized as follows:

- •We proposeAnswer\-Backtracked Credit Assignment, a fine\-grained credit assignment framework that rewards useful actions in failed trajectories while suppressing erroneous actions in successful ones\.
- •Based on ABC, we developABC\-SFT, which reweights the loss of each turn according to its step reward, andABC\-GRPO, which incorporates step\-level rewards into GRPO\.
- •We first trainABSeekerbased on Qwen3\.5\-4B usingABC\-SFT, and then further optimize it withABC\-GRPO, achieving 37\.3% on BrowseComp and 39\.1% on BrowseComp\-ZH\.

## 2Related Work

Search Agents\.The ReAct paradigm\(Yaoet al\.,[2022](https://arxiv.org/html/2608.05102#bib.bib1)\)established the standard recipe for LLM\-based web agents: interleave reasoning, tool calls, and observations to solve knowledge\-intensive tasks through dynamic interaction with external environments\. Recent work has scaled this framework to long\-horizon search, where agents execute dozens of sequential retrieve–browse–integrate cycles to locate fine\-grained information distributed across multiple sources\. Representative search agents, such as OpenAI’s Deep Research\(OpenAI,[2025](https://arxiv.org/html/2608.05102#bib.bib31)\)and Tongyi DeepResearch\(Teamet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib2)\), have been developed for these tasks\. Despite these advances, most existing agents are trained primarily with trajectory\-level supervision—successful trajectories are treated as uniformly positive and failed trajectories as uniformly negative—providing limited guidance on which intermediate decisions genuinely contribute to finding the correct answer\(Zhenget al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib7); Teamet al\.,[2026a](https://arxiv.org/html/2608.05102#bib.bib3); Chuet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib11); Teamet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib2); Duet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib18)\)\.

Credit Assignment for Search Agents\.Several recent methods have explored fine\-grained credit assignment to address this gap\. IGPO\(Wanget al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib12)\)assigns step\-level rewards based on the increase in the model’s likelihood of the ground\-truth answer, but the resulting credit depends on the model’s own belief estimation and fluctuates with policy updates\. CSO\(Liet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib19)\)identifies critical steps by testing alternative actions and verifying whether they lead to correct outcomes\. However, it assigns credit only to verified critical decisions, providing no direct supervision for the remaining steps in the trajectory\. SAPO\(Liuet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib13)\)and MindDR\(Team and Inc,[2026](https://arxiv.org/html/2608.05102#bib.bib14)\)assign step\-level credit based on intermediate entities, using either their graph proximity to the answer or their coverage in the trajectory\. However, such entity\-level signals cannot directly determine whether each search or reasoning decision is valid\. Collectively, these methods do not provide dense and reliable supervision that directly evaluates the correctness and contribution of every action throughout the trajectory\.

Our Approach\.We proposeAnswer\-Backtracked Credit Assignment \(ABC\), which uses verified answers to construct dense, fine\-grained process supervision\. ABC first applies Answer\-Backtracked Clue Recovery to recover intermediate evidence clues, and then uses Clue\-Anchored Step Scoring to assign each step a scalar score based on its contribution or error\. These scores reweight SFT losses inABC\-SFTand serve as rewards for GRPO inABC\-GRPO, from which we trainABSeeker\.

## 3Method

Figure[2](https://arxiv.org/html/2608.05102#S3.F2)provides an overview of the training pipeline, which consists of two core stages\. Section[3\.1](https://arxiv.org/html/2608.05102#S3.SS1)first formalizes the search trajectory and step\-level credit assignment problem\. Given a query and its verified answer,Answer\-Backtracked Clue Recoveryrecovers a set of intermediate evidence clues that define meaningful progress toward the answer \(Section[3\.2](https://arxiv.org/html/2608.05102#S3.SS2)\)\. Search trajectories are then rolled out, retaining both successful and failed trajectories\. Next,Clue\-Anchored Step Scoringevaluates every step against the recovered clue set and produces dense, fine\-grained step scores \(Section[3\.3](https://arxiv.org/html/2608.05102#S3.SS3)\)\. Finally, these step\-level scores directly serve as the reward signal for policy optimization \(Section[3\.4](https://arxiv.org/html/2608.05102#S3.SS4)\)\.

![Refer to caption](https://arxiv.org/html/2608.05102v1/x2.png)Figure 2:Overview of the training pipeline, which consists of two core stages\.Stage 1: Answer\-Backtracked Clue Recoveryrecovers intermediate evidence clues from the query and its verified answer\.Stage 2: Clue\-Anchored Step Scoringevaluates each step against the recovered clues and produces dense, fine\-grained rewards for each step, in contrast to the sparse answer\-level reward\.### 3\.1Search Trajectory and Credit Assignment

We consider a set of training questions, each consisting of a queryqqand a verified answera∗a^\{\*\}\. A search agent interacts with a web environment overTTturns to produce a search trajectory:

τ=\(s1,s2,…,sT,a\),\\tau=\(s\_\{1\},s\_\{2\},\\ldots,s\_\{T\},\\,a\),\(1\)whereaais the final answer submitted by the agent\. Each stepsts\_\{t\}contains the agent’s reasoning, the issued tool call, and the corresponding tool response returned by the environment\.

