arXiv subjects

Ke Zeng

Publications and source records attributed to Ke Zeng.

At least 19 recordsLinked to original sources

arXiv2026

CARE: Contrastive Anchor-based Rubric Evolution for Large Language Model Post-Training

Rubric-based reinforcement learning decomposes open-ended instructions into prompt-specific, flexible rubrics, making it better suited than reinforcement learning with verifiable rewards for post-training LLMs on open-ended tasks. However, static rubrics are inevitably hacked as the policy evolves, and existing dynamic approaches introduce new problems: undirected rubric extraction, unreliable hack detection, and unbounded rubric proliferation. We propose $\textbf{CARE}$ ($\textbf{C}$ontrastive $\textbf{A}$nchor-based $\textbf{R}$ubric $\textbf{E}$volution), which grounds every rubric evolution step in a high-quality anchor response generated by a frontier model conditioned on the prompt and its rubrics. At each training step, CARE contrasts the highest-scoring rollout against the anchor, enabling two complementary mechanisms: an Adaptive branch that reactively repairs reward misspecification; and a Chase branch that proactively converts frontier-level quality gaps into sharper rubrics. Together, the two branches $\textbf{maintain discriminative accuracy in the high-reward region}$---the precise region where reward over-optimization mostly originates. Experiments on WildChecklist-9K with Qwen2.5-7B-Base and Qwen2.5-7B-Instruct show that CARE achieves state-of-the-art performance on Arena-Hard-2.0, InfoBench, and FollowBench, and is the $\textbf{only}$ method whose win rate against GPT-4.1 anchor responses shows sustained improvement throughout 300 training steps; additional results on Llama-3.1-8B-Instruct and Qwen3-8B further indicate that CARE generalizes across model families.

cs.AI↗

arXiv2026

Uncovering and Mitigating Aggregation-Induced Reward Hacking in Multi-Reward Reinforcement Learning

Reinforcement learning fine-tuning of large language models increasingly adopts multiple reward dimensions, including verifiable rules, task-specific evaluators, and learned reward models, to provide richer supervision across diverse capabilities. These dimensions are commonly scalarized with fixed aggregation weights. We identify a failure mode in which aggregation itself induces reward hacking: static projection aliases qualitatively different reward profiles into a single scalar, steering optimization toward whichever dimensions are easiest, densest, or systematically favored by the reward signal. Over training, this traps the policy in suboptimal profiles and prevents convergence to better-balanced ones that would yield higher task performance. To address this, we propose Adaptive Multi-Reward Projection (AMRP), a lightweight online method that reallocates aggregation weights using three signals, relative shortfall, reward volatility, and recent progress, increasing pressure on lagging, unstable, or stagnant dimensions while relieving saturated ones. Across structured reasoning, citation-grounded generation, and open-ended alignment under GRPO, AMRP consistently improves reward-profile balance and downstream performance over fixed and dynamic weighting baselines; it also remains effective with GDPO and PPO, supporting compatibility across RL algorithms. Our code is available at https://github.com/yyhappier/AMRP.git.

cs.CL↗

arXiv2026

Retrieval Grounding Latent Reasoning for Dense Retrieval

Reasoning-intensive retrieval requires text representations to capture not only semantic similarity, but also the reasoning needed to determine relevance under a given retrieval instruction. Existing reasoning-enhanced embedding models improve retrieval by incorporating reasoning information into dense representations, yet their supervision is typically dominated by the final retrieval objective. As a result, latent reasoning trajectories may learn shortcut reasoning patterns that preserve retrieval performance without producing meaningful incremental retrieval gains. We propose Retrieval Grounding Latent Reasoning (RGLT), a latent reasoning framework for dense retrieval that explicitly connects intermediate latent transitions with retrieval improvements. RGLT performs non-autoregressive reasoning in hidden space through an instruction-conditioned latent reasoning trajectory constructed from silent tokens. It combines process-supervised explicit-to-implicit distillation with retrieval-grounded supervision, using stage-wise CoT reconstruction to shape intermediate latent states and retrieval-effect credit to optimize incremental retrieval gains across the latent reasoning trajectories. Experiments on reasoning-intensive retrieval benchmarks show that RGLT consistently outperforms strong baselines while preserving efficient embedding inference.

cs.AI↗

arXiv2026

ToolVerse: Unlocking Massive Environments and Long-Horizon Tasks for Agentic Reinforcement Learning

