arXiv subjects

Yixin Cao

Publications and source records attributed to Yixin Cao.

At least 19 recordsLinked to original sources

arXiv2026

Before the Next Rollout: Diagnosing Skill Utility Shifts from Policy Updates

Skill-evolution methods typically analyze execution trajectories to guide agent skill refinement. However, policy updates can change the behavioral effects and marginal utility of an unchanged skill. Attribution based on earlier trajectories may therefore no longer reflect the skill's contribution under the updated policy. Moreover, trajectory failures can arise from multiple factors, making individual skill contributions difficult to isolate. We evaluate whether policy-update signals contain predictive information about changes in skill marginal utility. At fixed contexts from pre-update training rollouts, we measure how policy updates change the contrast in token preferences between skill and control conditions. Combining these interaction shifts with advantage estimates yields reward-directed readouts for diagnosing utility changes. The readouts reuse existing training rollouts and require only forward passes for single-step action scoring, without additional post-update rollout evaluation. Our experiments show that policy updates can alter skill marginal utility and that policy-update readouts carry diagnostic information for identifying skills with declining utility. Integrating readout-guided prioritization into skill evolution improves downstream performance while reducing policy and editor token consumption.

cs.AI↗

arXiv2026

ACWM-RM: A Reference-Free Reward Model from Complementary Mismatch Preferences

Reinforcement learning has gained attention for improving action-conditioned world models (ACWMs), as it enables reward-driven optimization of visual quality (VQ) and action-following fidelity (AF) beyond flow-matching optimization. Realizing this potential requires reward signals that reliably rank videos by both VQ and AF. However, existing rewards capture only one dimension: metric-based rewards assess VQ but require ground-truth videos and do not directly assess AF, while inverse dynamics model (IDM) rewards assess AF without references but do not guarantee VQ. Moreover, IDMs, typically trained on real videos, may provide unreliable rewards for generated videos containing visual artifacts and physical inconsistencies. In this work, we introduce the Action and Video Mismatch Benchmark (AVM-Eval), comprising 67.2k pairs that evaluate how reliably IDMs assess AF on real and generated videos. We then propose ACWM-RM, a reward model that jointly assesses VQ and AF without ground-truth videos during reward computation. To train it, we first construct the Action and Video Mismatch Preference Dataset (AVM-Pref) with 506.3k pairs providing complementary supervision: action-mismatch pairs supervise AF, while video-mismatch pairs supervise VQ and expose the model to generated videos. We then pretrain an action-conditioned predictor on real videos to learn forward dynamics before jointly training it with a patch-weighted score head on AVM-Pref to transfer these dynamics to AF assessment on generated videos. Evaluation on AVM-Eval reveals the limited reliability of IDMs on generated videos, while downstream RL experiments show that ACWM-RM outperforms reward baselines in both criteria.

cs.CV↗

arXiv2026

SERO: Learning from Executable Oracle Continuations for Long-Horizon Agentic Reinforcement Learning

Long-horizon agentic RL is constrained by the futures visited by the current policy. When rollouts fail to reach useful behaviors, improved credit assignment cannot provide supervision for states and actions that were never encountered. On-policy distillation (OPD) introduces teacher guidance on policy-induced states, and recent multi-turn variants improve how this guidance is scheduled, replayed, refined, or validated. However, they do not directly use the oracle-controlled future itself as the training trajectory. We introduce \textbf{State-Rooted Executable Rollout Optimization (SERO)}, which augments on-policy agentic RL with executable oracle continuations. SERO restores a state actually reached by the current policy and lets a training-only oracle continue interacting with the real environment, making the downstream observations and oracle actions themselves available for learning. It then applies \textbf{policy-reachability weighting}, which weights each oracle step by the frozen behavior policy's compatibility with the preceding oracle actions. This allows deeper supervision to become effective as the policy becomes more compatible with the oracle branch. The oracle loss is optimized separately from the base RL estimator and no oracle is used at inference time. We evaluate SERO on two challenging interactive-agent benchmarks, ALFWorld and WebShop, using Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct. On ALFWorld, learning the full executable continuation is substantially more effective than supervising only the root oracle action, while policy-reachability weighting further improves out-of-distribution performance.