During training, the agent typically receives a reward based solely on whether the final answer matches the ground truth:

rans​\(τ\)=\{1,if​a=a∗,0,otherwise\.r\_\{\\text\{ans\}\}\(\\tau\)=\\cases\{1\},&\\text\{if \}a=a^\{\*\},\\\\ 0,&\\text\{otherwise\}\.\(2\)This trajectory\-level signal is sparse and coarse, leading to two fundamental credit\-assignment failures\. First, an incorrect trajectory may contain several useful intermediate steps—such as correct evidence discovery, verification, or candidate filtering—yet the final reward of zero provides no positive signal for these actions\. Second, a correct trajectory may contain erroneous intermediate conclusions or steps that discard useful evidence, yet the final reward of one does not distinguish these flawed actions from genuinely informative ones\. In both cases, trajectory\-level outcome supervision provides no per\-step signal indicating which decisions advanced or impeded progress toward the verified answer\.

ABCaddresses this limitation by constructing a step rewardrtr\_\{t\}for every step in each trajectory\. These rewards are produced byClue\-Anchored Step Scoring\(Section[3\.3](https://arxiv.org/html/2608.05102#S3.SS3)\) based on the clues recovered throughAnswer\-Backtracked Clue Recovery\(Section[3\.2](https://arxiv.org/html/2608.05102#S3.SS2)\)\. The recovered clues provide fixed anchors for evaluating each rollout, enabling stable, dense, and fine\-grained supervision that distinguishes useful actions from erroneous ones regardless of the final outcome\.

### 3\.2Answer\-Backtracked Clue Recovery

The verified answera∗a^\{\*\}specifies where the search should end, but provides little supervision about how the agent should reach it\. To evaluate the quality of intermediate steps, we require a set of answer\-backtracked anchors—concrete pieces of evidence that a valid search process should establish and that can serve as reference points for step\-level evaluation\.Answer\-Backtracked Clue Recoverymaps each training question\(q,a∗\)\(q,a^\{\*\}\)to a set of clues:

𝒞=\{c1,c2,…,cK\},\\mathcal\{C\}=\\\{c\_\{1\},c\_\{2\},\\ldots,c\_\{K\}\\\},\(3\)where eachckc\_\{k\}is a verifiable piece of intermediate evidence relevant to answeringqq, such as a specific entity, fact, attribute, or relationship connecting the query to the verified answer\.

The recovery process leverages the fact that benchmarks such as BrowseComp\(Weiet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib20)\)provide unique and verifiable answers, which offer a clear endpoint for evidence backtracking\. Each question contains a set of constraints that, together with the verified answer, implicitly define a single valid evidence path\. Given the query and its verified answer, we prompt an LLM to reconstruct the evidence chain by identifying the intermediate entities and facts that must be discovered, verified, and cross\-referenced\. Crucially, this backtracking is itself an active ReAct loop: the recovery model conducts web searches and visits pages through the same tool\-call protocol as the forward agent, tracing evidence from the answer back toward the query and anchoring each clue in actual web content\. Clues that survive this verification serve as reliable, answer\-backtracked reference points for subsequent step scoring\.

Figure[3](https://arxiv.org/html/2608.05102#S3.F3)\(left\) illustrates this process using a concrete question from the training set\. Given a four\-constraint query and the verified answerCeraVe, the recovery model produces six clues, includingCeramidesas the clinically supported ingredient,L’Oréalas the acquiring company, andEugène Schuelleras its founder who graduated in 1904\. Together, these clues form a verified evidence chain connecting the query constraints to the answer\. See Appendix[C\.1](https://arxiv.org/html/2608.05102#A3.SS1)for the full recovery prompt\.

![Refer to caption](https://arxiv.org/html/2608.05102v1/x3.png)Figure 3:An illustrative example ofAnswer\-Backtracked Clue RecoveryandClue\-Anchored Step Scoring\.Left:Given a multi\-constraint query and the verified answer, the recovery produces six intermediate evidence clues \(c1c\_\{1\}–c6c\_\{6\}\) that form a verified evidence chain connecting the query to the answer\.Right:Selected steps from a sampled trajectory are evaluated against the recovered clue set, with rewards reflecting the quality and contribution of each step\.
### 3\.3Clue\-Anchored Step Scoring

Given the recovered clue set𝒞\\mathcal\{C\}for a training question,Clue\-Anchored Step Scoringevaluates every step in each collected trajectory\. For each stepsts\_\{t\}, the scorer receives three inputs: \(1\) the current step, including its reasoning, tool call, and tool response; \(2\) the original queryqq; and \(3\) the complete clue set𝒞\\mathcal\{C\}\. It outputs a step rewardrtr\_\{t\}together with a brief rationale explaining the applied criteria and any relevant clues\. Each step starts with a base score of1\.01\.0, ensuring that reasonable exploration without an obvious error is not penalized\. The specific scoring criteria are listed in Table[1](https://arxiv.org/html/2608.05102#S3.T1)\.

Table 1:Clue\-Anchored Step Scoringrubric\.A step may exhibit multiple scored behaviors, and the same behavior may occur multiple times when several clues are affected\. The corresponding deltas in Table[1](https://arxiv.org/html/2608.05102#S3.T1)are accumulated on top of the base score and clipped to\[0,2\.0\]\[0,2\.0\]:

rt=\\operatorname​c​l​i​p​\(1\.0\+∑j∈𝒜tΔj,0,2\.0\),r\_\{t\}=\\operatorname\{clip\}\\left\(1\.0\+\\sum\_\{j\\in\\mathcal\{A\}\_\{t\}\}\\Delta\_\{j\},\\,0,\\,2\.0\\right\),\(4\)where𝒜t\\mathcal\{A\}\_\{t\}denotes the set of scored behavior instances detected at steptt\. A step that discovers a correct clue in a trajectory that ultimately fails still receives positive credit, whereas a step that incorrectly dismisses a correct clue in a trajectory that ultimately succeeds still receives a penalty\.