While LLM agents demonstrate strong reasoning abilities in compact and well-defined scenarios, they struggle to maintain robustness and effectiveness when faced with large-scale, diverse, and dynamic real-world environments that demand seamless tool integration. To address this gap, we introduce ToolVerse, a comprehensive framework that scales up agentic RL environments and enables agents to perform complex long-horizon reasoning in Tool-Integrated Reasoning (TIR) tasks. First, ToolVerse automatically builds the massive executable agent training environments from nearly 400 real-world Model Context Protocols (MCPs) that contain about 4500 tools. Second, we propose a task design strategy based on a tool dependency graph, utilizing Dynamic Unlocking Sampling Algorithm to generate long-horizon tasks, and produce GUST (Graph Unlocking Sampling Tasks) dataset. Third, to alleviate the credit assigment problem in long-horizon agentic RL, we propose a fine-grained Turn-Aware Relative Advantage algorithm. We conduct extensive Agentic RL training using ToolVerse and evaluate our framework on serveral agentic benchmarks. Experimental results demonstrate that our framework significantly strengthens LLMs' capabilities in long-horizon tool use, achieving a marked performance boost and showcasing robust reasoning within dynamic environments.

cs.AI↗

arXiv2026

BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents

Stepwise group-based RL is an attractive way to train long-horizon LLM agents without a learned critic: it reuses multiple sampled rollouts to estimate local advantages. Its weakness is less visible but more fundamental: every group-relative estimator assumes that the steps it compares are equivalent for credit assignment. We show that current agentic variants violate this assumption through a state-action credit mismatch. The observation-hash partition is overly fine on the state side, creating singleton groups with zero step-level signal, while a single within-group mean is too coarse on the action side, mixing state-value estimation with action-specific credit. We introduce BiPACE (Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation), a drop-in advantage estimator that fixes both sides without adding a critic, auxiliary loss, or extra rollouts. BiGPO clusters steps by cosine distance in the actor's own hidden-state geometry, an empirical policy-induced proxy for bisimulation that substantially lowers the singleton rate left by observation hashing. PACE then recenters returns within each behavioral cluster using action-conditioned peer baselines; its Q-style instance estimates a local Q(s,a)-V(s) nonparametrically. On ALFWorld/Qwen2.5-7B, BiPACE_Q raises overall validation success from GiGPO's 90.8 to $97.1\pm0.9$ over three seeds, and crosses the 95% threshold on every seed, which GiGPO never does within the same budget. On Qwen2.5-1.5B it reaches $93.5\pm1.2$ versus GiGPO's 86.7, and on WebShop and TextCraft it improves over GRPO and GiGPO at both model scales. The measured BiPACE-specific overhead is 11.3% of a single training-step wall time. Yet it changes the estimator's comparison unit from surface identity to approximate behavioral equivalence plus action-side counterfactuals. The code is available at https://github.com/TianxiangZhao/BiPACE.

cs.CL↗

arXiv2026

Achieving Heisenberg limit under noisy conditions with quantum Zeno dynamics and dynamical decoupling

Quantum Zeno dynamics (QZD) and dynamical decoupling (DD) are useful tools that enable the effective suppression of noise in quantum systems. We consider the problem of when (i) noise can be suppressed and (ii) Heisenberg limit (HL) can be achieved in quantum metrology, and prove necessary and sufficient conditions for when QZD and DD are useful for achieving these two goals. We further show this condition can hold even when the Hamiltonian-not-in-Lindblad-span (HNLS) condition fails, allowing QZD/DD to recover Heisenberg scaling beyond what is certified by the usual HNLS-based quantum error correction (QEC) criterion. Finally, we demonstrate that the combination of both techniques can allow individually imperfect QZD and DD strategies to saturate HL.

quant-ph↗

arXiv2026

STAGE-Claw: Automated State-based Agent Benchmarking for Realistic Scenarios

Large language models are increasingly used to power personal agents for everyday applications, but evaluating these agents remains a challenge. Existing benchmarks still rely on sandboxed artifacts, static task design, and coarse scoring, which hinder scalability and limit progress toward reliable personal-agent evaluation. This paper introduces STAGE-Claw, an automated framework for building and evaluating realistic personal-agent scenarios in state-based personal-computing environments. Given a task hint, STAGE-Claw automatically creates and validates a realistic benchmark task with its environment, task prompts, ground truth, and related verification programs. Agents are then evaluated in realistic operating environments, where performance is measured by the correctness of the final system state rather than only the textual response. Using STAGE-Claw, this paper creates a benchmark with 40 challenging real scenario agent tasks, evaluates 11 frontier models, and analyzes their task scores, costs, tool-call reliability, and common failure patterns. Overall, STAGE-Claw offers a scalable, state-based way to evaluate agents in realistic user scenarios.