cs.AI↗

arXiv2026

Travel Back To Move Forward: Retrospective Exploration and Evolving for Long Horizon Agents

Large language model agents increasingly operate in long-horizon environments where early mistakes can trigger cascading failures and trap agents in dead-end trajectories. Agents must therefore be able to proactively recover from such states and turn their exploratory experience into guidance for subsequent decisions. We introduce ReTRAIL, a framework that enables agents to rewind the environment to earlier states while preserving valid progress and carrying forward knowledge learned from abandoned exploration. The environment state moves backward while the agent's knowledge continues to move forward. ReTRAIL trains a shared policy with two complementary signals: State-Anchored GRPO compares alternative continuations from revisited states to improve execution, while a branch-advancement reward guides reflection through the progress of subsequent exploration. Experiments across four interactive benchmark MineSweeper, Sokoban, ALFWorld and ScienceWorld show that simply adding a rewind primitive provides limited benefit and can even degrade performance. In contrast, ReTRAIL enables agents to rewind from unproductive states and use insights from abandoned branches to guide subsequent exploration. These results demonstrate the value of coupling environment rewinding with experience reuse within an ongoing task for long-horizon task solving. The code is available at https://github.com/whi497/ReTRAIL.

cs.AI↗

arXiv2026

DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement

Online agent deployments accumulate execution trajectories at massive scale and behavioral diversity, for which predefined annotation criteria hardly exist. Extracting useful evidence therefore demands costly manual annotation or verifier signals that fail to scale, leaving valuable evidence buried among redundant, incomplete, and failed executions. This raises a question: without post-execution rewards or correctness labels, how can reusable experience be distilled from the trajectories themselves? To address this challenge, we introduce DENSE (Distilling Evidence from Nested Subtask Executions), which organizes trajectory-derived evidence into nested shortcut trees. By consolidating redundant attempts, identifying resolved subtasks, and retaining useful steps alongside outstanding requirements, DENSE transforms noisy execution traces into structured and reusable task-solving feedback. To evaluate whether such feedback helps agents retry the same task, we design REFIT, which measures success-rate changes between the initial attempt and feedback-guided retries. Among feedback methods without external outcome supervision, DENSE achieves the highest strict pass rate across four agent models on Terminal-Bench 2.1, improving over initial attempts by 7.12-21.81 percentage points with 19.0-43.6% fewer agent tokens on retries. In addition, on hard tasks DENSE consistently outperforms self-reflection in cumulative pass rate across multiple feedback iterations on all four models, demonstrating its strong potential for continual agent self-improvement.

cs.AI↗

arXiv2026

Belief-Trajectory Energy: Measuring the Path to a Prediction

Large language models (LLMs) progressively revise their predictions across Transformer layers, yet we typically observe only the final output, discarding the trajectory through which it is formed. We introduce Belief-Trajectory Energy(BTE), a model-grounded measure that characterizes an input through the layerwise predictive revisions it induces in a model. By mapping intermediate states into a shared predictive space, BTE provides a principled measure of belief change that can be summarized as either a scalar or a structured depth profile. Theoretically, we show that local BTE corresponds to predictive revision under the Fisher-Rao geometry, while the sequence of revisions captures information beyond the initial-to-final belief change. Empirically, scalar BTE provides a model-relative signal of difficulty across diverse reasoning tasks, while richer BTE representations support human-LLM review detection and fine-grained generator attribution, reaching up to $0.998$ macro-AUROC and $95.6\%$ eight-way attribution accuracy. Further analysis shows that BTE develops throughout pretraining and is selectively reshaped by targeted training, demonstrating that the resulting measurement reflects what the scoring model has learned. Together, our results establish belief trajectories as a principled model-grounded signal and suggest a broader perspective in which learned models can themselves serve as instruments for characterizing the data they process. More demonstrations can be found at https://yingjiahao14.github.io/BTE-web/.