Figure[3](https://arxiv.org/html/2608.05102#S3.F3)\(right\) illustrates the scoring process\. After recovering the six clues \(c1c\_\{1\}–c6c\_\{6\}\) in Section[3\.2](https://arxiv.org/html/2608.05102#S3.SS2), selected steps from a sampled trajectory are evaluated against the clue set\. Step 22 receives1\.81\.8, consisting of the base score of1\.01\.0and a\+0\.8\+0\.8reward for discoveringCeramidesand linking them toCeraVe\. Step 35 verifies both theL’Oréalacquisition andEugène Schueller’s graduation, coveringc2c\_\{2\}andc3c\_\{3\}; the two positive deltas are accumulated, and the resulting score is clipped to2\.02\.0\. Step 56 receives0\.20\.2because the agent abandons the accumulated evidence supportingCeraVeand returns to an incorrect candidate\. Finally, Step 64 receives0\.00\.0after the agent submits an incorrect final answer\. See Appendix[C\.2](https://arxiv.org/html/2608.05102#A3.SS2)for the full scoring prompt\.

### 3\.4Training with Step Rewards

Given the step rewardrtr\_\{t\}assigned to every step in each trajectory, we train the agent in two consecutive stages: reward\-weighted supervised fine\-tuning \(ABC\-SFT\) followed by step\-level reinforcement learning \(ABC\-GRPO\)\. Both successful and failed trajectories are retained, allowing high\-quality steps to be reinforced while erroneous steps are down\-weighted or penalized\.

#### 3\.4\.1Reward\-Weighted Supervised Fine\-Tuning \(ABC\-SFT\)

In the first stage, we perform SFT on all collected trajectories\. For a trajectoryτ\\tauof lengthTT, letxt,jx\_\{t,j\}denote thejj\-th policy\-generated token at stepsts\_\{t\}, with environment\-provided tool responses excluded from optimization\. The training objective is

ℒSFT​\(θ\)=−∑t=1Tw​\(rt\)​∑jlog⁡pθ​\(xt,j∣xt,<j\),\\mathcal\{L\}\_\{\\text\{SFT\}\}\(\\theta\)=\-\\sum\_\{t=1\}^\{T\}w\(r\_\{t\}\)\\sum\_\{j\}\\log p\_\{\\theta\}\\left\(x\_\{t,j\}\\mid x\_\{t,<j\}\\right\),\(5\)where the step\-level weight is computed via a sigmoid functionw​\(rt\)=σ​\(α⋅\(rt−β\)\)w\(r\_\{t\}\)=\\sigma\\big\(\\alpha\\cdot\(r\_\{t\}\-\\beta\)\\big\), withα\\alphacontrolling the sharpness of the mapping andβ\\betathe neutral baseline\. High\-scoring steps thus contribute more strongly to the gradient, whereas low\-scoring steps contribute little training signal\.

#### 3\.4\.2Step\-Level Reinforcement Learning \(ABC\-GRPO\)

In the second stage, we further optimize the agent through online RL with step\-level rewards\. For each sampled rolloutii, the reward at stepttis defined as

Ri,t=ri,t,R\_\{i,t\}=r\_\{i,t\},\(6\)whereri,tr\_\{i,t\}is the clue\-anchored step score from Section[3\.3](https://arxiv.org/html/2608.05102#S3.SS3)\. We normalize rewards within each rollout group to obtainR^i,t\\widehat\{R\}\_\{i,t\}and compute the discounted step\-level advantage:

Ai,t=∑k=tTiγk−t​R^i,k,A\_\{i,t\}=\\sum\_\{k=t\}^\{T\_\{i\}\}\\gamma^\{k\-t\}\\,\\widehat\{R\}\_\{i,k\},\(7\)whereγ\\gammacontrols how future step rewards are propagated to earlier decisions\. The resulting advantageAi,tA\_\{i,t\}is assigned to all policy\-generated tokens at steptt, while environment\-provided tool responses are masked from optimization\. We optimize the policy using the standard clipped GRPO objective\(Shaoet al\.,[2024](https://arxiv.org/html/2608.05102#bib.bib9)\), replacing its trajectory\-level advantage with the step\-specific advantageAi,tA\_\{i,t\}\.

## 4Experiments

### 4\.1Experimental Setup

Training Setup\.We use OpenSeeker\(Duet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib18)\)as the training data, collecting trajectories with both correct and incorrect final answers\. The maximum number of steps per trajectory is capped at 200\. We use Qwen3\.5\-4B\(Yanget al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib21)\)as the backbone model\. For SFT, we train on 8\.5K trajectories for 3 epochs\. For RL, we sample 1000 questions, each with 8 rollouts, starting from the resulting SFT checkpoint\. For both Answer\-Backtracked Clue Recovery and Clue\-Anchored Step Scoring, we use DeepSeek\-V4\-Flash\(Xuet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib22)\)as the backbone LLM\. Refer to Appendix[A](https://arxiv.org/html/2608.05102#A1)for additional details\.

