Searcharxiv⌕ Search

arXiv subjects

Zhanke Zhou

Publications and source records attributed to Zhanke Zhou.

At least 19 recordsLinked to original sources

RewardFlow: Can Action-Level Reward Be Estimated from State-Graph Topology in Agentic RL?

In agentic tasks such as household assistance and web shopping, large language model (LLM) agents take multiple actions before receiving a final outcome indicating success or failure. For reinforcement learning (RL), this outcome obscures individual actions' contributions: a shopping trajectory may find the right product but fail to complete the purchase. To address this credit-assignment problem, we introduce RewardFlow, which connects multiple trajectories for the same task through shared states (states grouped by task-relevant equivalence or semantic similarity). The resulting directed state graph links observed trajectory segments. Therefore, an action in a failed trajectory may reach a state from which another trajectory succeeds. It propagates reward signals backward from states where the task succeeds, assigning state potentials that decrease with shortest-path distance to success in the graph. Potential differences then reward actions that reduce this distance, bringing the agent closer to success in the graph. These estimated rewards yield action-level advantages that complement trajectory-level advantages from outcome rewards during policy optimization, without a learned process reward model or manual action-level annotations. Experiments on ALFWorld, WebShop, Sokoban, and search tasks show gains over RL baselines across model scales. With 3B policies, graph construction and reward estimation take $1.47$-$2.39$ seconds per training step on ALFWorld, WebShop, and Sokoban (Fig. 1). Code is available at https://github.com/tmlr-group/RewardFlow.

cs.AI↗

When Credit Turns Against Tools: Transition-Aware Credit Assignment in Agentic Reinforcement Learning

Agentic reinforcement learning with verifiable rewards trains language models through multi-turn tool interactions, yet their reasoning performance and tool-calling ability can decline during training. We trace this degradation to three coupled mechanisms: (1) tool-token imbalance produces negative overall tool credit when incorrect trajectories contain more tool tokens per turn than correct ones; (2) aggregation amplification can further strengthen this negative signal as token-mean aggregation gives longer incorrect trajectories greater weight in policy updates; and (3) failure-driven retries prolong trajectories as models tend to call tools again after execution errors, extending potential negative credit to recovery attempts. To counter this suppression, we propose Transition-Aware Credit Assignment (TACA), which targets the transition from reasoning to tool use through supervision of both call initiation and generation. Specifically, TACA uses Advantage Projection to enforce a nonnegative mean advantage on tool-call tokens within each prompt group through a minimal uniform shift. Additionally, TACA guides tool-use decisions with Decision Contrast, comparing rewards for tool-assisted and direct answers to the same prompt. Through Supervised Replay, it reinforces tool-call generation using calls from previously sampled correct trajectories. Across models and tasks, TACA improves performance and sustains more reliable tool interaction during training, raising mean accuracy over MT-GRPO on six reasoning benchmarks by 4.61 percentage points for Qwen2.5-1.5B and 3.61 for Qwen2.5-3B. Our code is available at https://github.com/tmlr-group/TACA.

cs.LG↗

Spatial Harness: Grounding Spatial Reasoning in Executable World States with Physics Engines

Spatial reasoning requires understanding how objects and observers are arranged in a scene and how their relations change with viewpoint or movement. Vision-language models (VLMs) typically perform both steps through model-internal reasoning, forming an implicit scene interpretation from partial views and imagining the consequences of spatial transformations. A misread scene can lead to a wrong answer, while even a correct interpretation does not ensure that imagined transformations satisfy spatial constraints. We introduce Spatial Harness, which represents scenes as explicit and executable world states to ground spatial reasoning in physics-engine feedback. A Perceiver integrates scene constraints extracted from the question with specialized visual estimates to place objects and cameras in a shared coordinate frame. The resulting state is instantiated in a physics engine, where a Reasoner executes viewpoint changes and uses computed measurements and rendered views to guide reasoning. We evaluate on MindCube, SAT, MMSI-Bench, and SpatialRGPT-Bench with open- and closed-weight VLMs. On MindCube, Spatial Harness improves Qwen3.5-27B from 52.70% to 71.14%, surpassing pySpatial by 7.52 percentage points. Given identical annotated object and camera poses, it reaches 98.25% versus 71.81% for model-internal reasoning. Code is publicly available at: https://github.com/tmlr-group/Spatial-Harness