cs.CL↗

arXiv2026

Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks

Large Language Models demonstrate remarkable proficiency in static reasoning, yet training them as autonomous agents through Reinforcement Learning (RL) for long-horizon tasks is often hindered by severe reward sparsity. While conventional \textit{agent-side warming} up via supervised fine-tuning (SFT) can alleviate this, it is frequently limited by data scarcity and constrained exploration. To address this, we propose a paradigm shift to \textit{environment-side adaptation} by constructing \textbf{F}eedback-\textbf{E}nriched \textbf{E}nvironments (\textbf{FEEs}). Through a pilot study, we establish a feedback design strategy that reformulates environments by transitioning from action guidance to observation enrichment during the later stages of both intra-episode exploration and inter-episode evolution. Large-scale experiments on SciWorld and BFCL benchmarks using various Qwen3 model scales and RL algorithms such as GRPO, GSPO, and DAPO demonstrate that FEEs consistently yield performance improvements over standard settings. Furthermore, our analysis reveals that training with FEEs \textbf{(1)} stabilizes training dynamics by reducing entropy volatility, \textbf{(2)} facilitates proactive state-space exploration in difficult tasks, \textbf{(3) }ensures the internalization of environmental guidance into policy weights rather than acting as a mere inference-time prior, and \textbf{(4) }identifies intra-group feedback consistency as a critical boundary for stable optimization.

cs.LG↗

arXiv2026

From Concentration to Differentiation and Back: Routing Effective Rank in MoE Reasoning Cohorts

Test-time scaling produces cohorts of reasoning rollouts, yet there is no standard label-free account of how their internal computation reorganizes as inference unfolds. We introduce routing effective rank deff, the entropy-effective dimensionality of a cross-rollout graph built from MoE expert-routing similarity. Across ten MoE configurations and five math/science benchmarks, deff exhibits a reproducible low-high-low trajectory, with a prominent interior maximum in 98.5% of 3,105 model-question cohorts: routing similarity is concentrated early, maximally differentiated at intermediate budgets, and reconcentrated later, and the timing of this maximum varies systematically with architecture and reasoning effort. An exact decomposition separates cohort-wide common-mode mass from residual spectral dimensionality: common-mode reallocation accounts for about two thirds of the trajectory, while the residual spectrum contributes about one quarter and retains substantial variation beyond the common mode. The decomposition further localizes behavior: among non-unanimous cohorts, increases in common-mode concentration strongly predict same-answer recoverability, and higher reasoning effort delays the maximum by 2.59 octaves (doublings of the token budget) and consistently expands the high-rank period across all four tested architectures, locating the effort effect in timing and duration rather than peak amplitude. Correctness comparisons separate structural monitoring from answer selection, positioning routing effective rank as a decomposable, label-free diagnostic of cohort organization - a principled spectral lens on how MoE reasoning cohorts differentiate and reconcentrate over inference time.