Evaluations\.We benchmark ABSeeker against the following four evaluation suites, which together span deep research, persistent web navigation, and general\-purpose agent reasoning: \(1\) BrowseComp\(Weiet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib20)\), designed to probe long\-horizon English browsing and information\-seeking through complex multi\-constraint queries; \(2\) BrowseComp\-ZH\(Zhouet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib23)\), a Chinese\-language counterpart that poses similarly difficult retrieval tasks on the Chinese web; \(3\) xbench\(Chenet al\.,[2025a](https://arxiv.org/html/2608.05102#bib.bib24)\), which tests deep research competence—planning, reasoning, and cross\-source synthesis across professional real\-world scenarios; and \(4\) GAIA \(Text\-Only\)\(Mialonet al\.,[2024](https://arxiv.org/html/2608.05102#bib.bib25)\), a suite of general assistant tasks demanding combined web browsing, tool use, and multi\-hop inference\. For all benchmarks the agent is allowed up to 200 tool calls\. We run each evaluation three times and report the averaged score\.

Baselines\.To validate the effectiveness of our method, we compare ABSeeker against three categories of baselines\. \(1\)Frontier foundation modelswith search capability: Gemini\-3\.1\-Pro\(Google DeepMind,[2026](https://arxiv.org/html/2608.05102#bib.bib32)\), Seed\-2\.0\-Pro\(Bytedance Seed,[2026](https://arxiv.org/html/2608.05102#bib.bib30)\), GLM\-5\(Zenget al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib28)\), DeepSeek\-V4\-Pro\-Max\(Xuet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib22)\), and GPT\-5 High\(Singhet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib27)\)\. \(2\)Search agents at∼\\sim30B: MiroThinker\-1\.7\-mini\(Teamet al\.,[2026a](https://arxiv.org/html/2608.05102#bib.bib3)\), RedSearcher\(Chuet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib11)\), DeepMiner\(Tanget al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib29)\), Tongyi\-DeepResearch\(Teamet al\.,[2025](https://arxiv.org/html/2608.05102#bib.bib2)\), and OpenSeeker\(Duet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib18)\)\. \(3\)Search agents at 4B: QUEST\-4B\(Xieet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib15)\), DR\-Venus\(Teamet al\.,[2026b](https://arxiv.org/html/2608.05102#bib.bib16)\), and AgentCPM\-Explore\(Chenet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib17)\)\.

### 4\.2Experimental Results

Table 2:Performance comparison across five benchmarks\. For BrowseComp and BrowseComp\-ZH, \* denotes results obtained without context management\. Within each model category, the bold score denotes the best result on each benchmark\. “–” indicates that the result is not reported\.ModelParamBrowseCompBrowseComp\-ZHxbench\-2505xbench\-2510GAIA\-text*Foundation Model with Tools*Gemini\-3\.1\-Pro––85\.9–53\.053\.080\.6Seed\-2\.0\-Pro–77\.377\.382\.482\.4––78\.678\.6GLM\-5358B75\.975\.972\.772\.7–––DeepSeek\-V4\-Pro\-Max1\.6T83\.4––80\.0–GPT\-5 High–54\.954\.963\.063\.077\.977\.975\.075\.076\.476\.4*Search Agent \(∼\\sim30B\)*MiroThinker\-1\.7\-mini30B67\.972\.3–57\.257\.280\.3RedSearcher30B57\.457\.458\.258\.2––80\.180\.1DeepMiner32B33\.533\.540\.140\.162\.062\.0–58\.758\.7Tongyi\-DeepResearch30B43\.443\.4\*46\.746\.7\*75\.0–70\.970\.9OpenSeeker30B29\.529\.5\*48\.448\.4\*74\.074\.0––*Search Agent \(4B\)*QUEST\-4B4B40\.040\.0–––77\.777\.7DR\-Venus4B29\.129\.1\*37\.737\.7\*74\.774\.740\.740\.764\.464\.4AgentCPM\-Explore4B24\.124\.1\*29\.129\.1\*70\.070\.034\.034\.063\.963\.9ABSeeker4B37\.3\*/55\.339\.1\*/52\.977\.046\.081\.6

Main Results\.Table[2](https://arxiv.org/html/2608.05102#S4.T2)compares ABSeeker with foundation models and representative search agents across five benchmarks\. Among 4B search agents, ABSeeker achieves the best performance on every benchmark, scoring55\.3%on BrowseComp,52\.9%on BrowseComp\-ZH,77\.0%on xbench\-2505,46\.0%on xbench\-2510, and81\.6%on GAIA\-text\. Despite its smaller model size, ABSeeker also remains competitive with substantially larger search agents\. It outperforms all reported 30B agents on xbench\-2505 and GAIA\-text, while surpassing several 30B systems on both BrowseComp and BrowseComp\-ZH\. Notably, although our method is trained exclusively on BrowseComp\-style questions, it generalizes effectively to xbench and GAIA, indicating strong cross\-benchmark generalization\. These results demonstrate that through answer\-backtracked step\-level credit assignment—which rewards useful actions even in failed trajectories and penalizes redundant or erroneous behaviors in successful ones—the model learns to search more efficiently and more deliberately, validating the effectiveness of our approach\.