cs.AI↗

arXiv2026

Rethinking Continual Experience Internalization for Self-Evolving LLM Agents

Experience internalization converts contextual experience from past interactions into reusable parametric capability, offering a promising path toward continual learning in large language models (LLMs). While prior work has predominantly focused on single-iteration transfer, we discover that under multi-iteration experience learning, existing methods suffer from a progressive capability collapse rather than compounding improvement. We systematically examine this failure through three vital dimensions of experience internalization: (1) Experience Granularity: We find that principle-level experience is more durable than instance-level experience, as it effectively abstracts transferable strategies away from trajectory-specific details. (2) Experience Injection Pattern: Our analysis reveals that step-wise injection significantly outperforms global injection by aligning experience with intermediate decision states, a property that is critical for long-horizon tool use. (3) Internalization Regime: We demonstrate that off-policy context-distillation on high-quality teacher trajectories provides a substantially more stable training signal than on-policy context-distillation, which is inherently limited by local corrections on student-induced flawed states. Together, these insights yield a simple yet robust recipe for stable and sustainable experience internalization, providing concrete guidance for engineering self-evolving and continually learning LLMs.

cs.CL↗

arXiv2026

SIRI: Self-Internalizing Reinforcement Learning with Intrinsic Skills for LLM Agent Training

Long-horizon LLM agents can benefit from reusable skills, yet existing skill-based methods often rely on external skill generators during training or persistent skill retrieval at inference, increasing engineering complexity, context length, and deployment latency. We propose Self-Internalizing Reinforcement learning with Intrinsic skills (SIRI), a three-phase framework that enables agents to discover, validate, and internalize skills without external skill generators or inference-time skill banks. SIRI first warms up the policy with GiGPO to acquire basic interaction ability and collect successful skill-free trajectories. It then performs self-skill mining, where the current policy summarizes compact skills from its own successful plain rollouts and validates them through paired skill-augmented and skill-free rollouts. Finally, SIRI distills only beneficial skill-guided action tokens into the plain policy using trajectory-level utility and action-level advantage. At inference, the agent runs with the original prompt only. On ALFWorld and WebShop with Qwen2.5-7B-Instruct, SIRI improves GiGPO from 0.908 to 0.930 on ALFWorld and from 0.728 to 0.813 on WebShop, outperforming prompt-based, RL-based, and memory-augmented baselines. Further analysis shows that our self-mining strategy can achieve performance comparable to distillation with closed-source large model. Our code is available at https://github.com/kirito618/SIRI.

cs.AI↗

arXiv2026

Skill or Skip? Learning Selective Skill Invocation in Agentic Tasks via Dual-Granularity Preference Learning

Agent skills are callable procedural modules that provide reusable knowledge and execution policies for complex agentic tasks. However, existing methods mainly focus on selecting relevant skills or improving the skills themselves, while overlooking whether a relevant skill should actually be invoked at the current decision point. Unhelpful invocations may introduce irrelevant context and disrupt an otherwise correct execution process. To address this issue, we propose SelSkill, a dual-granularity preference-learning framework for selective skill invocation. SelSkill formulates skill use as a skill-or-skip decision, uses predictive uncertainty to prioritize candidate decision points, and constructs controlled invoke-skip preference pairs from shared trajectory prefixes. It further combines episode-level outcome preferences with step-level invocation preferences to capture both overall trajectory quality and the local effectiveness of skill invocation. On ALFWorld with Qwen3-8B, SelSkill improves task success by 10.9 points over the skill-enabled baseline and 7.8 over No-Skill, with 29.1-point higher execution precision. On BFCL, task success and execution precision improve by 5.7 and 29.5 points, respectively. Zero-shot results on Tau-bench and PopQA suggest partial transfer to unseen domains and skills.

cs.CL↗

arXiv2026

Asymmetric On-Policy Distillation: Bridging Exploitation and Imitation at the Token Level

