跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1084新增论文
+883其中 AI 相关
+130已有论文修订
+59新增 AI 导读

arXiv 最近完成扫描:2026-11-18 04:28 · 全站最近入库:2026-11-18 19:45各信源进度 →

统计口径与数据新鲜度

统计 2026-11-17 23:42 — 2026-11-18 23:42(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-18 04:28最近发现新增版本:2026-11-18 19:45库内论文最新更新时间:2026-11-18 08:00

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

26,870 已收录论文22,374 AI 相关8,156 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 16 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-18 04:27(北京时间)
版本更新 · v3arXiv:2610.16859

ANCHOR: Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding

Wenqiang Zhu, Xinping Lei, Zhiyi Lai, Zhengshuo Gong, Tianzhuang He, Songwei Yu, Weihao Li, Jingxuan Xu 等 21 位作者

原文摘要

For repository-grounded coding agents, a functionally correct patch is not necessarily an acceptable one: developers also constrain how the task must be carried out-modification scope, workflow order, validation, tool use, and reporting. This capability, which we call engineering constraint following, remains largely invisible to current evaluations. We introduce ANCHOR (Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding), a benchmark built on naturalistic, multilingual repository tasks spanning eight language groups. ANCHOR combines (i) a hierarchical taxonomy of 10 categories and 78 leaf types distilled from real developer interactions, (ii) an automated pipeline that instantiates task-grounded atomic constraints and composes them into typed AND/THEN/PRIORITY relations, and (iii) a trajectory-grounded evaluation framework that checks both the code diff and the execution trajectory with rule-based and LLM judges. Across 19 agent configurations spanning 2 scaffolds and 10 foundation models on 358 tasks, the strongest measured system, Claude Opus 4.8 under Cline, reaches 70.1% atomic compliance and 63.7% relation compliance, while 19.0% of tasks pass every scored checklist item under the strict all-item FCCR criterion. A five-engineer audit of constraints from 50 instances reports 94% feasibility, 94% repository consistency, and 96% realism. In the primary paired set (14 configurations), the mean signed change in task completion after adding constraints is +0.7 percentage points; individual changes vary in both directions. These results distinguish functional success from constraint compliance, with failures concentrated in workflow ordering, tool-use discipline, and reporting, and substantial interactions between models and scaffolds.

cs.CL

版本更新 · v2arXiv:2610.11171

VAMR: Multi-Question Agentic Reasoning for Efficient Long-Form Video Understanding

Runquan Gui, Hanzhu Chen, Zehao Wang, Hanxin Zhu, Xin Li, Zhibo Chen

原文摘要

Long-form video understanding often involves multiple questions about different aspects of the same recording. Yet existing video agents typically process each question through an isolated tool-use trajectory. This repeatedly restarts video exploration and memory construction, missing opportunities to acquire evidence jointly and progressively build a shared understanding that supports the complete question set. We introduce \textbf{VAMR} (\textbf{V}ideo \textbf{A}gent for \textbf{M}ulti-Question \textbf{R}easoning), which coordinates all questions about a video through one shared tool-use trajectory. At each round, a persistent policy model can invoke tools for one or more unresolved questions and submit answers for questions with sufficient evidence. Question-conditioned visual perception retrieves fine-grained clues for several questions in one call, while layered multi-question memory integrates reusable context into a shared video story and preserves separate evidence for individual questions. After supervised fine-tuning initializes this interaction protocol, we propose question-horizon policy optimization (\qhpo) to optimize shared trajectories in which questions progress and finish at different rounds. Specifically, a question-level critic estimates the value of each active question, while round alignment maps each question advantage to the rounds that directly serve it before the aligned advantages are aggregated to optimize the shared actor. Across LVBench, Video-Holmes, and LongVideoBench, VAMR achieves the highest accuracy overall and the fewest reasoning rounds among iterative methods. On LVBench, it reaches 62.1\% accuracy, exceeding VideoARM by \textbf{4.3} points while reducing reasoning rounds and processed frames by \textbf{85.9\%} and \textbf{61.4\%}.