cs.LG↗

AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks

Agents pair reasoning models with harnesses that organize model calls, execute tools, load skill instructions, and retain history in memory. Understanding agent performance requires linking execution conditions to outcomes, behavior, and resource use. We introduce AlphaDiana, a measurement system with modular interfaces for datasets, agents, environments, and scorers. It preserves harness-native control loops while controlling file reuse and memory retention separately. AlphaDiana supports macro analyses of complete agents and micro interventions on agent components such as tools, skills, and memory under matched conditions. We study three open-source models and three open-source harnesses on five answer-based benchmarks, with two models also tested on two interactive benchmarks. We find that harnesses do not consistently improve the accuracy of reasoning: gains on questions unsolved by the model alone can be outweighed by losses on questions it answers correctly without tools. This inconsistency extends to component interventions: neither skill guidance nor retained history reliably improves accuracy. Behavior and output-token use must be assessed with correctness: low token entropy can accompany wrong answers, and less output per execution can mean more output tokens per correct answer when accuracy falls. AlphaDiana links answer correctness and execution status to traces and resource records for comparing agents and diagnosing failures under shared task and resource policies.

cs.AI↗

Bayes-Sufficient Compression Is Not Enough: How Does Communication Help Multi-Agent Systems?

Multi-agent LLM systems pair a sender with broad context and an executor with a limited local view. We study when a short message improves the executor's next decision, when raw context is preferable, and when a stronger sender helps. Our framework, \emph{receiver-relative bounded coordination}, expresses message utility as receiver gain minus protocol tax. Compression beats raw context when tax savings exceed losses from omitted information and decoder mismatch. Even \emph{Bayes-sufficient} compression can fail when a bounded executor cannot use its surface form. A three-stage decomposition separates externalization, absorption, and \emph{action closure}, explaining how errors remain after the correct content reaches the receiver. Under a single-crossing condition, sender upgrades help above a receiver-burden threshold. Across six benchmarks, the same Qwen protocol raises ContextBench joint accuracy from $0.633$ to $0.775$ but lowers ToolSandbox from $0.889$ to $0.653$. Fixed-message replay reveals closure failures despite correct artifact recovery. These results guide an inference-time selector that improves the accuracy-cost frontier on the evaluated communication regimes.

cs.LG↗

Model Inversion Attacks: A Survey of Approaches and Countermeasures

Deep neural networks have enabled numerous studies and applications on both Euclidean data, such as images and text, and non-Euclidean data, such as graphs. Because these networks may process private data, their deployment raises concerns about privacy leakage. Model inversion attacks (MIAs) exploit access to a trained model to reconstruct training examples or infer privacy-sensitive characteristics represented by the model. The effectiveness of MIAs has been demonstrated in various domains, including images, text, and graphs. These attacks highlight the vulnerability of neural networks and raise awareness about the risk of privacy leakage within the research community. This survey provides a threat-model- and assumption-aware synthesis of attacks and defenses. We compare exposed interfaces, attacker knowledge, reconstruction priors and targets, failure modes, deployment constraints, and privacy-utility trade-offs, while highlighting modeling principles, optimization challenges, and future directions. We also maintain an evolving repository of relevant research at https://github.com/AndrewZhou924/Awesome-model-inversion-attack.

cs.LG↗

DICE: Entropy-Regularized Equilibrium Selection for Stable Multi-Agent LLM Coordination

Multi-agent large language model (LLM) systems often fail to reliably outperform a single strong model equipped with best-of-N sampling. We argue that a core source of this instability is ill-posed equilibrium selection: current systems specify what information agents share, but not which coordination convention should be selected. We formalize a broad class of such systems as discounted incomplete-information Markov games and show that two common pathologies, oscillation between competing conventions and drift across them, can both induce unstable learning and linear Bayesian regret. To obtain a well-posed target, we introduce the Heterogeneous Quantal Response Equilibrium (HQRE), an entropy-regularized equilibrium concept with agent- and state-dependent temperatures. Under a monotonicity condition, HQRE is unique, admits linearly convergent mirror updates, and yields bounded Bayesian regret; the same condition yields rollout-measurable stability diagnostics. We instantiate this objective in two algorithms: DICE-PC, which coordinates frozen models through prompt-control actions, and DICE-FT, which performs parameter-efficient mirror fine-tuning. Across eleven benchmarks in four domains, DICE improves accuracy-cost trade-offs over strong within-class baselines; on reasoning and planning tasks, DICE-PC improves by 4.3 percentage points on average and DICE-FT by 8.5 points.

cs.LG↗

The Easy, the Hard, and the Learnable: Confidence and Difficulty-Adaptive Policy Optimization for LLM Reasoning

