Tree-of-Experience: A Structured Experience-Management Solution for Self-Evolving Agents under Low-Repetition and Implicit-Reward Environments

# Tree-of-Experience: A Structured Experience-Management Solution for Self-Evolving Agents under Low-Repetition and Implicit-Reward Environments
Source: [https://arxiv.org/html/2606.06960](https://arxiv.org/html/2606.06960)
Zihao Deng, Yining Zhu, Leiming Wang, Jingfei Lu Junbo Wang, Chuncheng Ran, Yu Yang, Dixuan Yang, Jikun Shen

###### Abstract

Experience\-based self\-evolution is crucial for LLM agents, but existing benchmarks often assume explicit goals, stable task patterns, and clear feedback\. We study a more challenging setting: low\-repetition tasks with implicit rewards, where past experience is difficult to reuse and feedback is delayed, noisy, and outcome\-level\. We introduceFinEvolveBench, a temporally controlled benchmark for financial sentiment prediction that links daily news\-driven predictions to future excess returns\. We further propose Tree\-of\-Experience \(ToE\), a structured experience\-management method that organizes, retrieves, validates, and updates agent experience\. Experiments show that general\-purpose experience mechanisms do not consistently outperform no\-experience baselines, while ToE achieves stronger overall performance\. These results highlight the importance of structured experience management for self\-evolving agents in implicit\-reward environments\.

Tree\-of\-Experience: A Structured Experience\-Management Solution for Self\-Evolving Agents under Low\-Repetition and Implicit\-Reward Environments

Zihao Deng, Yining Zhu, Leiming Wang, Jingfei LuJunbo Wang, Chuncheng Ran, Yu Yang, Dixuan Yang, Jikun Shen

## 1Introduction

Evaluating whether language\-model agents can improve through experience requires environments where past interactions can inform future decisions\. Existing self\-evolution benchmarks have made important progress in testing experience accumulation, reflection, and strategy revisionShinnet al\.\([2023](https://arxiv.org/html/2606.06960#bib.bib21)\); Zhenget al\.\([2025](https://arxiv.org/html/2606.06960#bib.bib23)\)\. However, many of them are built around procedural tasks with relatively stable execution patterns, explicit goals, and verifiable outcomes, such as software engineering, web navigation, tool use, or embodied instruction followingJimenezet al\.\([2024](https://arxiv.org/html/2606.06960#bib.bib9)\); Mialonet al\.\([2023](https://arxiv.org/html/2606.06960#bib.bib11)\); Shridharet al\.\([2020](https://arxiv.org/html/2606.06960#bib.bib3)\)\. These settings are useful for studying whether agents can reuse successful trajectories, but they only partially reflect real\-world decision problems where task instances are weakly repetitive, feedback is indirect, and environmental regularities change over time\.

We focus on this under\-evaluated setting, where agents must adapt across complex and heterogeneous tasks while interacting with environments that provide indirect and unstable feedback, i\.e\.,low\-repetition tasks with implicit rewards\. In such tasks, historical experience cannot be reused directly, because each instance may involve different information sources, latent causal factors, and decision rationales\. Feedback is typically observed only at the outcome level, rather than as step\-wise supervision for individual reasoning or information\-selection decisions\. This creates a challenging credit\-assignment problem: success or failure cannot be reliably traced back to a specific factor, reasoning step, or piece of evidence\. For example, financial markets are exposed to numerous positive and negative news signals on each trading day, making it difficult to determine whether a correct or incorrect market sentiment prediction results from any particular analysis\. Moreover, the environment may be noisy and non\-stationary, making previously useful experience unreliable or even harmful\. These properties challenge a common assumption behind many experience\-based agent methods: that semantically similar past tasks provide directly reusable guidance for future tasks\.

Financial sentiment prediction provides a natural testbed for this setting\. Given temporally ordered market news, an agent must infer sentiment signals that are predictive of future market movement\. Unlike standard sentiment classification, prediction quality is not supervised by explicit labels attached to individual news items, but is reflected indirectly through subsequent market outcomes\. The reward is therefore delayed, noisy, and implicit\. Meanwhile, the task is inherently low\-repetition: even when news items share similar entities or topics, the surrounding market context, investor expectations, macroeconomic conditions, and firm\-specific states may differ substantially\. Thus, historical experience cannot be reused based on surface similarity alone\. An agent must determine whether a past reasoning pattern remains valid under the current market regime, and decide what experience to retrieve, revise, or discard\. This makes financial sentiment prediction a suitable benchmark for testing whether agents can manage experience and adapt under low repeatability, implicit rewards, and non\-stationary environments\.

Thus, we introduceFinEvolveBench, a benchmark for evaluating experience\-based self\-evolution in financial sentiment prediction and propose Tree\-of\-Experience\(ToE\), a structured experience\-management method as a reference solution for low\-repetition and implicit\-reward environments\. We conduct a comprehensive evaluation across multiple foundation models and self\-evolution strategies, including a no\-experience pipeline, general\-purpose experience mechanisms, and the proposed structured method\. The results show that general\-purpose experience mechanisms do not consistently improve over the no\-experience setting, suggesting that experience reuse in low\-repetition financial environments is non\-trivial\. In contrast, structured experience management provides stronger overall performance under the current protocol\.

Our contributions are summarized as follows: \(1\) we formulate low\-repetition tasks with implicit rewards as an evaluation setting for self\-evolving agents, highlighting the combined challenges of weak task repeatability, outcome\-level feedback, and non\-stationary; \(2\) we constructFinEvolveBench, a temporally ordered financial sentiment prediction benchmark with strict time control, recent market data, delayed feedback, and financial evaluation metrics based on future excess returns\. \(3\) we propose Tree\-of\-Experience\(ToE\), a structured experience\-management reference method and diagnostic evaluation, showing that effective self\-evolution inFinEvolveBenchrequires agents to selectively retrieve, validate, and update experience rather than reuse historical patterns directly\.

## 2Related Work

A growing body of work has studied runtime self\-evolution for LLM agents, aiming to enable agents to improve after deployment through reflection, memory, and experience reuse\. We first introduce representative self\-evolution methods and the evaluation benchmarks adopted in these studies, and then review existing financial NLP benchmarks and explain why they are insufficient for validating agent self\-evolution ability under low\-repetition tasks with implicit rewards\.

##### Runtime Self\-Evolution and General Benchmarks\.

Runtime learning studies how agents improve post\-deployment through interaction streams rather than offline supervised data, echoing the broader consensus that future agents must learn from grounded experience\(Silver and Sutton,[2025](https://arxiv.org/html/2606.06960#bib.bib25)\)\. External memory offers a highly effective, non\-parametric alternative to model weight updates, bypassing traditional cost and stability bottlenecks\. Systems such as Reflexion\(Shinnet al\.,[2023](https://arxiv.org/html/2606.06960#bib.bib21)\), Mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib22)\), Evo\-Memory\(Weiet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib24)\), and MemRL\(Zhanget al\.,[2026](https://arxiv.org/html/2606.06960#bib.bib26)\)explore how agents can store, retrieve, and reuse past successes and failures to optimize future reasoning\. To evaluate these capabilities, language models are increasingly tested as autonomous agents across various general benchmarks\. Software and tool\-use benchmarks such as SWE\-Bench\(Jimenezet al\.,[2024](https://arxiv.org/html/2606.06960#bib.bib9)\), BigCodeBench\(Zhuoet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib4)\), GAIA\(Mialonet al\.,[2023](https://arxiv.org/html/2606.06960#bib.bib11)\), XBench\(Chen and others,[2025](https://arxiv.org/html/2606.06960#bib.bib12)\), andτ2\\tau^\{2\}\-Bench\(Barres and others,[2025](https://arxiv.org/html/2606.06960#bib.bib13)\)measure whether agents can solve realistic tasks through reasoning, tool invocation, and environment interaction\. Benchmarks closer to self\-evolving or long\-horizon agents, including LifelongAgentBench\(Zhenget al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib23)\), HLE\(Phanet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib2)\), and ALFWorld\(Shridharet al\.,[2020](https://arxiv.org/html/2606.06960#bib.bib3)\), further stress lifelong learning, frontier\-level knowledge, and interactive planning in simulated environments\. While these benchmarks are crucial for measuring task completion and experience reuse, they operate in environments that largely assume explicit success signals, stable rules, and clear task recurrence\.

##### Financial NLP and financial benchmarks\.

Existing financial benchmarks, including FinQA\(Chenet al\.,[2021](https://arxiv.org/html/2606.06960#bib.bib17)\), TAT\-QA\(Zhuet al\.,[2021](https://arxiv.org/html/2606.06960#bib.bib18)\), FinBen\(Xie and others,[2024](https://arxiv.org/html/2606.06960#bib.bib8)\), and target\-based sentiment datasets\(Muhammadet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib7)\), mainly provide static, instance\-level supervision for financial understanding and reasoning\. They do not require agents to process time\-indexed information, maintain evolving experience, or adapt to delayed market feedback\. Trading\-oriented frameworks such as StockBench\(Chenet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib19)\)introduce temporal simulation, but mainly assess the profitability of trading actions rather than memory\-driven self\-evolution\. They also often rely on limited retrieved evidence, reducing the breadth and noise of real market information streams\.FinEvolveBenchinstead evaluates agents in a broader market\-wide information environment with dense, noisy, and weakly structured public opinion signals\.