Reward Distribution Analysis\.We further analyze the step\-reward distribution over the 8\.5K trajectories used forABC\-SFT\. As shown in Figure[4](https://arxiv.org/html/2608.05102#S4.F4), even successful trajectories contain approximately 4% low\-quality steps with rewards below1\.01\.0\. More importantly, nearly 10% of the steps in failed trajectories receive rewards above1\.01\.0, indicating that they still discover or verify useful clues despite ultimately producing an incorrect answer\. Trajectory\-level supervision would assign the same outcome signal to all of these steps, thereby reinforcing erroneous actions in successful trajectories while penalizing useful actions in failed ones\. In contrast,Answer\-Backtracked Credit Assignmentevaluates each step independently, preserving productive exploration and suppressing incorrect decisions\. This precise supervision accounts for the consistent gains over the corresponding baselines\. See Section[4\.3](https://arxiv.org/html/2608.05102#S4.SS3)for a detailed ablation analysis\.

![Refer to caption](https://arxiv.org/html/2608.05102v1/Workspace/reward.png)Figure 4:Distribution of step rewards across the 8\.5K SFT trajectories\. The left half shows successful trajectories, while the right half shows failed trajectories\. Each bar represents one\-tenth of the steps in a trajectory, ordered from earliest to latest; for example, 0–10% represents the first 10% of steps in a trajectory\. Red, gray, and green denote low\-quality steps withrt<1\.0r\_\{t\}<1\.0, neutral steps withrt=1\.0r\_\{t\}=1\.0, and high\-quality steps withrt\>1\.0r\_\{t\}\>1\.0, respectively\.RL Training Dynamics\.To better understand the source of these gains, we compareABC\-GRPOwith standard trajectory\-level GRPO during training, both evaluated on a validation set of 200 randomly sampled BrowseComp questions\. As shown in Figure[5](https://arxiv.org/html/2608.05102#S4.F5),ABC\-GRPOachieves consistently stronger BrowseComp performance after training begins while producing longer search trajectories\. Together, these results show that step\-level credit assignment improves both search accuracy and exploratory behavior\.

Effect of Context Management\.Following MiroThinker\(Teamet al\.,[2026a](https://arxiv.org/html/2608.05102#bib.bib3)\)and LongSeeker\(Luet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib26)\), we set the maximum context length to 256K tokens and apply the discard\-all strategy for up to five rounds on BrowseComp and BrowseComp\-ZH\. As shown in Figure[6](https://arxiv.org/html/2608.05102#S4.F6), ABSeeker improves from 37\.3% to 55\.3% on BrowseComp and from 39\.1% to 52\.9% on BrowseComp\-ZH\.

![[Uncaptioned image]](https://arxiv.org/html/2608.05102v1/Workspace/rl_step_vs_bc_acc.png)

![[Uncaptioned image]](https://arxiv.org/html/2608.05102v1/Workspace/rl_step_vs_num_turn.png)

Figure 5:RL training dynamics\.Left: Performance during training\.Right: Average number of interaction turns\.

![Refer to caption](https://arxiv.org/html/2608.05102v1/Workspace/context.png)
Figure 6:Performance under different context budgets\.

### 4\.3Ablation Studies

Table 3:Ablation study of ABC\-SFT and ABC\-GRPO\. Results are evaluated without context management\.Table[3](https://arxiv.org/html/2608.05102#S4.T3)evaluates ABC\-SFT and ABC\-GRPO across all five benchmarks, with all methods tested without context management\. Compared with standard SFT, ABC\-SFT improves performance on BrowseComp, BrowseComp\-ZH, xbench\-2510, and GAIA\-text, while remaining comparable on xbench\-2505\. Building on this initialization, ABC\-GRPO consistently outperforms standard trajectory\-level GRPO across all benchmarks\. These results demonstrate that fine\-grained step\-level credit assignment improves both SFT and RL by enabling the model to emphasize useful actions and suppress erroneous ones throughout training\.

## 5Conclusion

We presentAnswer\-Backtracked Credit Assignment \(ABC\), a fine\-grained credit assignment framework for training long\-horizon search agents\. Instead of treating all steps within a trajectory uniformly, ABC recovers intermediate evidence clues from verified answers and uses them to evaluate each search step\. Based on the resulting clue\-anchored rewards, we developABC\-SFT, which reweights the loss of each turn, andABC\-GRPO, which uses step\-level scores as rewards in GRPO\. These methods reward useful actions in failed trajectories while suppressing erroneous or redundant actions in successful ones\. Building on this framework, we trainABSeekerbased on Qwen3\.5\-4B\. Experiments across BrowseComp, BrowseComp\-ZH, xbench, and GAIA\-text show that ABSeeker outperforms same\-scale baselines and remains competitive with substantially larger search agents\. Further analyses demonstrate that answer\-backtracked step\-level supervision improves reward quality, training dynamics, and long\-horizon exploration, highlighting the importance of explicit process supervision for scalable search\-agent training\.

Future Work\.Due to computational constraints, our experiments focus on a compact 4B model\. A natural next step is to scale ABSeeker to larger backbone models and examine whether answer\-backtracked credit assignment brings stronger gains under higher model capacity\. Beyond web search, we also plan to extend this framework to other long\-horizon agent tasks where final outcomes can be backtracked into intermediate evidence, subgoals, or decision points to provide fine\-grained process supervision\.