RL with verifiable rewards can substantially improve LLM reasoning, yet standard GRPO-style training often treats easy, hard, and learnable questions alike through uniform sampling and weighting, leading to inefficient compute allocation. We study GRPO by tracking token log-probabilities, group-normalized advantages, and the induced token-level update weights. This reveals three recurring dynamics as training proceeds: (1) confidence inflation, (2) advantage contraction, and (3) hierarchical convergence. These findings suggest that the utility of each update depends strongly on both question difficulty and the model's current competence. Motivated by this, we propose Confidence and Difficulty-adaptive Policy Optimization (CoDaPO), which assigns each question a bounded value from rollout confidence and empirical difficulty. CoDaPO then uses this value to reweight policy updates and resample high-value learnable questions within mini-batches, thereby increasing discovery within the learnable band under a fixed compute budget. Across twelve benchmarks, CoDaPO consistently improves accuracy over existing RL methods. Our code is publicly available at https://github.com/tmlr-group/CoDaPO.

cs.LG↗

Beyond Homophily: Towards Generalized Graph Reconstruction Attack and Defense

Graph neural networks (GNNs) are widely deployed on relational data, yet they can leak sensitive or proprietary information about the training graph adjacency, e.g., social ties, transactions, and interactions. This work studies graph reconstruction attacks (GRA), a form of model inversion that reconstructs the training adjacency from a trained GNN, given different levels of attacker-side information. We first provide a systematic characterization of when and why adjacency becomes recoverable through features, labels, embeddings, and predictions, with leakage modulated by graph homophily, heterophily, and the model's inductive bias. Motivated by these findings, we view GNN inference through a Markov chain approximation lens, treating the layered forward computation as a chain of topology-dependent representations. Building on this view, we develop complementary attack and defense methods. On the attack side, we propose MC-GRA (+), which reconstructs the adjacency by optimizing a surrogate adjacency whose GNN-induced representations align with those of the target model at each layer. On the defense side, we propose MC-GPB (+), which suppresses adjacency-dependent information throughout the representation chain while aiming to preserve classification accuracy under a privacy-utility trade-off. Experiments across homophilic/heterophilic graph benchmarks and GNNs show that our attacks improve reconstruction fidelity over prior methods, while our defenses reduce reconstruction success with only minor accuracy loss.

cs.LG↗

Deliberate Evolution: Agentic Reasoning for Sample-Efficient Symbolic Regression with LLMs

Symbolic regression (SR) discovers compact mathematical expressions from data, yet recent LLM-based evolutionary methods remain sample-inefficient because they rely mainly on scalar feedback such as MSE. We identify a core limitation: existing methods conflate candidate proposal with search guidance, requiring the LLM to infer how to evolve an expression, diagnose its errors, and reuse past experience from a single score. To address this, we propose Deliberate Evolution (DE), an agentic framework that decouples symbolic generation from search control. DE guides LLM proposals with adaptive operators for search direction, analytical tools for structural diagnosis, and reflective memory for trajectory-level experience. Experiments on LLM-SRBench show that DE consistently outperforms representative LLM-based SR baselines across diverse scientific domains while using only 40% of the standard sample budget.

cs.CL↗

What Do Evolutionary Coding Agents Evolve?

Recent work pairs LLMs with evolutionary search to iteratively generate, modify, and select code using task-specific feedback. These systems have produced strong results in mathematical discovery and algorithm design, yet a fundamental question remains: what do they actually evolve? Progress is typically summarized by the best score a run reaches under a task-specific evaluator, but that score can reflect several different mechanisms: new algorithmic structure, re-tuning an existing strategy, recombining ideas already in the model's internal knowledge, or overfitting to the evaluator. Distinguishing these mechanisms requires inspecting the search process itself, not only its final outcome. We introduce EvoTrace, a dataset of evolutionary coding traces spanning four evolutionary frameworks, reasoning and non-reasoning models, and 16 tasks across mathematics and algorithm design. To analyze these traces, we develop EvoReplay, a replay-based methodology that reconstructs the local search states behind high-scoring solutions and tests controlled interventions, including adjusting constants, removing program components and substituting models or prompting contexts. We annotate every code edit in EvoTrace with one of nine recurring edit types using an LLM-as-judge pipeline validated against blind human re-annotation. Across EvoTrace, most score gains come from a small subset of these edit types. We further find a deterministic cycling pattern: about 30% of code lines added during search are byte-identical re-introductions of previously-deleted lines, present throughout nearly every run. These results show that benchmark gains in evolutionary coding agents can arise from qualitatively different mechanisms, only some of which correspond to new algorithmic structure. EvoTrace enables more diagnostic evaluation of evolutionary coding agents beyond final benchmark scores.