## 3FinEvolveBench

In this section, we presentFinEvolveBench, a benchmark consisting of two major components: a temporally interactive financial market environment and a well\-defined self\-evolving agent workflow\. Together, these components provide a flexible testbed where users can conveniently plug in and evaluate different experience\-based self\-evolution methods\. The overall framework ofFinEvolveBenchis illustrated in Figure[1](https://arxiv.org/html/2606.06960#S3.F1)\.

![Refer to caption](https://arxiv.org/html/2606.06960v1/x1.png)Figure 1:Overview ofFinEvolveBench\.FinEvolveBenchconsists of a market environment dataset with implicit rewards and a workflow that converts a backbone LLM into an agent equipped with a general self\-evolution mechanism\.### 3\.1Environment Construction

We constructFinEvolveBenchfrom 31 Shenwan first\-level industry indices in the Chinese A\-share market, time\-stamped public financial news, and daily market records\. For each trading day, we align the news stream with open/close prices, trading volume, and transaction data, and convert the aligned records into chronologically ordered sentiment prediction instances\. The environment is temporally interactive: the agent can access only information available up to the current date, while each prediction is logged and converted into delayed market feedback once its horizon becomes observable\. We include details of the financial market environment and data structures in Appendix\.

### 3\.2Self\-Evolving Agent Workflow

On top of the interactive environment inFinEvolveBench, we design a lightweight self\-evolving agent workflow\. Inspired by mainstream memory and experience\-based self\-evolution mechanisms such as Mem0Chhikaraet al\.\([2025](https://arxiv.org/html/2606.06960#bib.bib22)\)and MemRLZhanget al\.\([2026](https://arxiv.org/html/2606.06960#bib.bib26)\), we decompose experience management into two stages: retrieval and updating\. This abstraction allows researchers to evaluate their own experience\-based self\-evolution methods under a unified benchmark interface\.

As shown in the right part of Fig\.[1](https://arxiv.org/html/2606.06960#S3.F1), the workflow contains four main steps\.

##### Step 1: Information Perception and Task Formulation\.

The agent first perceives the available environmental information and extracts task\-relevant content, thereby formulating the task required for the current interaction\. Specifically, given the industry associated with the target to be analyzed, the agent retrieves and extracts important news relevant to that industry from the available information sources, thereby formulating a concrete financial sentiment analysis task\.

##### Step 2: Task\-Relevant Experience Retrieval\.

The experience management agent retrieves from the experience bank a set of experiences that may be useful for completing the current task\.

##### Step 3: Task Execution\.

Given the task description and the retrieved experiences, the agent executes the task\. Following prior work on financial sentiment analysis such as FinBERTAraci \([2019](https://arxiv.org/html/2606.06960#bib.bib27)\)and FinGPTYanget al\.\([2023](https://arxiv.org/html/2606.06960#bib.bib30)\), the agent generates a continuous sentiment score in the range of\[−1,1\]\[\-1,1\]for the target under analysis\.

##### Step 4: Validation and Reflection\.

After task execution, the environment records the prediction trajectory and computes feedback once the corresponding market outcome becomes observable\. The feedback\-augmented trajectory is then used to update the experience pool, and the simulation advances to the next trading day\.

This workflow offers a unified interface for evaluating experience\-based self\-evolution methods without prescribing how experiences should be organized, retrieved, or updated\. In the following section, we introduce our structured experience\-management self\-evolve solution for low\-repetition tasks with implicit rewards\.

## 4Methodology

In this section, we introduce Tree\-of\-Experience: a structured experience\-management method and show howFinEvolveBenchcan compare and diagnose different experience mechanisms under low\-repetition and implicit\-reward environments\.

### 4\.1Problem Formulation

Following the self\-evolving agent workflow defined in Sec\.[3](https://arxiv.org/html/2606.06960#S3), we formulate experience adoption as a non\-parametric reinforcement learning process\. At each time steptt, the agent observes a statests\_\{t\}, retrieves external experiencesℰt\\mathcal\{E\}\_\{t\}, and produces an actionata\_\{t\}\. After the corresponding outcomeoto\_\{t\}becomes observable, the environment returns a delayed feedback signal

rt=R​\(at,ot\),r\_\{t\}=R\(a\_\{t\},o\_\{t\}\),\(1\)whereR​\(⋅\)R\(\\cdot\)is a task\-specific reward function\. The feedback is used to update the utility of retrieved experiences without modifying model parameters\. The overall framework is illustrated in Fig\.[2](https://arxiv.org/html/2606.06960#S4.F2)\.

![Refer to caption](https://arxiv.org/html/2606.06960v1/x2.png)Figure 2:Overview ofTree\-of\-Experience\. The upper panel shows the end\-to\-end learning workflow: given a taskτt\\tau\_\{t\}, the agent retrieves an experience contextmctxm\_\{\\mathrm\{ctx\}\}from the experience memoryℰ\\mathcal\{E\}, generates the outputyty\_\{t\}, and updates the utility valuesQQof the retrieved experiences according to the environment rewardRR\. The lower\-left panel illustrates the two\-stage experience query process, where a predefined experience selection algorithm first recalls candidate experiences, and an LLM\-as\-judge module then determines whether the recalled experiences are usable; usable experiences are reranked, while unusable cases trigger the creation of new experiences\. The lower\-right panel shows the utility update process, where environment feedback is used to update experience utilitiesQQ\. The right panel visualizes the tree\-structured organization of the experience memory\.
### 4\.2Tree\-of\-Experience Self\-Evolution Framework

we propose an experience\-centric self\-evolution framework that optimizes the selection and organization of external experience without updating the frozen LLM parameters\. The framework consists of four components: \(i\) prior experience induction, \(ii\) depth\-constrained representation, \(iii\) hierarchical selection with adaptive expansion, and \(iv\) runtime utility estimation\.

#### 4\.2\.1Experience Representation

We organize the external experience poolℰ\\mathcal\{E\}as a depth\-constrained, width\-expandable tree, where each root\-to\-leaf path defines an executable analytical perspective:

ℰ=\{\(𝐏i,𝐐i,𝐌i\)\}i=1\|ℰ\|\.\\mathcal\{E\}=\\left\\\{\\Big\(\\mathbf\{P\}\_\{i\},\\mathbf\{Q\}\_\{i\},\\mathbf\{M\}\_\{i\}\\Big\)\\right\\\}\_\{i=1\}^\{\|\\mathcal\{E\}\|\}\.\(2\)
Each tuple contains three components\.𝐏i=\(pi\(1\),…,pi\(L\)\)\\mathbf\{P\}\_\{i\}=\(p\_\{i\}^\{\(1\)\},\\dots,p\_\{i\}^\{\(L\)\}\)is a depth\-LLexperience path, whose upper levels encode abstract task patterns and lower levels encode concrete reasoning principles\. New experiences follow the same interface by adding leaf nodes\.𝐐i\\mathbf\{Q\}\_\{i\}stores experience\-specific utility estimates for the path\. In our settings, it illustrates how important or correct is this experience’s analytical perspective\.𝐌i\\mathbf\{M\}\_\{i\}contains non\-semantic metadata, such as recall/hit counts and identifiers, used only for bookkeeping and utility updates\. We provide an example experience tree in the Appendix to illustrate the hierarchical organization of experiences\. High\-level factors, such as “policy planning”, are used as root nodes, while more specific analytical directions, such as “supply\-side administrative constraints”, are organized as lower\-level nodes\.

#### 4\.2\.2Construction of Experience Tree

In this subsection, we introduce how we construct our ToE automatically from historical tasks that strictly precede the test window, ensuring no leakage from test\-period news, returns, or feedback\.

For each historical taskτk\\tau\_\{k\}, the induction operator rewrites its reasoning trajectoryTkT\_\{k\}into a compact logical skeleton:

T~k=ℱrewrite​\(Tk\),∀τk∈𝒟\.\\tilde\{T\}\_\{k\}=\\mathcal\{F\}\_\{\\text\{rewrite\}\}\(T\_\{k\}\),\\quad\\forall\\tau\_\{k\}\\in\\mathcal\{D\}\.\(3\)Given the skeleton set𝒯~\\tilde\{\\mathcal\{T\}\}, the system builds a prior paradigm tree\. The top\-level paradigm is extracted globally, and lower\-level paradigms are induced conditioned on their ancestor path:

p\(1\)∼pextract\(⋅∣𝒯~\),p^\{\(1\)\}\\sim p\_\{\\text\{extract\}\}\\left\(\\cdot\\mid\\tilde\{\\mathcal\{T\}\}\\right\),\(4\)p\(l\)←gind\(l\)​\(p\(1\),…,p\(l−1\),𝒯~sub\),l∈\{2,…,L\}\.p^\{\(l\)\}\\leftarrow g\_\{\\text\{ind\}\}^\{\(l\)\}\\bigl\(p^\{\(1\)\},\\dots,p^\{\(l\-1\)\},\\tilde\{\\mathcal\{T\}\}\_\{\\text\{sub\}\}\\bigr\),\\\\ l\\in\\\{2,\\dots,L\\\}\.\(5\)