On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy distillation and standard reinforcement learning. However, we find that its standard advantage weighted policy gradient suffers from three structural weaknesses, including high variance updates, vanishing gradients in zero-advantage regions, and exploration bottlenecks when corrective signals are insufficient. We therefore propose Asymmetric On-Policy Distillation (AOPD), which replaces ineffective negative reinforcement with localized divergence minimization in non-positive advantage regions while preserving positive reinforcement learning. Experiments on mathematical reasoning benchmarks show that AOPD consistently outperforms standard OPD, with average gains of 4.09 / 8.34 under strong / weak initialization, respectively. AOPD also maintains higher policy entropy during training and better capability retention during sequential tool-use adaptation.

cs.LG↗

arXiv2026

Purifying Multimodal Retrieval: Fragment-Level Evidence Selection for RAG

Multimodal Retrieval-Augmented Generation (MRAG) is widely adopted for Multimodal Large Language Models (MLLMs) with external evidence to reduce hallucinations. Despite its success, most existing MRAG frameworks treat retrieved evidence as indivisible documents, implicitly assuming that all content within a document is equally informative. In practice, however, sometimes only a small fraction of a document is relevant to a given query, while the remaining content introduces substantial noise that may lead to performance degradation. We address this fundamental limitation by reframing MRAG as a fine-grained evidence selection problem. We propose Fragment-level Evidence Selection for RAG (FES-RAG), a framework that selects atomic multimodal fragments rather than entire documents as grounding evidence. FES-RAG decomposes retrieved multimodal documents into sentence-level textual fragments and region-level visual fragments, enabling precise identification of evidence that directly supports generation. To guide fragment selection, we introduce Fragment Information Gain (FIG), a principled metric that measures the marginal contribution of each fragment to the MLLM's generation confidence. Based on FIG, we distill fragment-level utility judgments from a high-capacity MLLM into a lightweight selector, achieving accurate evidence selection with low inference overhead. Experiments on the M2RAG benchmark show that FES-RAG consistently outperforms state-of-the-art document-level MRAG methods, achieving up to 27 percent relative improvement in CIDEr. By selecting fewer yet more informative fragments, our approach substantially reduces context length while improving factual accuracy and generation coherence.

cs.IR↗

arXiv2026

Factorized Latent Reasoning for LLM-based Recommendation

Large language models (LLMs) have recently been adopted for recommendation by framing user preference modeling as a language generation problem. However, existing latent reasoning approaches typically represent user intent with a single latent vector, which struggles to capture the inherently multi-faceted nature of user preferences. We propose Factorized Latent Reasoning (FLR), a novel framework for LLM-based sequential recommendation that decomposes latent reasoning into multiple disentangled preference factors. FLR introduces a lightweight multi-factor attention module that iteratively refines a latent thought representation, where each factor attends to distinct aspects of the user's interaction history. To encourage diversity and specialization, we design orthogonality, attention diversity, and sparsity regularization objectives, and dynamically aggregate factor contributions for the final prediction. We further integrate FLR with an efficient reinforcement learning strategy based on group-relative policy optimization, enabling stable alignment directly in the latent reasoning space. Experiments on multiple benchmarks show that FLR consistently outperforms strong baselines while improving robustness and interpretability.

cs.IR↗

arXiv2026

Global Context or Local Detail? Adaptive Visual Grounding for Hallucination Mitigation

Vision-Language Models (VLMs) are frequently undermined by object hallucination--generating content that contradicts visual reality--due to an over-reliance on linguistic priors. We introduce Positive-and-Negative Decoding (PND), a training-free inference framework that intervenes directly in the decoding process to enforce visual fidelity. PND is motivated by our key finding of a critical attention deficit in VLMs, where visual features are empirically under-weighted. Our framework corrects this via a dual-path contrast: The positive path amplifies salient visual evidence using multi-layer attention to encourage faithful descriptions, directly counteracting the attention deficit. Simultaneously, the negative path identifies and degrades the core object's features to create a strong counterfactual, which penalizes ungrounded, prior-dominant generation. By contrasting the model's outputs from these two perspectives at each step, PND steers generation towards text that is not just linguistically probable, but visually factual. Extensive experiments on benchmarks like POPE, MME, and CHAIR show that PND achieves state-of-the-art performance with up to 6.5% accuracy improvement, substantially reducing object hallucination while also enhancing descriptive detail--all without requiring any model retraining. The method generalizes effectively across diverse VLM architectures including LLaVA, InstructBLIP, InternVL, and Qwen-VL.

cs.CV↗

arXiv2026

V-tableR1: Process-Supervised Multimodal Table Reasoning with Critic-Guided Policy Optimization