cs.NE↗

Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models

While reinforcement learning with verifiable rewards (RLVR) is effective to improve the reasoning ability of large language models (LLMs), its reliance on human-annotated labels leads to the scaling up dilemma, especially for complex tasks. Recent self-rewarding methods investigate a label-free alternative to unlock the reasoning capabilities of LLMs, yet they frequently encounter the non-negligible training collapse issue, as the single-view supervision signal easily forms the self-consistent illusion, yielding the reward hacking. Inspired by the success of self-supervised learning, we propose \textit{Co-rewarding}, a novel self-supervised RL framework that improves training stability by seeking complementary supervision from another views. Specifically, we instantiate Co-rewarding in two ways: (1) \textit{Co-rewarding-I} is a data-side instantiation that derives reward signals from contrastive agreement across semantically analogous questions; and (2) \textit{Co-rewarding-II} is a model-side instantiation that maintains a slowly-updated reference teacher with pseudo labels to realize self-distillation. Intuitively, such instantiations introduce different levels of discrepancy to increase the difficulty of training collapse on trivial reasoning solutions. We also explore their orthogonally combined version to further boost the performance. Empirically, Co-rewarding exhibits stable training across various setups, and outperforms other self-rewarding baselines by $+3.31\%$ improvements on average on multiple mathematical reasoning benchmarks, especially by $+7.49\%$ on Llama-3.2-3B-Instruct. Notably, Co-rewarding reaches or even surpasses RLVR with ground-truth (GT) label in several cases, such as a Pass@$1$ of $94.01\%$ on GSM8K with Qwen3-8B-Base remarkably higher than GT. Our code is released at https://github.com/tmlr-group/Co-rewarding.

cs.LG↗

AlphaApollo: A System for Deep Agentic Reasoning

We present AlphaApollo, an agentic reasoning system that targets two bottlenecks in foundation-model reasoning: (1) limited reasoning capacity for complex, long-horizon problem solving and (2) unreliable test-time evolution without trustworthy verification. AlphaApollo orchestrates models and tools via three components: (i) multi-turn agentic reasoning, which formalizes model-environment interaction with structured tool calls and responses; (ii) multi-turn agentic learning, which applies turn-level reinforcement learning to optimize tool-use reasoning while decoupling actions from tool responses for stable training; and (iii) multi-round agentic evolution, which refines solutions through a propose-judge-update loop with tool-assisted verifications and long-horizon memory. Across seven math reasoning benchmarks and multiple model scales, AlphaApollo improves performance through reliable tool use (> 85% tool-call success), substantial gains from multi-turn RL (Avg@32: Qwen2.5-1.5B-Instruct 1.07% -> 9.64%, Qwen2.5-7B-Instruct 8.77% -> 20.35%), and improvements from evolution (e.g., Qwen2.5-3B-Instruct 5.27% -> 7.70%, Qwen2.5-14B-Instruct 16.53% -> 21.08%). This project is still ongoing. We welcome feedback from the community and will frequently update the source code and technical report.

cs.AI↗

Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning

Large language models (LLMs) benefit substantially from supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) in reasoning tasks. However, these recipes perform poorly in instruction-based molecular optimization, where each data point typically provides only a single optimized reference molecule and no step-by-step optimization trajectory. We reveal that answer-only SFT on the reference molecules collapses reasoning, and RLVR provides sparse feedback under similarity constraints due to the model's lack of effective exploration, which slows learning and limits optimization. To encourage the exploration of new molecules while balancing the exploitation of the reference molecules, we introduce Reference-guided Policy Optimization (RePO), an optimization approach that learns from reference molecules without requiring trajectory data. At each update, RePO samples candidate molecules with their intermediate reasoning trajectories from the model and trains the model using verifiable rewards that measure property satisfaction under similarity constraints in an RL manner. Meanwhile, it applies reference guidance by keeping the policy's intermediate reasoning trajectory as context and training only the answer in a supervised manner. Together, the RL term promotes exploration, while the guidance term mitigates reward sparsity and stabilizes training by grounding outputs to references when many valid molecular edits exist. Across molecular optimization benchmarks, RePO consistently outperforms SFT and RLVR baselines (e.g., GRPO), achieving improvements on the optimization metric (Success Rate $\times$ Similarity), improving balance across competing objectives, and generalizing better to unseen instruction styles. Our code is publicly available at https://github.com/tmlr-group/RePO.

cs.LG↗

Landscape of Thoughts: Visualizing the Reasoning Process of Large Language Models