#### 4\.2\.3Hierarchical Experience Selection and Adaptive Expansion

Given a task statess, the experience policyμ​\(𝐏∣s,ℰ\)\\mu\(\\mathbf\{P\}\\mid s,\\mathcal\{E\}\)selects an experience path through a cascaded retrieval operator\. For each layerll, we define a matching function

fmatch\(l\):𝒮×𝒫\(l\)→\[0,1\],f\_\{\\text\{match\}\}^\{\(l\)\}:\\mathcal\{S\}\\times\\mathcal\{P\}^\{\(l\)\}\\rightarrow\[0,1\],\(6\)which measures the compatibility between the current state and a layer\-llexperience node\. For a complete path𝐏=\(p\(1\),…,p\(L\)\)\\mathbf\{P\}=\(p^\{\(1\)\},\\dots,p^\{\(L\)\}\), its contextual compatibility is computed as

ρ​\(s,𝐏\)=\(∏l=1Lfmatch\(l\)​\(s,p\(l\)\)\)1/L\.\\rho\(s,\\mathbf\{P\}\)=\\left\(\\prod\_\{l=1\}^\{L\}f\_\{\\text\{match\}\}^\{\(l\)\}\(s,p^\{\(l\)\}\)\\right\)^\{1/L\}\.\(7\)We combine this compatibility with the task\-specific historical utility:

S​\(s,𝐏\)=ρ​\(s,𝐏\)⋅Uτ​\(𝐐​\(𝐏\)\),S\(s,\\mathbf\{P\}\)=\\rho\(s,\\mathbf\{P\}\)\\cdot U\_\{\\tau\}\(\\mathbf\{Q\}\(\\mathbf\{P\}\)\),\(8\)whereUτ​\(⋅\)U\_\{\\tau\}\(\\cdot\)maps the structured utility vector to the scalar objective required by the current task, e\.g\., the utility dimension of the target prediction horizon\.

Selection proceeds in a coarse\-to\-fine manner\. Starting from𝒞\(0\)=ℰ\\mathcal\{C\}^\{\(0\)\}=\\mathcal\{E\}, each internal layer keeps the top\-klk\_\{l\}candidate paths according to the corresponding layer\-level matching score:

𝒞\(l\)​\(τ\)\\displaystyle\\mathcal\{C\}^\{\(l\)\}\(\\tau\)=TopK𝐏i∈𝒞\(l−1\)​\(τ\)\(\\displaystyle=\\operatorname\{TopK\}\_\{\\mathbf\{P\}\_\{i\}\\in\\mathcal\{C\}^\{\(l\-1\)\}\(\\tau\)\}\\Bigl\(\(9\)fmatch\(l\)\(s,pi\(l\)\),kl\),l=1,…,L−1\.\\displaystyle\\quad f\_\{\\text\{match\}\}^\{\(l\)\}\(s,p\_\{i\}^\{\(l\)\}\),\\,k\_\{l\}\\Bigr\),\\quad l=1,\\dots,L\-1\.Keeping multiple candidates avoids premature commitment to a single high\-level reasoning route while bounding the search space\.

At the leaf level, the system retrieves fine\-grained candidates by semantic similarity:

𝒞ℓ​\(τ\)=\{𝐏i∈𝒞\(L−1\)​\(τ\)∣cos⁡\(e​\(s\),e​\(pi\(L\)\)\)≥λ\}\.\\mathcal\{C\}\_\{\\ell\}\(\\tau\)=\\\{\\mathbf\{P\}\_\{i\}\\in\\mathcal\{C\}^\{\(L\-1\)\}\(\\tau\)\\mid\\cos\(e\(s\),e\(p\_\{i\}^\{\(L\)\}\)\)\\geq\\lambda\\\}\.\(10\)A fixed\-prompt judgegjudgeg\_\{\\text\{judge\}\}then determines whether the recalled leaves sufficiently cover the current analytical principle:

zexpand=gjudge​\(τ,𝒞ℓ​\(τ\)\)∈\{0,1\}\.z\_\{\\text\{expand\}\}=g\_\{\\text\{judge\}\}\(\\tau,\\mathcal\{C\}\_\{\\ell\}\(\\tau\)\)\\in\\\{0,1\\\}\.\(11\)If𝒞ℓ​\(τ\)≠∅\\mathcal\{C\}\_\{\\ell\}\(\\tau\)\\neq\\emptysetandzexpand=0z\_\{\\text\{expand\}\}=0, the agent reuses the path with the largest selection scoreS​\(s,𝐏\)S\(s,\\mathbf\{P\}\)\. Otherwise, it expands the selected internal prefix with a new leaf generated bygexpandg\_\{\\text\{expand\}\}and initializes its utility with𝐐init\\mathbf\{Q\}\_\{\\text\{init\}\}\. Both selection and expansion use only the observable context before prediction; the utility of newly added leaves is updated only after delayed feedback becomes available\.

#### 4\.2\.4Runtime Utility Estimation and Update

The experience module is updated from delayed environmental feedback\. We compare two update strategies: explicit numerical calibration and LLM\-based utility rewriting\.

Formula\-based update\.Each experience path𝐏\\mathbf\{P\}maintains a utility vector𝐐​\(𝐏\)∈\[0,2​q0\]d\\mathbf\{Q\}\(\\mathbf\{P\}\)\\in\[0,2q\_\{0\}\]^\{d\}, whereq0q\_\{0\}is the neutral prior\. The utility controls the activation strength of an experience path rather than rewriting its semantic content\.

After executing the selected path𝐏∗\\mathbf\{P\}^\{\*\}, the system receives delayed outcome\-level feedback\. We convert it into an experience\-level update signal:

𝐫env=ψτ​\(at,ot\+h\),\\mathbf\{r\}\_\{\\text\{env\}\}=\\psi\_\{\\tau\}\(a\_\{t\},o\_\{t\+h\}\),\(12\)whereot\+ho\_\{t\+h\}denotes the delayed task outcome andψτ\\psi\_\{\\tau\}is a task\-specific feedback mapping\.

We update the utility as

𝐐new​\(𝐏∗\)←𝐐old​\(𝐏∗\)\+η​1−\(𝐐old​\(𝐏∗\)−q0q0\)21\+log⁡\(1\+nhit\)⊙𝐫env,\\mathbf\{Q\}\_\{\\text\{new\}\}\(\\mathbf\{P\}^\{\*\}\)\\leftarrow\\mathbf\{Q\}\_\{\\text\{old\}\}\(\\mathbf\{P\}^\{\*\}\)\\\\ \+\\eta\\frac\{1\-\\left\(\\frac\{\\mathbf\{Q\}\_\{\\text\{old\}\}\(\\mathbf\{P\}^\{\*\}\)\-q\_\{0\}\}\{q\_\{0\}\}\\right\)^\{2\}\}\{1\+\\log\(1\+n\_\{\\text\{hit\}\}\)\}\\odot\\mathbf\{r\}\_\{\\text\{env\}\},\(13\)and clip𝐐new​\(𝐏∗\)\\mathbf\{Q\}\_\{\\text\{new\}\}\(\\mathbf\{P\}^\{\*\}\)to\[0,2​q0\]d\[0,2q\_\{0\}\]^\{d\}\. Here,η\\etais the base learning rate andnhitn\_\{\\text\{hit\}\}is the historical usage count\. The denominator decays the update step for frequently used paths, while the boundary\-aware term reduces updates near the utility bounds\. This yields soft forgetting, where ineffective paths are gradually selected less often instead of being explicitly removed\.

LLM\-based update\.As a comparison, we also test an LLM\-based updater:

𝐐new​\(𝐏∗\)←gllm​\(𝐏∗,𝐐old,at,𝐫env\),\\mathbf\{Q\}\_\{\\text\{new\}\}\(\\mathbf\{P\}^\{\*\}\)\\leftarrow g\_\{\\text\{llm\}\}\\bigl\(\\mathbf\{P\}^\{\*\},\\mathbf\{Q\}\_\{\\text\{old\}\},a\_\{t\},\\mathbf\{r\}\_\{\\text\{env\}\}\\bigr\),\(14\)where the frozen LLM directly revises the utility representation while preserving the experience path\. This variant is used to examine whether constrained numerical updates are more stable than direct LLM\-based rewriting under delayed feedback\.

#### 4\.2\.5Instantiation for Financial Sentiment Prediction

InFinEvolveBench, each taskτt\\tau\_\{t\}contains a prediction date, a target industry index, and the pre\-prediction news context\. For horizonhh, the agent predicts a sentiment scorey^t,h\\hat\{y\}\_\{t,h\}\. The task reward is instantiated as

rt,h=y^t,h⋅αt,h,αt,h=closet\+hcloset−1,r\_\{t,h\}=\\hat\{y\}\_\{t,h\}\\cdot\\alpha\_\{t,h\},\\qquad\\alpha\_\{t,h\}=\\frac\{\\mathrm\{close\}\_\{t\+h\}\}\{\\mathrm\{close\}\_\{t\}\}\-1,\(15\)whereαt,h\\alpha\_\{t,h\}denotes the realized horizon\-hhreturn\.

Tree\-of\-Experience is implemented as a two\-level hierarchy, with keyword\-matched financial factor categories and similarity\-retrieved analytical principles using thresholdλ=0\.8\\lambda=0\.8\.