## References

- Seed2\.0 model card: towards intelligence frontier for real\-world complexity\.arXiv preprint arXiv:2607\.00248\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2607.00248)Cited by:[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- H\. Chen, X\. Cong, S\. Fan, Y\. Fu, Z\. Gong, Y\. Lu, Y\. Li, B\. Niu, C\. Pan, Z\. Song,et al\.\(2026\)AgentCPM\-explore: realizing long\-horizon deep exploration for edge\-scale agents\.arXiv preprint arXiv:2602\.06485\.Cited by:[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- K\. Chen, Y\. Ren, Y\. Liu, X\. Hu, H\. Tian, T\. Xie, F\. Liu, H\. Zhang, H\. Liu, Y\. Gong,et al\.\(2025a\)Xbench: tracking agents productivity scaling with profession\-aligned real\-world evaluations\.arXiv preprint arXiv:2506\.13651\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p2.1)\.
- M\. Chen, L\. Sun, T\. Li, H\. Sun, Y\. Zhou, C\. Zhu, H\. Wang, J\. Pan, W\. Zhang, H\. Chen,et al\.\(2025b\)ReSearch: learning to reason with search for llms via reinforcement learning\. arxiv 2025\.arXiv preprint arXiv:2503\.19470\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p2.1)\.
- Z\. Chu, X\. Wang, J\. Hong, H\. Fan, Y\. Huang, Y\. Yang, G\. Xu, C\. Zhao, C\. Xiang, S\. Hu,et al\.\(2026\)Redsearcher: a scalable and cost\-efficient framework for long\-horizon search agents\.arXiv preprint arXiv:2602\.14234\.Cited by:[§2](https://arxiv.org/html/2608.05102#S2.p1.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- Y\. Du, R\. Ye, S\. Tang, X\. Zhu, Y\. Lu, Y\. Cai, and S\. Chen \(2026\)Openseeker: democratizing frontier search agents by fully open\-sourcing training data\.arXiv preprint arXiv:2603\.15594\.Cited by:[Appendix A](https://arxiv.org/html/2608.05102#A1.p1.7),[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§2](https://arxiv.org/html/2608.05102#S2.p1.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- J\. Gao, W\. Fu, M\. Xie, S\. Xu, C\. He, Z\. Mei, B\. Zhu, and Y\. Wu \(2025\)Beyond ten turns: unlocking long\-horizon agentic search with large\-scale asynchronous rl\.arXiv preprint arXiv:2508\.07976\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p2.1)\.
- Google DeepMind \(2026\)Gemini 3\.1 Pro model card\.Note:[https://deepmind\.google/models/model\-cards/gemini\-3\-1\-pro/](https://deepmind.google/models/model-cards/gemini-3-1-pro/)Cited by:[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- B\. Jin, H\. Zeng, Z\. Yue, J\. Yoon, S\. Arik, D\. Wang, H\. Zamani, and J\. Han \(2025\)Search\-r1: training llms to reason and leverage search engines with reinforcement learning\.arXiv preprint arXiv:2503\.09516\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p2.1)\.
- M\. Li, Q\. Zeng, T\. Fang, Z\. Liang, L\. Song, Q\. Liu, H\. Mi, and D\. Yu \(2026\)Verified critical step optimization for llm agents\.arXiv preprint arXiv:2602\.03412\.Cited by:[§2](https://arxiv.org/html/2608.05102#S2.p2.1)\.
- Y\. Liu, Y\. Feng, L\. Qin, J\. Chen, J\. Yu, S\. Gao, S\. Yang, and W\. Xu \(2026\)Beyond trajectory rewards: step\-level credit assignment for agentic search via graph modeling\.arXiv preprint arXiv:2605\.29697\.Cited by:[§2](https://arxiv.org/html/2608.05102#S2.p2.1)\.
- Y\. Lu, R\. Ye, Y\. Du, J\. Wang, S\. Liu, and S\. Chen \(2026\)LongSeeker: elastic context orchestration for long\-horizon search agents\.arXiv preprint arXiv:2605\.05191\.Cited by:[§4\.2](https://arxiv.org/html/2608.05102#S4.SS2.p4.1)\.
- G\. Mialon, C\. Fourrier, T\. Wolf, Y\. LeCun, and T\. Scialom \(2024\)Gaia: a benchmark for general ai assistants\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 9025–9049\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p2.1)\.
- OpenAI \(2025\)Deep research system card\.Note:[https://cdn\.openai\.com/deep\-research\-system\-card\.pdf](https://cdn.openai.com/deep-research-system-card.pdf)February 25, 2025Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p1.1),[§2](https://arxiv.org/html/2608.05102#S2.p1.1)\.
- Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. Li, Y\. Wu,et al\.\(2024\)Deepseekmath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Cited by:[§3\.4\.2](https://arxiv.org/html/2608.05102#S3.SS4.SSS2.p1.8)\.