cs.AI↗

arXiv2026

From Base Rollouts to RL Reasoning: A Budgeted Search Perspective

Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but how these gains relate to inference-time decoding and search remains unclear. Does RL create reasoning the base model lacks, or shift the rollout distribution toward trajectories it can already reach but rarely samples? We study this behaviorally with a Unified Decoding Framework (UDF), which expresses token-level sampling, beam-like search, tree search, and sequence-level resampling as executable policies over a shared budgeted operating space, scored post hoc with pass@$k$, self-consistency, best-of-$N$, and first-finish success. Using paired Base/RL checkpoints from SimpleRL-Zoo, we ask whether an RL default-policy curve can be approximated by a structured path of Base operating points. On Math500, AIME, GPQA, and IFEval, the pass@$k$ recovery path follows a Budgeted Operating-Point Transition Rule (BOPTR), $N_{\mathrm{Base}} \approx αN_{\mathrm{RL}}^β$, with benchmark-conditioned exponents. On Qwen2.5-7B, BOPTR gives the lowest transfer error among the non-oracle rules we test, 3.41 pp (95% CI [2.32, 5.53]); a three-seed replication gives 3.07 $\pm$ 0.39 pp. The rule extends to ten models across four families (3.28 to 4.87 pp on checkpoints added after fitting), to four benchmarks it was never fitted on (5.03 pp vs. 4.44 pp in fit), and holds without an RL checkpoint for the target model (4.19 pp) or without RL supervision of any kind (5.08 pp). These results support a qualified internalized-search reading: under the recipe we test, much of the measured RL gain corresponds to a change in sampling efficiency toward operating points the base model can already reach under search. We treat the scaling patterns as descriptive of this recipe and cohort, report where they break down, and use UDF and BOPTR as behavioral diagnostics rather than evidence of parameter-level equivalence.

cs.CL↗

arXiv2026

EMemBench: Interactive Benchmarking of Episodic Memory for VLM Agents

We introduce EMemBench, a programmatic benchmark generator for evaluating long-term episodic memory of agents through interactive games. Rather than using a fixed set of questions, EMemBench generates questions from environment-grounded trajectories, covering both text-only and visual game environments. Each template computes verifiable ground truth from underlying game signals, with controlled answerability and balanced coverage over memory skills: single/multi-hop recall, induction, temporal, spatial, logical, and adversarial. We evaluate memory agents with strong LMs/VLMs as backbones, using in-context prompting as baselines. Across 15 text games and multiple visual seeds, results are far from saturated: induction and spatial reasoning are persistent bottlenecks, especially in visual settings. Persistent memory yields clear gains for open backbones on text games, but improvements are less consistent for VLM agents, suggesting that visually grounded episodic memory remains an open challenge. A human study further contextualizes the difficulty and interpretability of EMemBench.

cs.CL↗

arXiv2026

Skill-as-Pseudocode: Refactoring Skill Libraries to Pseudocode for LLM Agents

Markdown skill libraries for LLM agents ship as free-form prose, forcing the agent to re-derive both the input schema and the concrete invocation syntax on every retrieval. This produces a "confused $\to$ re-retrieve $\to$ still confused" loop: the agent issues a partially-correct action, receives uninformative feedback, and re-retrieves the same prose. We propose Skill-as-Pseudocode (SaP), an automatic conversion of markdown skill libraries into typed pseudocode with deterministic quality control. From each cluster of similar procedural passages, SaP extracts a typed contract and filters it through a four-check deterministic verifier (coverage, binding, replacement, risk). Promoted contracts are inlined into a rewritten skill skeleton alongside restored action templates, giving the agent two complementary signals: a typed signature for what a skill does and a concrete template for how to invoke it. On the ALFWorld unseen split (134 games, gpt-4o-mini, three seeds), SaP wins 82/402 paired games versus 47/402 for the Graph-of-Skills (GoS) baseline (pooled McNemar $p = 8.2 \times 10^{-5}$), at $-22.8 \pm 6.4$% input tokens and $-14.5 \pm 4.1$% LLM calls per game. A bundle-component ablation attributes the gain to the pairing of typed contracts with concrete action templates: the contract alone falls below the prose baseline.

cs.PL↗

arXiv2026

Beyond the Trace: Coupling an Interpretable Reasoning-State Readout to Native MoE Routing