## 5Experiments

### 5\.1Experimental Setup

We briefly introduce our experimental setup and include more details in Appendix\. Our main experiments useDeepSeek\-V4\-Flashas the backbone model, withQwen3\.6\-35B\-A3Bincluded as an additional backbone\-model ablation\. Across all experiments, we set the experience\-tree depth toL=2L=2, the first\-level candidate\-retention size tok1=8k\_\{1\}=8, the leaf\-level retrieval cosine\-similarity threshold toλ=0\.8\\lambda=0\.8, the neutral utility toq0=1q\_\{0\}=1, and the learning rate toη=10\\eta=10\.

##### Compared methods\.

We compareBaseline, an average aggregation of LLM\-assigned news sentiment scores\(Wanget al\.,[2018](https://arxiv.org/html/2606.06960#bib.bib34); Mohanet al\.,[2019](https://arxiv.org/html/2606.06960#bib.bib33); Lopez\-Lira and Tang,[2023](https://arxiv.org/html/2606.06960#bib.bib29); Wang and Ma,[2024](https://arxiv.org/html/2606.06960#bib.bib32)\);Pipe, which predicts from industry\-filtered news;Pipe\+mem0andPipe\+MemRL, which add mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib22)\)and MemRL\(Zhanget al\.,[2026](https://arxiv.org/html/2606.06960#bib.bib26)\)as experience modules; andPipe\+ToE, which uses our tree\-structured experience retrieval with formula\-based utility updating\.

##### Evaluation metrics\.

We use the information coefficient \(IC\), a standard metric for evaluating factor quality in quantitative trading, as the main evaluation metric\. IC is appropriate forFinEvolveBenchbecause the task is not to predict a categorical ground\-truth label for each news item, but to produce a sentiment factor whose values should align with future market movements\. Moreover, we add two IC\-based metrics,Cross\-sectional IC \(csIC\)that measures the correlation between predicted factor values and future returns across different targets on the same trading day andTime\-series IC \(tsIC\)that measures the correlation between predicted factor values and future returns across time for the same trading target\. The detailed definition is in Appendix\.

### 5\.2Performance Comparison

Figure[3](https://arxiv.org/html/2606.06960#S5.F3)shows the results on the 20\-trading\-day prediction horizon\. This experiment compares three factors: basic sentiment scoring \(Baseline\), structured news processing without experience retrieval \(Pipe\), and different experience\-management mechanisms added to the same prediction pipeline\. The additional backbone\-model ablation is also included in the same figure\.

We make the following observations:

- •Structured news processing is an important foundation\.OnDeepSeek\-V4\-Flash, Pipe outperforms Baseline on both tsIC and csIC\. This suggests that structured annotation and filtering before the prediction module are important components for this task\. Therefore, the gains of experience\-augmented methods should be compared primarily against Pipe rather than only against Baseline\.
- •Structured experience further improves time\-series and cross\-sectional predictiveness\.OnDeepSeek\-V4\-Flash, Pipe\+ToE achieves higher tsIC \(0\.0741\) and csIC \(0\.0528\) than Pipe \(0\.0517 and 0\.0408\)\. OnQwen3\.6\-35B\-A3B, the completed 20\-day Pipe\+ToE results reach a similar level, with a tsIC of 0\.0749 and a csIC of 0\.0524\. This indicates that adding structured experience retrieval and formula\-based updating to the same news\-processing pipeline improves both time\-series prediction and same\-day target ranking\.
- •General\-purpose experience systems do not consistently outperform the no\-experience Pipe\.OnDeepSeek\-V4\-Flash, both Pipe\+mem0 and Pipe\+MemRL obtain lower tsIC and csIC than Pipe\. This suggests that directly transferring general semantic retrieval or reinforcement\-learning\-based experience management is not necessarily suitable for low\-repetition financial sentiment analysis with implicit rewards\.

![Refer to caption](https://arxiv.org/html/2606.06960v1/x3.png)Figure 3:Main results on the 20\-trading\-day prediction horizon\. Higher values indicate stronger predictive ability\.
### 5\.3Effect of Prediction Horizon

Table[1](https://arxiv.org/html/2606.06960#S5.T1)reports the tsIC prediction\-horizon ablation across 1\-, 5\-, 10\-, and 20\-trading\-day horizons\. This experiment tests whether experience retrieval provides consistent gains across time scales and backbone models\.

We make the following observations:

- •The benefit of structured experience is horizon\-dependent\.OnDeepSeek\-V4\-Flash, Pipe\+ToE obtains the highest tsIC on the 1\-, 10\-, and 20\-day horizons, but it is lower than Pipe and Pipe\+MemRL on the 5\-day horizon\. This shows that experience retrieval is not monotonically effective across all time scales\. Its effect may depend jointly on the speed of news\-event propagation, delayed market reactions, and retrieval noise\.
- •Structured experience is more beneficial for longer\-horizon prediction\.On the 20\-day horizon withDeepSeek\-V4\-Flash, Pipe\+ToE reaches a tsIC of 0\.0741, higher than Pipe’s 0\.0517\. The completed Pipe\+ToE result onQwen3\.6\-35B\-A3Balso shows the same long\-horizon advantage, with a 20\-day tsIC of 0\.0749\. Compared with short\-horizon results, the long\-horizon results better match the design motivation of structured experience: historical experience is more likely to help the model identify industry narratives and trend effects that accumulate over multiple days\.
- •The 5\-day results reveal remaining failure modes of experience retrieval\.OnDeepSeek\-V4\-Flash, Pipe\+ToE underperforms the no\-experience Pipe on the 5\-day horizon, suggesting that retrieved experience may introduce irrelevant or outdated information for some intermediate\-horizon predictions\. This limits any strong claim that experience is always beneficial, and points to the need for further analysis of retrieval quality and horizon\-adaptive weighting\.

DeepSeek\-V4\-FlashQwen3\.6\-35B\-A3BMethod1d5d10d20d1d5d10d20dBaseline0\.01850\.02310\.02870\.03550\.02120\.02610\.04170\.0554Pipe0\.01540\.03200\.02430\.05170\.02330\.02970\.03490\.0676Pipe\+mem00\.01900\.02340\.02020\.03250\.01020\.01530\.01320\.0336Pipe\+MemRL0\.01420\.03170\.01750\.01820\.01130\.01760\.01860\.0481Pipe\+ToE0\.02420\.02340\.03130\.07410\.03020\.02830\.03220\.0749Table 1:Prediction\-horizon ablation for tsIC across 1\-, 5\-, 10\-, and 20\-trading\-day horizons\.Overall, theQwen3\.6\-35B\-A3Bresults corroborate the main finding fromDeepSeek\-V4\-Flash: Pipe\+ToE is most effective for long\-horizon prediction, while shorter horizons show more limited separation\.

### 5\.4Effect of Utility Update Strategy

Table[2](https://arxiv.org/html/2606.06960#S5.T2)compares two experience\-update strategies on the 20\-trading\-day prediction horizon\. The two settings share the same experience\-retrieval and prediction pipeline, differing only in how experience utility is updated: Pipe\+formula is the Pipe\+ToE setting used in the main experiments, which applies the formula\-based update in Section[4](https://arxiv.org/html/2606.06960#S4); Pipe\+llm directly asks the LLM to rewrite experience weights according to feedback\. It can be observed that

- •Formula\-based updating outperforms direct LLM updating on both metrics\.Compared with Pipe\+llm, Pipe\+formula improves tsIC from 0\.0598 to 0\.0741 and csIC from 0\.0431 to 0\.0528\. This result suggests that, in the current task, converting implicit market feedback into constrained numerical updates is more effective than fully relying on the LLM to rewrite experience weights\.
- •The results support the importance of numerical constraints for experience evolution\.Financial sentiment feedback is delayed and noisy\. If the LLM directly decides the update magnitude, the update scale may become inconsistent across samples\. Formula\-based updating controls the step size through hit\-count decay and boundary constraints, making experience\-utility evolution smoother and easier to interpret\.

DeepSeek\-V4\-FlashStrategytsICcsICPipe\+formula0\.07410\.0528Pipe\+llm0\.05980\.0431Table 2:Update\-strategy ablation on the 20\-trading\-day prediction horizon\. Pipe\+formula is the formula\-based weight update used in this paper, while Pipe\+llm directly uses the LLM to update weights\.Overall, the experiments support that, under the currentFinEvolveBenchsetting, the effectiveness and stability of structured, formula\-updated experience management, while the horizon ablation reveals that its benefits remain conditional on retrieval quality and prediction horizon\.

## 6Conclusion

We introducedFinEvolveBench, a financial sentiment\-analysis benchmark for evaluating experience\-based self\-evolution in low\-repetition and implicit\-reward settings\. Unlike existing benchmarks that primarily examine static financial knowledge or procedural task execution,FinEvolveBenchplaces agents in a real continuous\-time market environment\. Agents must generate future market\-sentiment predictions from observable news under strict temporal splits, and their effectiveness is evaluated with financial metrics such as information coefficients\. On top of this benchmark, we further provided a structured experience\-management framework as a reference solution\. The framework represents experience as selectable and updatable external analytical patterns, and performs experience selection, expansion, and utility calibration while keeping the task model parameters frozen\.

## Ethical Considerations

We use publicly available financial news and market data only for research and benchmark evaluation\.FinEvolveBenchis not intended for investment advice, automated trading, or other real\-world financial decision\-making\. Users must comply with the terms of use of the original data sources\.