- A\. Singh, A\. Fry, A\. Perelman, A\. Tart, A\. Ganesh, A\. El\-Kishky, A\. McLaughlin, A\. Low, A\. Ostrow, A\. Ananthram,et al\.\(2025\)Openai gpt\-5 system card\.arXiv preprint arXiv:2601\.03267\.Cited by:[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- H\. Song, J\. Jiang, Y\. Min, J\. Chen, Z\. Chen, W\. X\. Zhao, L\. Fang, and J\. Wen \(2025\)R1\-searcher: incentivizing the search capability in llms via reinforcement learning\.arXiv preprint arXiv:2503\.05592\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p2.1)\.
- Q\. Tang, H\. Xiang, L\. Yu, B\. Yu, Y\. Lu, X\. Han, L\. Sun, W\. Zhang, P\. Wang, S\. Liu,et al\.\(2025\)Beyond turn limits: training deep search agents with dynamic context window\.arXiv preprint arXiv:2510\.08276\.Cited by:[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- M\. Team and L\. A\. Inc \(2026\)Mind deepresearch technical report\.arXiv preprint arXiv:2604\.14518\.Cited by:[§2](https://arxiv.org/html/2608.05102#S2.p2.1)\.
- M\. Team, S\. Bai, L\. Bing, L\. Lei, R\. Li, X\. Li, X\. Lin, E\. Min, L\. Su, B\. Wang,et al\.\(2026a\)Mirothinker\-1\.7 & h1: towards heavy\-duty research agents via verification\.arXiv preprint arXiv:2603\.15726\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p1.1),[§2](https://arxiv.org/html/2608.05102#S2.p1.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1),[§4\.2](https://arxiv.org/html/2608.05102#S4.SS2.p4.1)\.
- T\. D\. Team, B\. Li, B\. Zhang, D\. Zhang, F\. Huang, G\. Li, G\. Chen, H\. Yin, J\. Wu, J\. Zhou,et al\.\(2025\)Tongyi deepresearch technical report\.arXiv preprint arXiv:2510\.24701\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p1.1),[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§2](https://arxiv.org/html/2608.05102#S2.p1.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- V\. Team, S\. Dai, Y\. Deng, J\. Lin, Y\. Song, G\. Wang, X\. Wu, Y\. Zhou, S\. Yang, Z\. Ying,et al\.\(2026b\)DR\-venus: towards frontier edge\-scale deep research agents with only 10k open data\.arXiv preprint arXiv:2604\.19859\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- G\. Wang, S\. Dai, G\. Ye, Z\. Gan, W\. Yao, Y\. Deng, X\. Wu, and Z\. Ying \(2026\)Information gain\-based policy optimization: a simple and effective approach for multi\-turn search agents\.InThe Fourteenth International Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2608.05102#S2.p2.1)\.
- J\. Wei, Z\. Sun, S\. Papay, S\. McKinney, J\. Han, I\. Fulford, H\. W\. Chung, A\. T\. Passos, W\. Fedus, and A\. Glaese \(2025\)Browsecomp: a simple yet challenging benchmark for browsing agents\.arXiv preprint arXiv:2504\.12516\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§3\.2](https://arxiv.org/html/2608.05102#S3.SS2.p2.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p2.1)\.
- J\. Xie, T\. Lin, Z\. Wang, Y\. Ning, Y\. Yao, T\. Xue, Z\. Zhang, Z\. Li, K\. Zhang, Y\. Wu,et al\.\(2026\)QUEST: training frontier deep research agents with fully synthetic tasks\.arXiv preprint arXiv:2605\.24218\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p2.1),[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- A\. Xu, B\. Lin, B\. Xue, B\. Wang, B\. Xu, B\. Wu, B\. Zhang, C\. Lin, C\. Dong, C\. Ling,et al\.\(2026\)Deepseek\-v4: towards highly efficient million\-token context intelligence\.arXiv preprint arXiv:2606\.19348\.Cited by:[Appendix B](https://arxiv.org/html/2608.05102#A2.p1.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p1.1)\.
- S\. Yao, J\. Zhao, D\. Yu, I\. Shafran, K\. R\. Narasimhan, and Y\. Cao \(2022\)React: synergizing reasoning and acting in language models\.InNeurIPS 2022 Foundation Models for Decision Making Workshop,Cited by:[§2](https://arxiv.org/html/2608.05102#S2.p1.1)\.
- A\. Zeng, X\. Lv, Z\. Hou, Z\. Du, Q\. Zheng, B\. Chen, D\. Yin, C\. Ge, C\. Huang, C\. Xie,et al\.\(2026\)Glm\-5: from vibe coding to agentic engineering\.arXiv preprint arXiv:2602\.15763\.Cited by:[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p3.1)\.
- Y\. Zheng, D\. Fu, X\. Hu, X\. Cai, L\. Ye, P\. Lu, and P\. Liu \(2025\)Deepresearcher: scaling deep research via reinforcement learning in real\-world environments\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 414–431\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p2.1),[§2](https://arxiv.org/html/2608.05102#S2.p1.1)\.
- P\. Zhou, B\. Leon, X\. Ying, C\. Zhang, Y\. Shao, Q\. Ye, D\. Chong, Z\. Jin, C\. Xie, M\. Cao,et al\.\(2025\)Browsecomp\-zh: benchmarking web browsing ability of large language models in chinese\.arXiv preprint arXiv:2504\.19314\.Cited by:[§1](https://arxiv.org/html/2608.05102#S1.p5.1),[§4\.1](https://arxiv.org/html/2608.05102#S4.SS1.p2.1)\.