Numerous applications of large language models (LLMs) rely on their ability to perform step-by-step reasoning. However, the reasoning behavior of LLMs remains poorly understood, posing challenges to research, development, and safety. To address this gap, we introduce landscape of thoughts (LoT), the first landscape visualization tool to inspect the reasoning trajectories with certain reasoning methods on any multi-choice dataset. We represent the textual states in a trajectory as numerical features that quantify the states' distances to the answer choices. These features are then visualized in two-dimensional plots using t-SNE. Qualitative and quantitative analysis with the landscape of thoughts effectively distinguishes between strong and weak models, correct and incorrect answers, as well as different reasoning tasks. It also uncovers undesirable reasoning patterns, such as low consistency and high uncertainty. Additionally, users can adapt LoT to a model that predicts the property they observe. We showcase this advantage by adapting LoT to a lightweight verifier that evaluates the correctness of trajectories. Empirically, this verifier boosts the reasoning accuracy and the test-time scaling effect. The code is publicly available at: https://github.com/tmlr-group/landscape-of-thoughts.

cs.LG↗

SATBench: Benchmarking LLMs' Logical Reasoning via Automated Puzzle Generation from SAT Formulas

We introduce SATBench, a benchmark for evaluating the logical reasoning capabilities of large language models (LLMs) through logical puzzles derived from Boolean satisfiability (SAT) problems. Unlike prior work that focuses on inference rule-based reasoning, which often involves deducing conclusions from a set of premises, our approach leverages the search-based nature of SAT problems, where the objective is to find a solution that fulfills a specified set of logical constraints. Each instance in SATBench is generated from a SAT formula, then translated into a puzzle using LLMs. The generation process is fully automated and allows for adjustable difficulty by varying the number of clauses. All 2100 puzzles are validated through both LLM-based and solver-based consistency checks, with human validation on a subset. Experimental results show that even the strongest model, o4-mini, achieves only 65.0% accuracy on hard UNSAT problems, close to the random baseline of 50%. Our error analysis reveals systematic failures such as satisfiability bias, context inconsistency, and condition omission, highlighting limitations of current LLMs in search-based logical reasoning. Our code and data are publicly available at https://github.com/Anjiang-Wei/SATBench

cs.AI↗

From Debate to Equilibrium: Belief-Driven Multi-Agent LLM Reasoning via Bayesian Nash Equilibrium

Multi-agent frameworks can substantially boost the reasoning power of large language models (LLMs), but they typically incur heavy computational costs and lack convergence guarantees. To overcome these challenges, we recast multi-LLM coordination as an incomplete-information game and seek a Bayesian Nash equilibrium (BNE), in which each agent optimally responds to its probabilistic beliefs about the strategies of others. We introduce Efficient Coordination via Nash Equilibrium (ECON), a hierarchical reinforcement-learning paradigm that marries distributed reasoning with centralized final output. Under ECON, each LLM independently selects responses that maximize its expected reward, conditioned on its beliefs about co-agents, without requiring costly inter-agent exchanges. We mathematically prove that ECON attains a markedly tighter regret bound than non-equilibrium multi-agent schemes. Empirically, ECON outperforms existing multi-LLM approaches by 11.2% on average across six benchmarks spanning complex reasoning and planning tasks. Further experiments demonstrate ECON's ability to flexibly incorporate additional models, confirming its scalability and paving the way toward larger, more powerful multi-LLM ensembles. The code is publicly available at: https://github.com/tmlr-group/ECON.

cs.LG↗

From Passive to Active Reasoning: Can Large Language Models Ask the Right Questions under Incomplete Information?

While existing benchmarks probe the reasoning abilities of large language models (LLMs) across diverse domains, they predominantly assess passive reasoning, providing models with all the information needed to reach a solution. By contrast, active reasoning-where an LLM must interact with external systems to acquire missing evidence or data-has received little systematic attention. To address this shortfall, we present AR-Bench, a novel benchmark designed explicitly to evaluate an LLM's active reasoning skills. AR-Bench comprises three task families-detective cases, situation puzzles, and guessing numbers-that together simulate real-world, agentic scenarios and measure performance across commonsense, logical, and symbolic reasoning challenges. Empirical evaluation on AR-Bench demonstrates that contemporary LLMs exhibit pronounced difficulties with active reasoning: they frequently fail to acquire or leverage the information needed to solve tasks. This gap highlights a stark divergence between their passive and active reasoning abilities. Moreover, ablation studies indicate that even advanced strategies, such as tree-based searching or post-training approaches, yield only modest gains and fall short of the levels required for real-world deployment. Collectively, these findings highlight the critical need to advance methodology for active reasoning, e.g., incorporating interactive learning, real-time feedback loops, and environment-aware objectives for training. The benchmark is publicly available at: https://github.com/tmlr-group/AR-Bench.

cs.LG↗