## 7Limitations

First, short\-horizon financial prediction is inherently noisy, so the effect of experience retrieval in short\-term settings should be interpreted with caution\. The current results show that the correlation between news\-based predictions and future returns is relatively weak on shorter prediction horizons\. This suggests that short\-term market movements contain substantial noise, and news signals may not be sufficient to explain price changes reliably within a short window\. Under such conditions, introducing historical experience may not always improve prediction performance, because the retrieved experience can be only superficially similar to the current case but limited in actual predictive value\. Therefore, although the proposed method shows clearer advantages on longer prediction horizons, its effectiveness remains less stable in short\-horizon settings, and experience retrieval should not be assumed to be beneficial across all time windows\.

Second, the proposed method still depends on the quality of upstream news structuring\. Both Pipe and Pipe\+tExp rely on steps such as news importance scoring, industry mapping, news aggregation, tag matching, and semantic reranking\. If errors occur in news filtering, industry assignment, or importance estimation, these errors may propagate into the experience base, affect experience retrieval, and eventually influence the final prediction\. Therefore, the structured experience mechanism improves how historical information is organized and used, but it does not eliminate the dependence on high\-quality financial text preprocessing\.

Third, the current experiments mainly focus on predictive evaluation metrics rather than a complete trading backtest\. IC, tsIC, and csIC can reflect the correlation between predicted factors and future returns, as well as the cross\-sectional ranking ability of the model\. However, they do not directly account for transaction costs, turnover, liquidity constraints, risk exposure, position sizing, or portfolio construction\. As a result, the current results should be understood as predictive\-evaluation evidence underFinEvolveBench, rather than as direct evidence that the proposed method can generate stable and realizable trading profits\.

## References

- FinBERT: financial sentiment analysis with pre\-trained language models\.arXiv preprint arXiv:1908\.10063\.Cited by:[§3\.2](https://arxiv.org/html/2606.06960#S3.SS2.SSS0.Px3.p1.1),[§8\.2](https://arxiv.org/html/2606.06960#S8.SS2.SSS0.Px3.p1.1)\.
- V\. Barreset al\.\(2025\)τ2\\tau^\{2\}\-Bench: evaluating conversational agents in a dual\-control setting\.arXiv preprint arXiv:2506\.07982\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- K\. Chenet al\.\(2025\)XBench: tracking agents productivity scaling with profession\-aligned real\-world evaluations\.arXiv preprint arXiv:2506\.13651\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Chen, Z\. Yao, Y\. Liu, A\. Xin, J\. Ye, J\. Yu, L\. Hou, and J\. Li \(2025\)StockBench: can llm agents trade stocks profitably in real\-world markets?\.arXiv preprint arXiv:2510\.02209\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Chen, W\. Chen, C\. Smiley, S\. Shah, I\. Borova, D\. Langdon, R\. Moussa, M\. Beane, T\. Huang, B\. Routledge, and W\. Y\. Wang \(2021\)FinQA: a dataset of numerical reasoning over financial data\.Proceedings of EMNLP\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px2.p1.1)\.
- P\. Chhikara, D\. Khant, S\. Aryan, T\. Singh, and D\. Yadav \(2025\)Mem0: building production\-ready ai agents with scalable long\-term memory\.External Links:2504\.19413Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.06960#S3.SS2.p1.1),[§5\.1](https://arxiv.org/html/2606.06960#S5.SS1.SSS0.Px1.p1.1),[3rd item](https://arxiv.org/html/2606.06960#S8.I1.i3.p1.1)\.
- C\. E\. Jimenez, J\. Yang, A\. Wettig, S\. Yao, K\. Pei, O\. Press, and K\. Narasimhan \(2024\)SWE\-bench: can language models resolve real\-world github issues?\.arXiv preprint arXiv:2310\.06770\.Cited by:[§1](https://arxiv.org/html/2606.06960#S1.p1.1),[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- A\. Lopez\-Lira and Y\. Tang \(2023\)Can ChatGPT forecast stock price movements? return predictability and large language models\.arXiv preprint arXiv:2304\.07619\.Cited by:[§5\.1](https://arxiv.org/html/2606.06960#S5.SS1.SSS0.Px1.p1.1),[1st item](https://arxiv.org/html/2606.06960#S8.I1.i1.p1.3)\.
- G\. Mialon, C\. Fourrier, C\. Swift, T\. Wolf, Y\. LeCun, and T\. Scialom \(2023\)GAIA: a benchmark for general ai assistants\.arXiv preprint arXiv:2311\.12983\.Cited by:[§1](https://arxiv.org/html/2606.06960#S1.p1.1),[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Mohan, S\. Mullapudi, S\. Sammeta, P\. Vijayvergia, and D\. C\. Anastasiu \(2019\)Stock price prediction using news sentiment analysis\.In2019 IEEE Fifth International Conference on Big Data Computing Service and Applications,pp\. 205–208\.Cited by:[§5\.1](https://arxiv.org/html/2606.06960#S5.SS1.SSS0.Px1.p1.1),[1st item](https://arxiv.org/html/2606.06960#S8.I1.i1.p1.3)\.
- I\. Muhammad, M\. Rospocher, T\. Knez, and S\. Žitnik \(2025\)Benchmarking large language models for target\-based financial sentiment analysis\.InProceedings of the Italian Conference on Computational Linguistics,Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px2.p1.1)\.
- L\. Phan, A\. Gatti, Z\. Han, N\. Li, J\. Hu, H\. Zhang, C\. B\. C\. Zhang, M\. Shaaban, J\. Ling, S\. Shi,et al\.\(2025\)Humanity’s last exam\.arXiv preprint arXiv:2501\.14249\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\)Reflexion: language agents with verbal reinforcement learning\.Advances in Neural Information Processing Systems36,pp\. 8634–8652\.Cited by:[§1](https://arxiv.org/html/2606.06960#S1.p1.1),[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- M\. Shridhar, X\. Yuan, M\. Côté, Y\. Bisk, A\. Trischler, and M\. Hausknecht \(2020\)Alfworld: aligning text and embodied environments for interactive learning\.arXiv preprint arXiv:2010\.03768\.Cited by:[§1](https://arxiv.org/html/2606.06960#S1.p1.1),[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- D\. Silver and R\. S\. Sutton \(2025\)Welcome to the era of experience\.Google AI\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- M\. Wang and T\. Ma \(2024\)MANA\-Net: mitigating aggregated sentiment homogenization with news weighting for enhanced market prediction\.InProceedings of the 33rd ACM International Conference on Information and Knowledge Management,Cited by:[§5\.1](https://arxiv.org/html/2606.06960#S5.SS1.SSS0.Px1.p1.1),[1st item](https://arxiv.org/html/2606.06960#S8.I1.i1.p1.3)\.
- Q\. Wang, W\. Xu, and H\. Zheng \(2018\)Combining the wisdom of crowds and technical analysis for financial market prediction using deep random subspace ensembles\.Neurocomputing299,pp\. 51–61\.Cited by:[§5\.1](https://arxiv.org/html/2606.06960#S5.SS1.SSS0.Px1.p1.1),[1st item](https://arxiv.org/html/2606.06960#S8.I1.i1.p1.3)\.
- T\. Wei, N\. Sachdeva, B\. Coleman, Z\. He, Y\. Bei,et al\.\(2025\)Evo\-memory: benchmarking llm agent test\-time learning with self\-evolving memory\.arXiv preprint arXiv:2511\.20857\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- Q\. Xieet al\.\(2024\)FinBen: a holistic financial benchmark for large language models\.Advances in Neural Information Processing Systems\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px2.p1.1)\.
- H\. Yang, X\. Liu, and C\. D\. Wang \(2023\)FinGPT: open\-source financial large language models\.arXiv preprint arXiv:2306\.06031\.Cited by:[§3\.2](https://arxiv.org/html/2606.06960#S3.SS2.SSS0.Px3.p1.1),[§8\.2](https://arxiv.org/html/2606.06960#S8.SS2.SSS0.Px3.p1.1)\.
- S\. Zhang, J\. Wang, R\. Zhou, J\. Liao, Y\. Feng,et al\.\(2026\)MemRL: self\-evolving agents via runtime reinforcement learning on episodic memory\.arXiv preprint arXiv:2601\.03192\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.06960#S3.SS2.p1.1),[§5\.1](https://arxiv.org/html/2606.06960#S5.SS1.SSS0.Px1.p1.1),[4th item](https://arxiv.org/html/2606.06960#S8.I1.i4.p1.1)\.
- J\. Zheng, X\. Cai, Q\. Li, D\. Zhang, Z\. Li, Y\. Zhang, L\. Song, and Q\. Ma \(2025\)LifelongAgentBench: evaluating llm agents as lifelong learners\.External Links:2505\.11942Cited by:[§1](https://arxiv.org/html/2606.06960#S1.p1.1),[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.
- F\. Zhu, W\. Lei, Y\. Huang, C\. Wang, S\. Zhang, J\. Lv, F\. Feng, and T\. Chua \(2021\)TAT\-QA: a question answering benchmark on a hybrid of tabular and textual content in finance\.InProceedings of ACL\-IJCNLP,Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px2.p1.1)\.
- T\. Y\. Zhuo, M\. C\. Vu, J\. Chim, H\. Hu, W\. Yu, R\. Widyasari, I\. N\. B\. Yusuf, H\. Zhan, J\. He, I\. Paul,et al\.\(2025\)Bigcodebench: benchmarking code generation with diverse function calls and complex instructions\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 66602–66656\.Cited by:[§2](https://arxiv.org/html/2606.06960#S2.SS0.SSS0.Px1.p1.1)\.