## Appendix ATraining Details

Supervised Fine\-Tuning\.We implement SFT using Slime and initialize the model from Qwen3\.5\-4B\. We randomly select 8\.5K trajectories from OpenSeeker\(Duet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib18)\), consisting of 5\.5K correct and 3\.0K incorrect trajectories\. We train on these 8\.5K scored trajectories for 3 epochs with a global batch size of 64\. The optimizer is Adam with a learning rate of5×10−55\\times 10^\{\-5\}, cosine decay, a warmup ratio of0\.10\.1, a minimum learning rate of1×10−61\\times 10^\{\-6\}, and weight decay of0\.10\.1\. We use tensor model parallel size 2 and context parallel size 8 for long\-context training\. Tool responses are masked from the loss\. For ABC\-SFT, we map each step reward to a loss weight byw​\(rt\)=2​σ​\(2​\(rt−1\)\)w\(r\_\{t\}\)=2\\sigma\(2\(r\_\{t\}\-1\)\), where the neutral rewardrt=1\.0r\_\{t\}=1\.0is mapped to weight1\.01\.0\.

Reinforcement Learning\.We implement RL using veRL and initialize from the SFT checkpoint\. We train on 1,000 questions filtered by the number of interaction turns, of which 200 have fewer than 100 turns and the remaining 800 have at least 100 turns\. We sample 8 rollouts per question with a training batch size of 16\. Each rollout allows up to 200 interaction turns\. We use asynchronous rollout with 16 agent\-loop workers\. The rollout policy uses temperature1\.01\.0, top\-p=0\.95p=0\.95\. We optimize with ABC\-GRPO using clue\-anchored rewards plus a format penalty\. Rewards are normalized within each rollout group, and discounted step\-level advantages are computed withγ=0\.25\\gamma=0\.25\. The actor learning rate is1×10−61\\times 10^\{\-6\}and the KL loss coefficient is0\.0010\.001\.

## Appendix BEvaluation Details

Due to computational constraints, for the RL validation set, we sample 200 BrowseComp questions from the full set using random seed 42\. To reduce evaluation variance, we run each evaluation three times and report the averaged score\. All evaluations allow up to 200 interaction turns and use DeepSeek\-V4\-Flash\(Xuet al\.,[2026](https://arxiv.org/html/2608.05102#bib.bib22)\)as the evaluation model with the prompt in Appendix[C\.3](https://arxiv.org/html/2608.05102#A3.SS3)to judge final\-answer correctness\.

## Appendix CPrompt Templates

### C\.1Answer\-Backtracked Clue Recovery

This prompt recovers intermediate clues from the query and verified answer\.

\{promptbox\}

Prompt for Answer\-Backtracked Clue Recovery

System:

You are an expert investigator\. Use search\_web and visit\_web to gather evidence\. Do NOT provide the final answer directly\. Extract intermediate clues, such as dates, places, and events, from the question, search for them first, then piece together to verify the answer\.

User:

Reconstruct cite\-backed steps that prove the answer is correct\.

Question: \{query\}

Known Answer: \{answer\}

Your task: Find the intermediate clues that lead to the correct answer\. These clues are usually conditions in the question, such as events, dates, places, and entities\. Search for them first, then verify the answer\. You MUST explicitly mention these intermediate clues in your final answer and explain how they support the answer\.

### C\.2Clue\-Anchored Step Scoring

This prompt scores each trajectory step using the recovered clue set\.

\{promptbox\}

Prompt for Clue\-Anchored Step Scoring

System:

You are a trajectory scorer\. Your task is to evaluate each step in an AI agent’s problem\-solving trajectory based on the scoring rubric\.

Focus on INTERMEDIATE STEPS and KEY ENTITIES\. Base your scoring ONLY on the step’s content, including Context, Reasoning, and Tool Call\.

User:

Scoring Rubric:

Criteria 2–4 apply only to clues listed in “Core Clues” below\. Each step starts from base score 1\.0\. Multiple criteria may apply in one step\.

1\.Baseline \(1\.0\):Reasonable progress without an obvious error gets the base score 1\.0\.

2\.Finds or verifies a correct clue \(\+0\.8\):The step discovers, investigates, or verifies a clue that matches the core clue set\.

3\.Correctly rules out a wrong candidate \(\+0\.4\):The step correctly excludes, dismisses, or identifies as wrong an irrelevant or incorrect candidate\.

4\.Incorrectly dismisses a correct clue \(\-0\.8\):The step wrongly excludes or dismisses a clue that is correct according to the core clues\. Exploring wrong candidates does NOT count as this\.

5\.Correct final answer \(\+1\.0\):Only when the step explicitly gives the final answer and it matches the Correct Answer below\.

6\.Wrong final answer \(\-1\.0\):Only when the step explicitly submits an incorrect final answer\. If this applies, do NOT apply criterion 5\.

Question:

\{query\}

Correct Answer:

\{correct\_answer\}

Core Clues:

\{core\_clues\_text\}

Step to Score:

Step Number: \{step\_num\}

Context: \{user\_content\}

Reasoning: \{reasoning\}

Tool Call: \{action\}

Tool Response: \{response\}

Task:

Score this step based on the rubric above\.

Return ONLY a JSON object with this exact structure:

\{ “score”:<<float between 0\.0 and 2\.0; start from 1\.0, add/subtract applied deltas, then clip to \[0, 2\.0\]\>\>,

“criteria”:<<array of int 1–6, the rubric criterion numbers that apply, e\.g\., \[1\] or \[1, 2, 5\]\>\>,

“explanation”:<<brief explanation\. When applying criteria, MUST name the specific clue\(s\)/entity, e\.g\., clue X or entity Y\>\>\}

Important: Return ONLY valid JSON, no other text\.

### C\.3Evaluation Prompt

This prompt checks whether the predicted answer matches the ground truth\.

\{promptbox\}

Prompt for Final Answer Evaluation

System:

You are an evaluator\. Your ONLY task is to decide: does the Predicted Answer refer to the SAME entity or value as the Ground Truth?

User:

Question: \{query\}

Ground Truth: \{gt\}

Predicted Answer: \{pred\}

Output only a JSON object with this format:

\{“correct”: true/false, “reason”: “brief explanation”\}