What a reasoning model writes is only a partial record of the process that produces it. We introduce a two-level internal readout for mixture-of-experts reasoning. We first distill vocabulary-scale J-space into J64, a 64-axis semantic frame learned from the model's own reasoning states. J64 reveals readable process state that the emitted trace does not show: it separates inference effort from problem-induced strain. It also adds 0.096 to 0.135 held-out AUC over a baseline that reads the same rollout as token occupancy and aggregates it in exactly the same way. We then reconstruct J64 from native expert-routing statistics. The result is R64, a low-overhead proxy: its median per-axis correlation with J64 is 0.69 to 0.86 across three models and two families, and on gpt-oss-20b it preserves 95 to 100% of J64's predictive gain. The readout supports test-time decisions at two temporal resolutions. Over completed candidate sets, J64 and R64 improve single-branch selection, and R64-weighted voting improves plain majority voting in seven of eight settings. During generation, rolling readout windows drive a cumulative stop-and-resample policy whose operating point is fixed on training questions alone. J64 improves accuracy by 1.1 to 5.9 points over a sibling-permuted control, and the routing-only R64 proxy retains 0.9 to 3.2 of those points. Finally, router edits aimed at the mechanism J64 names induce the predicted reasoning behaviors and shift a diagnosed stall from numerical guessing toward exact symbolic execution. Together, J64 makes latent process state readable, while routing makes it deployable and actionable.

cs.AI↗

arXiv2026

Disagree to Explore, Agree to Commit: Routing-Guided Test-Time Scaling for Software Agents

Software-engineering agents solve repository-level tasks through long, stochastic tool-use trajectories, and repeated attempts often find fixes missed by one run. Test-time scaling is difficult because patches lack canonical answer forms, while sibling actions from a shared prefix are correlated. We study whether native MoE router traces can guide steering and selection without an external judge or selection-time test execution. Our analysis shows that routing provides a robust behavioral role signal; token-granular readouts and decision-matched comparison sets turn it into effective control. We therefore introduce Risa (Routing-Informed Steering and Arbitration): within trajectories, routing encourages diverse exploration and controlled convergence during patch commitment; across separately sampled trajectories, agreement at informative patch positions selects a final candidate. We evaluate on SWE-bench Verified using open-weight sparse MoE agents across scales and reasoning-effort settings. Risa's routing arbitration raises the macro-average resolved rate from 44.9% under uniform sampling to 48.2% on the gpt-oss family, matching text consensus without answer-string matching, and it transfers to Qwen3.6, where it improves on uniform choice and matches text consensus on the full 500-task benchmark.

cs.AI↗

arXiv2026

Cluster Deletion is as Hard to Approximate as Vertex Cover

Recent breakthroughs in Cluster Editing have motivated attempts to adapt these approaches to obtain better-than-$2$ approximations for Cluster Deletion. We rule out this possibility under the Unique Games Conjecture: Cluster Deletion is NP-hard to approximate within a factor of $2-ε$ for every fixed $ε>0$, matching the known $2$-approximation [Veldt et al., WWW 2018]. Our approximation-preserving reduction from Vertex Cover also implies NP-hardness of approximation within $\sqrt2-ε$. We also show that better-than-$2$ approximations are possible in restricted settings. We close the paper with a brief discussion of the relationship between Cluster Editing and Bad Triangle Transversal. In particular, we give a $31$-vertex graph~$G$ for which the two optimal values differ, answering an open question of Adriaens and Tatti [ICML 2026].

cs.DS↗

arXiv2026

DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes

Reinforcement learning has become a central paradigm for advancing reasoning in large language models, yet most existing methods still depend on stronger teacher models or heavily curated difficult datasets, limiting scalable capability improvement. In this paper, we introduce DenoiseRL, a reinforcement learning framework that substitutes external supervision with recovery-oriented optimization over failures from weak models. Instead of relying on stronger supervision or carefully engineered data, DenoiseRL learns directly from noisy reasoning prefixes by converting them into opportunities for improvement, while exercising fine-grained control over the noise intensity, making training more scalable and effective. This yields a richer and more diverse learning signal, improving exploration efficiency by leveraging imperfect model behavior. Empirically, DenoiseRL consistently outperforms strong on-policy RL baselines across competitive mathematical reasoning tasks and interactive decision-making tasks, while promoting stronger self-corrective behavior as training difficulty increases, highlighting an effective and scalable pathway for improving agentic reasoning capabilities of large language models.