## 8Appendix

### 8\.1Details of FinEvolveBench

This section supplements the main text with additional construction details ofFinEvolveBench, including the investment universe, news collection and annotation pipeline, task\-instance format, temporal replay protocol, corpus profile, and market\-feedback statistics\. These details clarify how the benchmark preserves temporal causality and supports delayed feedback during online evaluation\.

#### 8\.1\.1Interactive Financial Market Environment with Temporally Grounded Feedback

We first describe the construction of the interactive market environment and its temporally grounded feedback mechanism\.

First, the environment defines a set of investment assets, i\.e\., stocks or indices for which market sentiment needs to be analyzed\. These assets are predefined to standardize benchmark usage and evaluation\. Second, the environment contains a time\-stamped news corpus that may be relevant to the fluctuations of the corresponding assets\. The data collection process is detailed below\. Third, the environment provides market data, including open price, close price, trading volume, and transaction price\. These signals enable benchmark users to design different forms of implicit rewards and provide a rich space for evaluating self\-evolution strategies\. Fourth, the environment is temporally interactive: at each time step, the trading agent can only access news and historical market data available up to the current time, and its predictions are validated and converted into feedback only after the corresponding prediction horizon has elapsed\.

##### Investment Assets\.

We define a diverse market universe inFinEvolveBenchby selecting 31 Shenwan Hongyuan first\-level industry indices from the Chinese A\-share market as investment assets\. Figure[4](https://arxiv.org/html/2606.06960#S8.F4)shows the industry coverage of the selected assets\. As shown in the figure, these first\-level industry indices cover nearly all major sectors observable in the market, providing sufficient diversity for model evaluation and thereby improving the robustness of the benchmark\.

![Refer to caption](https://arxiv.org/html/2606.06960v1/x4.png)Figure 4:Industry distribution of the selected Shenwan Hongyuan first\-level industry indices\.
##### News Corpus\.

We construct a news corpus for the investment assets, enabling agents to reason about sentiment events under an information setting similar to that faced by investors responding to public market narratives\. For each stock or index, we collect news articles published within each 24\-hour window\. The primary sources include mainstream Chinese financial media outlets such as Yuncaijing, ifeng Finance, Sina Finance, Eastmoney, Tonghuashun, JRJ, Wallstreetcn, Yicai, and Cailian Press\. Since these sources do not provide public application programming interfaces, we implement Python\-based crawlers using Selenium for data collection\. During crawling, we strictly follow the robots\.txt protocols of the corresponding websites, collect only publicly accessible content that does not require payment or registration, and control the request frequency to avoid imposing excessive load on target servers\. For each raw news article, we collect four fields: title, body text, publication time, and source website\. We further preprocess the collected articles with an LLM to obtain metadata such as industry classification and importance score\.

To preserve the high\-noise characteristics of real financial information flows, we retain all publicly collected media articles in the default news corpus and do not apply deduplication, threshold\-based filtering, or outlier removal\. This design treats duplicate and near\-duplicate reports as part of the public information environment, since repeated coverage may itself reflect media attention intensity and investor exposure\. Industry labels and importance scores are uniformly generated through an LLM\-as\-Judge procedure and are provided only as metadata for downstream methods, rather than as filtering criteria for removing articles\.

Market data and news are organized into dailyYYYYMMDD\.csvfiles to support temporally causal day\-by\-day replay\. Missing price fields are retained asNaN; no forward filling or interpolation is applied, so that downstream methods can explicitly decide how to handle missing observations\. The computation ofreturnandexcess returnis described in Appendix[8\.5](https://arxiv.org/html/2606.06960#S8.SS5)\.

##### Benchmark Construction Protocol\.

To ensure the reproducibility ofFinEvolveBench, we convert the raw market environment into a chronologically organized collection of sentiment analysis tasks\. Each task instance consists of a trading date, an investment target, the list of news articles observable on that day, and a set of market state fields, includingdate,sector,newsList,open,close, andvolume\. All instances are generated in trading\-day order\. During prediction, an agent may only access news, market prices, and historical experiences available at or before datett\. Rewards derived from future returns are revealed only after the corresponding prediction horizon has ended, and can then be used for evaluation or experience updating\.

Table[3](https://arxiv.org/html/2606.06960#S8.T3)summarizes the core statistics ofFinEvolveBench\. It is important to note thatFinEvolveBenchis not designed as a conventional training dataset\. Instead, it provides a temporally controlled simulation environment for evaluating agents under complex tasks with weak feedback\. The environment includes both left and right temporal buffers\. The left buffer strictly precedes the evaluation window and can be used for cold start, experience initialization, or historical calibration without exposing any information from the test period\. The right buffer follows the evaluation window and is used to compute delayed feedback for prediction horizons of 1, 5, 10, and 20 trading days\. Online updates during the evaluation period also follow causal constraints: feedback for anhh\-day prediction made at datettcan only be written into the experience system aftert\+ht\+h\.

The construction of the test environment follows two principles\. First, we retain the news stream as completely as possible to preserve the high\-noise nature of real financial markets\. Second, we use a recent evaluation window to reduce the risk of data leakage from the pretraining corpora of mainstream LLMs\. The processed data inFinEvolveBenchspans from January 1, 2024 to May 3, 2026\. The period from January 1, 2025 to March 31, 2026 is fixed as the online evaluation window, covering 300 trading days\. Data outside this window is used only as temporal buffer: the left buffer supports cold start, experience initialization, and historical calibration, while the right buffer supports delayed feedback computation for multi\-horizon prediction tasks\.

StatisticValueMarket universe31 SW first\-level industry indicesProcessed span2024\-01\-01 to 2026\-05\-03Online evaluation span2025\-01\-01 to 2026\-03\-31Trading days in evaluation300Training splitNoneNews articles177,324Active news days696Avg\. news / day254\.78Avg\. news / industry / day6\.08Prediction horizons1, 5, 10, 20 trading daysTable 3:Core statistics ofFinEvolveBench\.
##### Corpus Profile\.

Beyond the core scale statistics, we characterize the news corpus along four dimensions: time, source, industry, and event topic\. As shown in Figure[5](https://arxiv.org/html/2606.06960#S8.F5), the news stream provides continuous coverage from January 2024 to April 2026, while May 2026 contains only right\-buffer data up to May 3\. The corpus covers 9 news sources\. Sina Finance accounts for 55\.99% of the articles, and the remaining articles come from Wallstreetcn, Tonghuashun, Eastmoney, JRJ, ifeng Finance, Yuncaijing, Yicai, and Cailian Press\. Industry labels exhibit a long\-tailed distribution, with higher news density in sectors such as computer, electronics, petroleum and petrochemicals, transportation, non\-bank finance, and banking\. The importance scores are concentrated in relatively high ranges, with 56\.63% of the articles receiving a score of at least 8\. The topic distribution covers macroeconomics, policy planning, industrial and supply chains, geopolitics, market sentiment behavior, natural climate, and other event categories\.

Figure[6](https://arxiv.org/html/2606.06960#S8.F6)further shows the monthly industry exposure intensity\. The coverage shifts across industries over time, indicating that the environment is non\-stationary and not dominated by a single industry or a single period\.

![Refer to caption](https://arxiv.org/html/2606.06960v1/x5.png)Figure 5:News corpus profile ofFinEvolveBench, including monthly source distribution, major industry coverage, importance score distribution, and primary topic distribution\.![Refer to caption](https://arxiv.org/html/2606.06960v1/x6.png)Figure 6:Industry\-month news coverage heatmap\. Colors indicate the number of tag assignments for each industry in each month\. Since one article may contain multiple industry tags, the figure measures industry exposure intensity rather than deduplicated article counts\.
##### Market Feedback Profile\.

In addition to the high\-noise and non\-stationary input news stream, the feedback signal inFinEvolveBenchis continuous, delayed, and weakly supervised\. For each industry index, we compute the close\-to\-close forward return afterh∈\{1,5,10,20\}h\\in\\\{1,5,10,20\\\}trading days\. We then derive market\-adjusted feedback by subtracting the same\-horizon return of the CSI1000 index \(000852\.SH\), i\.e\.,αt,h=rt,hsector−rt,hCSI1000\\alpha\_\{t,h\}=r^\{\\text\{sector\}\}\_\{t,h\}\-r^\{\\text\{CSI1000\}\}\_\{t,h\}\. Figure[7](https://arxiv.org/html/2606.06960#S8.F7)presents the market feedback distribution within the online evaluation window\. The adjusted feedback distribution is approximately centered around zero but exhibits substantial dispersion\. As the prediction horizon increases from 1 day to 20 days, the standard deviation rises from 1\.20% to 5\.50%, and the 5–95 percentile interval expands from\[−1\.81%,1\.95%\]\[\-1\.81\\%,1\.95\\%\]to\[−9\.08%,8\.71%\]\[\-9\.08\\%,8\.71\\%\]\. Meanwhile, the proportion of positive feedback across horizons is only approximately 42\.25%–45\.67%, indicating that the task does not contain a stable one\-sided label bias\. This feedback structure prevents experience updating from relying on explicit correct/incorrect labels; instead, agents must estimate the long\-term utility of an analysis pattern from delayed market outcomes\.