We introduce V-tableR1, a process-supervised reinforcement learning framework that elicits rigorous, verifiable reasoning from multimodal large language models (MLLMs). Current MLLMs trained solely on final outcomes often treat visual reasoning as a black box, relying on superficial pattern matching rather than performing rigorous multi-step inference. While Reinforcement Learning with Verifiable Rewards could enforce transparent reasoning trajectories, extending it to visual domains remains severely hindered by the ambiguity of grounding abstract logic into continuous pixel space. We solve this by leveraging the deterministic grid structure of tables as an ideal visual testbed. V-tableR1 employs a specialized critic VLM to provide dense, step-level feedback on the explicit visual chain-of-thought generated by a policy VLM. To optimize this system, we propose Process-Guided Direct Alignment Policy Optimization (PGPO), a novel RL algorithm integrating process rewards, decoupled policy constraints, and length-aware dynamic sampling. Extensive evaluations demonstrate that V-tableR1 explicitly penalizes visual hallucinations and shortcut guessing. By fundamentally shifting multimodal inference from black-box pattern matching to verifiable logical derivation, V-tableR1 4B establishes state-of-the-art accuracy among open-source models on complex tabular benchmarks, outperforming models up to 18x its size and improving over its SFT baseline

cs.AI↗

arXiv2026

Breaking the Illusion: When Positive Meets Negative in Multimodal Decoding

Vision-Language Models (VLMs) are frequently undermined by object hallucination, generating content that contradicts visual reality, due to an over-reliance on linguistic priors. We introduce Positive-and-Negative Decoding (PND), a training-free inference framework that intervenes directly in the decoding process to enforce visual fidelity. PND is motivated by our finding of an attention imbalance in VLMs, where visual features are under-weighted. Our framework introduces a dual-path contrast: a positive path that amplifies visual evidence and a negative path that constructs counterfactuals to penalize prior-dominant generation. By contrasting outputs from both paths during decoding, PND steers generation toward visually grounded results. Experiments on POPE, MME, and CHAIR demonstrate state-of-the-art performance without retraining.

cs.LG↗

arXiv2026

Meituan Merchant Business Diagnosis via Policy-Guided Dual-Process User Simulation

Simulating group-level user behavior enables scalable counterfactual evaluation of merchant strategies without costly online experiments. However, building a trustworthy simulator faces two structural challenges. First, information incompleteness causes reasoning-based simulators to over-rationalize when unobserved factors such as offline context and implicit habits are missing. Second, mechanism duality requires capturing both interpretable preferences and implicit statistical regularities, which no single paradigm achieves alone. We propose Policy-Guided Hybrid Simulation (PGHS), a dual-process framework that mines transferable decision policies from behavioral trajectories and uses them as a shared alignment layer. This layer anchors an LLM-based reasoning branch that prevents over-rationalization and an ML-based fitting branch that absorbs implicit regularities. Group-level predictions from both branches are fused for complementary correction. We deploy PGHS on Meituan with 101 merchants and over 26,000 trajectories. PGHS achieves a group simulation error of 8.80%, improving over the best reasoning-based and fitting-based baselines by 45.8% and 40.9% respectively.

cs.AI↗

arXiv2026

SCOPE: Signal-Calibrated On-Policy Distillation Enhancement with Dual-Path Adaptive Weighting

On-policy reinforcement learning has become the dominant paradigm for reasoning alignment in large language models, yet its sparse, outcome-level rewards make token-level credit assignment notoriously difficult. On-Policy Distillation (OPD) alleviates this by introducing dense, token-level KL supervision from a teacher model, but typically applies this supervision uniformly across all rollouts, ignoring fundamental differences in signal quality. We propose Signal-Calibrated On-Policy Distillation Enhancement (SCOPE), a dual-path adaptive training framework that routes on-policy rollouts by correctness into two complementary supervision paths. For incorrect trajectories, SCOPE performs teacher-perplexity-weighted KL distillation to prioritize instances where the teacher demonstrates genuine corrective capability, while down-weighting unreliable guidance. For correct trajectories, it applies student-perplexity-weighted MLE to concentrate reinforcement on low-confidence samples at the capability boundary rather than over-reinforcing already mastered ones. Both paths employ a group-level normalization to adaptively calibrate weight distributions, accounting for the intrinsic difficulty variance across prompts. Extensive experiments on six reasoning benchmarks show that SCOPE achieves an average relative improvement of 11.42% in Avg@32 and 7.30% in Pass@32 over competitive baselines, demonstrating its consistent effectiveness.

cs.LG↗