cs.AI↗

arXiv2026

ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders

The recent emergence of vibe-coding workflows is changing what coding agents are expected to do. Instead of merely completing code under fully specified instructions, agents are increasingly expected to transform incomplete product intent into working software by combining various abilities including planning, requirement clarification, tool use, debugging, and repository-level construction. Yet existing benchmarks have not fully caught up with this shift, evaluating agents on static, fully specified tasks. In this paper, we introduce ICAE-Bench, a benchmark for evaluating coding agents under interactive project-building settings. The basic idea is to start from a fuzzy product requirement, simulating the dynamic paradigm with an automated User Agent. To make this setting both realistic and evaluable, ICAE-Bench introduces three key designs. First, to avoid the ambiguity of unconstrained fuzzy requirements, each task derives ambiguity from a precise real open-source repository with executable behavior. Second, to ensure high-quality and reproducible user simulation, ICAE-Bench grounds interaction through User Agent Data, allowing the User Agent to reveal hidden constraints without inventing new requirements or leaking implementation artifacts. Third, to evaluate open-ended repositories fairly, ICAE-Bench uses standardized black-box tests together with multi-dimensional diagnostics, including functional correctness, semantic and API similarity, structural fidelity, design quality, and interaction quality.

cs.AI↗

arXiv2026

Making Image Editing Easier via Adaptive Task Reformulation with Agentic Executions

Instruction guided image editing has advanced substantially with recent generative models, yet it still fails to produce reliable results across many seemingly simple cases. We observe that a large portion of these failures stem not from insufficient model capacity, but from poorly formulated editing tasks, such as those involving small targets, implicit spatial relations, or under-specified instructions. In this work, we frame image editing failures as a task formulation problem and propose an adaptive task reformulation framework that improves editing performance without modifying the underlying model. Our key idea is to transform the original image-instruction pair into a sequence of operations that are dynamically determined and executed by a MLLM agent through analysis, routing, reformulation, and feedback-driven refinement. Experiments on multiple benchmarks, including ImgEdit, PICA, and RePlan, across diverse editing backbones such as Qwen Image Edit and Nano Banana, show consistent improvements, with especially large gains on challenging cases. These results suggest that task reformulation is a critical but underexplored factor, and that substantial gains can be achieved by better matching editing tasks to the effective operating regime of existing models.

cs.CV↗

arXiv2026

Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control

In sparse Mixture-of-Experts language models, does the same token id imply the same router state and the same experts producing it? Holding the emitted token id fixed at repeated anchors, we find it does not: the experts that produce it still separate task context, trajectory history, and reasoning-effort mode. This residual structure supports test-time control: near \emph{boundary} anchors (the final-response transition) and \emph{delimiter} anchors (which open the answer, e.g.\ \texttt{\textbackslash boxed\{} or code fences), routing neighborhoods already align with final-answer basins at a marker-only readout and strongest when the routing is read at the answer opening. We operationalize this as \textbf{RAD} (Routing Agreement Decoding), an answer-string-free multi-rollout selector: it locates a fixed anchor, represents each rollout by its anchor-window MoE routing states, and returns the densest Weighted-Jaccard $K$-NN route-basin center, without parsing, normalizing, executing, or voting over answer strings. Across 10 sparse-MoE configurations (gpt-oss, Qwen3-MoE) and 6 datasets spanning math, GPQA, and code, RAD is on par with Majority where string voting is well-posed, with small positive paired deltas (RAD $73.9$ / RAD+DC $74.2$ vs.\ Majority $73.6$). Like majority voting, RAD is not a verifier: a dense \emph{wrong} basin can still win. Its value is the interface: the same selector gives direct pass@1 on code, where exact-string voting is ill-defined, and the same routing-density principle, re-anchored to the agentic boundary, improves best-of-16 patch selection on SWE-bench Verified over random, where patches have no answer string to vote on.

cs.CL↗