![Refer to caption](https://arxiv.org/html/2606.06960v1/x7.png)Figure 7:Market feedback profile ofFinEvolveBench\. Raw forward returns are computed from sector close prices, and adjusted feedback subtracts the same\-horizon CSI1000 return\. The distributions are continuous, noisy, and horizon\-dependent, making feedback substantially weaker than explicit task labels\.

### 8\.2Self\-Evolving Agent Workflow

On top of the interactive environment inFinEvolveBench, we design a lightweight self\-evolving agent workflow\. Inspired by mainstream memory and experience\-based self\-evolution mechanisms such as Mem0 and MemRL, we decompose experience management into two stages: retrieval and updating\. This abstraction allows researchers to evaluate their own experience\-based self\-evolution methods under a unified benchmark interface\.

The workflow contains four main steps\.

##### Step 1: Information Perception and Task Formulation\.

The agent first perceives the available environmental information and extracts the subset needed for the current interaction\. Specifically, given the industry associated with the target to be analyzed, the agent retrieves and extracts important news relevant to that industry from the available information sources, thereby formulating a concrete financial sentiment analysis task\.

##### Step 2: Task\-Relevant Experience Retrieval\.

The experience management agent retrieves from the experience bank a set of experiences that may be useful for completing the current task\. These retrieved experiences may provide the downstream analysis agent with additional information about prior interactions with the environment, such as historical success rate, realized market reaction, or other outcome\-related statistics\. The exact information available to the analysis agent depends on the implementation of the experience management module\.

##### Step 3: Task Execution\.

Given the task description and the retrieved experiences, the agent executes the task\. Following prior work on financial sentiment analysis such as FinBERT\(Araci,[2019](https://arxiv.org/html/2606.06960#bib.bib27)\)and FinGPT\(Yanget al\.,[2023](https://arxiv.org/html/2606.06960#bib.bib30)\), the agent generates a continuous sentiment score in the range of\[−1,1\]\[\-1,1\]for the target under analysis\.

##### Step 4: Validation and Reflection\.

After task execution, the environment records the prediction trajectory but does not reveal feedback immediately\. Once the price corresponding to the prediction horizon becomes observable, the environment computes feedback for the prediction according to the predefined market\-feedback metric\. The feedback\-augmented trajectory is then returned to the experience management module for experience updating, and the updated experience can affect only subsequent trading days\. After these four steps are completed, the simulation advances to the next trading day\.

This workflow provides a unified interface for evaluating different experience\-based self\-evolution methods\. However, it does not prescribe how experiences should be organized, retrieved, or updated\. These design choices constitute the core mechanisms that determine the effectiveness of self\-evolution\. The following sections provide the concrete online procedure and structured experience representation used in our implementation\.

### 8\.3Online Procedure of Experience\-Augmented Self\-Evolution

Table[4](https://arxiv.org/html/2606.06960#S8.T4)summarizes the online procedure of the experience\-augmented self\-evolution framework\. The procedure separates prediction\-time experience selection and expansion from delayed feedback\-based utility updates\. This separation ensures that the agent uses only information observable at the current prediction time, while feedback from future market outcomes is written back to the experience store only after the corresponding prediction horizon has elapsed\.

Algorithm 1Online Experience\-Augmented Self\-EvolutionInput:pre\-test historical tasks𝒟\\mathcal\{D\}, online tasks\{τt\}\\\{\\tau\_\{t\}\\\}, frozen task modelpθp\_\{\\theta\}, prediction horizonsℋ\\mathcal\{H\}Output:predictions\{at\}\\\{a\_\{t\}\\\}and updated experience storeℰ\\mathcal\{E\}1Initialize the depth\-constrained experience treeℰ0\\mathcal\{E\}\_\{0\}from𝒟\\mathcal\{D\}, where𝒟\\mathcal\{D\}is strictly earlier than the online evaluation window\.2For each online taskτt\\tau\_\{t\}, observe only the task context available at timett\.3Estimate context compatibilityρ​\(st,𝐏\)\\rho\(s\_\{t\},\\mathbf\{P\}\)for candidate experience paths and select𝐏∗\\mathbf\{P\}^\{\*\}according to the effective utility𝐐~​\(st,𝐏\)\\widetilde\{\\mathbf\{Q\}\}\(s\_\{t\},\\mathbf\{P\}\)\.4If the leaf\-level expansion judge determines that existing candidates are insufficient, expand a new leaf using the current context and the selected prefix path only\.5Condition the frozen task modelpθp\_\{\\theta\}onsts\_\{t\}and𝐏∗\\mathbf\{P\}^\{\*\}to generate the task actionata\_\{t\}\.6Store a pending record\(𝐏∗,at,t,h\)\(\\mathbf\{P\}^\{\*\},a\_\{t\},t,h\)for each prediction horizonh∈ℋh\\in\\mathcal\{H\}\.7When the feedback for a pending record becomes observable, compute𝐫env\\mathbf\{r\}\_\{\\text\{env\}\}and update𝐐​\(𝐏∗\)\\mathbf\{Q\}\(\\mathbf\{P\}^\{\*\}\); the updated experience affects only subsequent tasks\.Table 4:Online procedure of the experience\-augmented self\-evolution framework\. The algorithm separates prediction\-time experience selection and expansion from delayed feedback\-based utility update, preserving temporal causality during evaluation\.
### 8\.4Structured Experience Representation

FactorAnalysis directionImpact horizonsRecall/HitEIDMacroeconomyPricing capacity utilization from aggregate\-demand cycles\[1\.0574, 1\.6602, 0\.5981, 0\.9529\]705/2081049Policy planningSupply\-side administrative constraints and capacity\-entry barriers\[1\.2198, 1\.1205, 0\.3548, 0\.2004\]1125/6021044GeopoliticsGeopolitical uncertainty, policy flexibility of oil\-producing countries, and supply\-side capacity disruptions\[1\.1273, 1\.4600, 0\.7345, 0\.8703\]22446/132091040Table 5:Examples of structured experience in financial sentiment analysis\. Impact horizons denote task\-related utility estimates for different prediction horizons\.Representative structured experience entries are shown in Table[5](https://arxiv.org/html/2606.06960#S8.T5)\. Each structured experience decomposes prior market reasoning into a reusablefactor, a concreteanalysis direction, horizon\-specific utility estimates, and non\-semantic metadata such as recall counts, hit counts, and experience identifiers\. Thefactorandanalysis directionfields form reusable paradigm paths, theimpact horizonscorrespond to task\-related utility estimates over different prediction horizons, andrecall,hit, andeidare used only for bookkeeping and utility updates\.

### 8\.5Experimental Details

##### Evaluated models and decoding\.

Our main experiments useDeepSeek\-V4\-Flashas the primary backbone model, withQwen3\.6\-35B\-A3Bincluded as an additional backbone\-model ablation\. For both backbones, the context window is set to 32,768 tokens, and we use the officially recommended decoding configuration\. Unless otherwise specified, each method is run independently three times, and the tables report the average performance across the three runs\.

##### Benchmark scope and temporal splits\.

FinEvolveBenchuses 31 Shenwan first\-level industry indices from the Chinese A\-share market as investment assets\. The processed environment spans from January 1, 2024 to May 3, 2026\. The online evaluation window is fixed to January 1, 2025–March 31, 2026, covering 300 trading days\. Data before the evaluation window is used only as the left temporal buffer for cold start, experience initialization, and historical calibration, while data after the evaluation window is used only as the right temporal buffer for delayed feedback computation\. We use no conventional training split\. The prediction horizons areh∈\{1,5,10,20\}h\\in\\\{1,5,10,20\\\}trading days\.

##### News and market data processing\.

For each investment target, the environment provides timestamped news articles and market data fields including open price, close price, trading volume, and transaction price\. Raw news articles are collected from public Chinese financial media sources and are retained to preserve the high\-noise characteristics of real financial information flows\. We do not apply deduplication, threshold\-based filtering, or outlier removal to the news corpus\. Industry labels and importance scores are generated through an LLM\-as\-Judge procedure and used as metadata for downstream methods, rather than as filtering criteria\. Market data and news are organized into dailyYYYYMMDD\.csvfiles to support temporally causal day\-by\-day replay\. Missing price fields are retained asNaN; no forward filling or interpolation is applied\.

##### Task construction and information access\.

Each task instance consists of a trading date, an investment target, the news articles observable for the target industry, and market state fields such asdate,sector,newsList,open,close,volume, andtransaction\-related fields\. All instances are generated in trading\-day order\. During prediction at datett, a method can only access news, market prices, and historical experiences observable at or beforett\. Feedback derived from future returns is revealed only after the corresponding prediction horizon has elapsed and can affect only subsequent tasks\.

##### Hyperparameters\.

Across all experiments, we set the experience\-tree depth toL=2L=2, the first\-level candidate\-retention size tok1=8k\_\{1\}=8, the leaf\-level retrieval cosine\-similarity threshold toλ=0\.8\\lambda=0\.8, the neutral utility toq0=1q\_\{0\}=1, and the learning rate toη=10\\eta=10\. These hyperparameters are fixed before the online evaluation window begins\. We do not use future test\-period news, future returns, or online evaluation results for hyperparameter tuning\.

##### Feedback mapping\.

For each industry index and prediction horizonhh, we compute the close\-to\-close sector forward return

rt,hsector=closet\+hsectorclosetsector−1\.r^\{\\mathrm\{sector\}\}\_\{t,h\}=\\frac\{\\mathrm\{close\}^\{\\mathrm\{sector\}\}\_\{t\+h\}\}\{\\mathrm\{close\}^\{\\mathrm\{sector\}\}\_\{t\}\}\-1\.\(16\)We then derive market\-adjusted excess return by subtracting the same\-horizon CSI1000 index return:

αt,h=rt,hsector−rt,hCSI1000,\\alpha\_\{t,h\}=r^\{\\mathrm\{sector\}\}\_\{t,h\}\-r^\{\\mathrm\{CSI1000\}\}\_\{t,h\},\(17\)where CSI1000 denotes index000852\.SH\. Given the predicted sentiment scorest,hs\_\{t,h\}, we use its sign as the predicted direction,

y^t,h=sign⁡\(st,h\),\\hat\{y\}\_\{t,h\}=\\operatorname\{sign\}\(s\_\{t,h\}\),\(18\)and construct the raw scalar feedback

xt,h=y^t,h⋅αt,h\.x\_\{t,h\}=\\hat\{y\}\_\{t,h\}\\cdot\\alpha\_\{t,h\}\.\(19\)In the formula\-based update setting, we use the identity feedback mapping

renv=ψτ​\(at,ot\+h\)=xt,h\.r\_\{\\mathrm\{env\}\}=\\psi\_\{\\tau\}\(a\_\{t\},o\_\{t\+h\}\)=x\_\{t,h\}\.\(20\)Thus, the feedback is continuous, delayed, horizon\-specific, and market\-adjusted\. For a prediction made at datett, the corresponding feedback can be written into the experience system only after datet\+ht\+h\.

##### Fairness controls\.

To ensure that the comparison focuses on the experience\-management mechanism itself, all methods share the same evaluated backbone model, decoding configuration, news\-structuring pipeline, prediction prompt, online evaluation order, prediction horizons, and evaluation metrics\. For experience\-augmented settings, including Pipe\+mem0, Pipe\+MemRL, Pipe\+ToE, and Pipe\+llm, we use the same pre\-test historical tasks as the source for experience initialization and enforce the same temporal\-causality constraint\. Differences among experience methods therefore come only from their experience representation, retrieval strategy, and update rule, rather than from additional test\-period data or different prediction models\.

##### Compared methods\.

We compare five main types of methods in RQ1 and RQ2\. Except for Baseline, all methods use the same structured news annotations and prediction prompt\. Thus, the difference between Baseline and Pipe mainly reflects the effect of structured news processing, while the difference between Pipe and the experience\-augmented methods is used to analyze the contribution of experience retrieval and updating\.

- •Baseline: Following prior financial market prediction studies that use news\-derived sentiment, crowd sentiment, and LLM\-extracted textual signals as predictive factors\(Wanget al\.,[2018](https://arxiv.org/html/2606.06960#bib.bib34); Mohanet al\.,[2019](https://arxiv.org/html/2606.06960#bib.bib33); Lopez\-Lira and Tang,[2023](https://arxiv.org/html/2606.06960#bib.bib29); Wang and Ma,[2024](https://arxiv.org/html/2606.06960#bib.bib32)\), the LLM performs sentiment analysis for each news item and discretizes the output assentiment∈\{−1,0,1\}\\text\{sentiment\}\\in\\\{\-1,0,1\\\}\. For the news set associated with an industry from dayT−4T\-4to dayTT, we compute an importance\-weighted average using each news item’simportance\.scoreas the forecast sentiment for that industry\. This method does not introduce experience retrieval or a complex reasoning framework\.
- •Pipe: The pipeline first produces structured annotations for each news item, such as importance score and related industry, keeps all news from dayT−4T\-4to dayTTfor the target industry, and sends the aggregated news to the LLM prediction module to generate horizon\-specific sentiment scores\. This pipeline does not include experience retrieval\.
- •Pipe\+mem0: After obtaining the aggregated news, this method uses mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2606.06960#bib.bib22)\), a general\-purpose experience system based on vector storage and semantic retrieval, for experience management\. Retrieved experience and aggregated news are passed together to the prediction module\.
- •Pipe\+MemRL: After obtaining the aggregated news, this method uses MemRL\(Zhanget al\.,[2026](https://arxiv.org/html/2606.06960#bib.bib26)\)for experience management\. It retrieves relevant experience using the MemRL retrieval policy, passes the retrieved experience and aggregated news to the prediction module, and updates experience with the MemRL update mechanism after the corresponding feedback becomes observable\.
- •Pipe\+ToE \(ours\): After obtaining the aggregated news, this method uses Tree\-of\-Experience \(ToE\) for experience management\. It retrieves experience from a structured experience tree through internal\-layer alignment and leaf\-level adaptation, conditions the prediction module on the retrieved experience and aggregated news, records the prediction, and applies formula\-based utility updating after delayed feedback becomes observable\. Unless otherwise specified in RQ3, Pipe\+ToE refers to the version that uses formula\-based utility updates\.

DeepSeek\-V4\-FlashQwen3\.6\-35B\-A3BMethod1d5d10d20d1d5d10d20dBaseline\-0\.00410\.01680\.02000\.01490\.01200\.03050\.03770\.0363Pipe0\.00200\.01920\.02080\.04080\.02450\.03560\.03420\.0542Pipe\+mem00\.00160\.00830\.01410\.02400\.00620\.01840\.00980\.0272Pipe\+MemRL0\.01680\.02960\.01630\.00600\.00050\.00240\.01350\.0514Pipe\+ToE0\.01280\.02750\.03450\.05280\.01380\.02850\.03000\.0524Table 6:Prediction\-horizon ablation for csIC across 1\-, 5\-, 10\-, and 20\-trading\-day horizons\.
##### Update\-strategy ablation\.

For RQ3, we additionally compare two update strategies under the same retrieval and prediction pipeline on the 20\-trading\-day horizon\.Pipe\+formulais the Pipe\+ToE setting used in the main experiments, which applies the formula\-based utility update\.Pipe\+llmdirectly asks the frozen LLM to update the utility representation after receiving delayed feedback\. This ablation isolates the effect of the utility\-update rule while keeping experience retrieval, prediction prompts, backbone model, and evaluation metrics unchanged\.

##### Evaluation metrics\.

We use the information coefficient \(IC\), a standard metric for evaluating factor quality in quantitative trading, as the main evaluation metric\. IC is appropriate forFinEvolveBenchbecause the task is not to predict a categorical ground\-truth label for each news item, but to produce a sentiment factor whose values should align with future market movements\.

We report two IC\-based metrics\.Cross\-sectional IC \(csIC\)measures the correlation between predicted factor values and future returns across different assets on the same trading day\. Suppose there areNtN\_\{t\}trading assets on daytt\. Letsi,t,hs\_\{i,t,h\}denote the predicted sentiment value for theii\-th target at horizonhh, and letαi,t,h\\alpha\_\{i,t,h\}denote its same\-horizon excess return\. The cross\-sectional IC is defined as:

c​s​I​Ct,h=Corri⁡\(si,t,h,αi,t,h\)\.csIC\_\{t,h\}=\\operatorname\{Corr\}\_\{i\}\(s\_\{i,t,h\},\\alpha\_\{i,t,h\}\)\.\(21\)The final csIC for horizonhhis the average over all evaluated trading days:

c​s​I​Ch=1T​∑t=1Tc​s​I​Ct,h\.csIC\_\{h\}=\\frac\{1\}\{T\}\\sum\_\{t=1\}^\{T\}csIC\_\{t,h\}\.\(22\)Time\-series IC \(tsIC\)measures the correlation between predicted factor values and future excess returns across time for the same trading target:

t​s​I​Ci,h=Corrt⁡\(si,t,h,αi,t,h\)\.tsIC\_\{i,h\}=\\operatorname\{Corr\}\_\{t\}\(s\_\{i,t,h\},\\alpha\_\{i,t,h\}\)\.\(23\)The final tsIC for horizonhhis the average over all trading assets:

t​s​I​Ch=1N​∑i=1Nt​s​I​Ci,h\.tsIC\_\{h\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}tsIC\_\{i,h\}\.\(24\)

### 8\.6Cross\-sectional prediction\-horizon ablation

Table[6](https://arxiv.org/html/2606.06960#S8.T6)reports the csIC results across 1\-, 5\-, 10\-, and 20\-trading\-day prediction horizons\. Cross\-sectional ranking ability is also stronger on longer horizons: Pipe\+ToE achieves the best results on the 10\- and 20\-day horizons withDeepSeek\-V4\-Flash, and the 20\-day csIC reaches 0\.0528, higher than Pipe’s 0\.0408\. OnQwen3\.6\-35B\-A3B, Pipe\+ToE also reaches a 20\-day csIC of 0\.0524\. This suggests that structured experience improves not only time\-series prediction for individual assets, but also relative ranking among assets on the same trading day\.