cs.CV · cs.AI

新论文 · v1arXiv:2611.16097

A Matching Lower Bound for Online Smooth Calibration

Kaixuan Ji, Quanquan Gu

原文摘要

We study smooth calibration error in sequential binary prediction, which measures how far a sequence of probabilistic forecasts is from being calibrated and, unlike the classical expected calibration error, is robust to small perturbations of the forecasts. Against adaptive adversaries, the minimax smooth calibration error is known to lie between $Ω(T^{1/3})$ and $O(\sqrt{T})$ (Qiao and Zheng, 2024), where $T$ is the number of rounds. In this paper, we close this gap by constructing a new adversary whose outcome mean moves with the forecaster's cumulative residual. This design makes compensating for past residuals costly, and forces every forecaster to incur an expected smooth calibration error of $Ω(\sqrt{T})$. This lower bound matches the known upper bound up to universal constants, and further implies $Ω(\sqrt{T})$ lower bounds for calibration distance and lower calibration distance. Overall, our results settle the minimax rates of these calibration measures in sequential prediction. The proof ideas are partially developed by GPT-6 Astra under the authors' guidance.

cs.LG · stat.ML

新论文 · v1arXiv:2611.13383

Settling the Burn-in Gap: Sharp Risk Bounds for Tail-Averaged SGD in Linear Regression

Qingyue Zhao, Xuheng Li, Quanquan Gu

原文摘要

We study the excess risk of constant-stepsize online stochastic gradient descent (SGD) with tail-averaging for high-dimensional linear regression. Building on the instance-dependent spectral framework of Zou et al. (2021), we establish matching initialization-explicit upper and lower bounds on the total excess risk of tail-averaged SGD in terms of the population covariance spectrum $\mathbf{H}$ and initialization $\mathbf{w}_0-\mathbf{w}^*$ whenever a trajectory spectral feedback condition holds. Existing analyses exhibit a fundamental gap in the effective variance: the upper bound of Zou et al. (2021) retains an undamped contribution from the initial error in the head eigenspace $\mathbf{I}_{0:k^\dagger}$, whereas their lower bound omits the variance induced by the head initialization altogether. We show that both the existing upper and lower bounds are loose: the single-hop fast-mode variance contribution from the head initialization is exponentially damped during the burn-in phase, whereas fourth-moment energy scattered from the head initialization into slow spectral modes during the early burn-in steps survives undamped into the averaging window. We prove that our spectral feedback condition is satisfied in two concrete constant-stepsize regimes: (i) for arbitrary covariance spectra $\mathbf{H}$ and any finite horizon, under an explicit horizon-dependent stepsize chosen prior to execution; and (ii) for polynomially decaying spectra $λ_i\asymp i^{-a}$ ($a>1$), under a single horizon-independent positive stepsize. Finally, we complement these positive results with explicit one- and two-dimensional separation constructions proving an exponential-versus-polynomial gap in the lower bound of Zou et al. (2021) and showing why fourth-moment assumptions alone cannot determine the asymptotic exponential decay rate without additional structural conditions.

cs.LG · math.OC · stat.ML

版本更新 · v3arXiv:2605.09214

Fast Rates for Offline Contextual Bandits with Forward-KL Regularization under Single-Policy Concentrability

Qingyue Zhao, Kaixuan Ji, Heyang Zhao, Quanquan Gu

原文摘要

\emph{Kullback-Leibler} (KL) regularization is ubiquitous in reinforcement learning algorithms in the form of \emph{reverse} or \emph{forward} KL. Recent studies have demonstrated $ε^{-1}$-type fast rates for decision making under reverse KL regularization, in contrast to the standard $ε^{-2}$-type sample complexity. However, for forward-KL-regularized objectives, existing statistical analyses are either not applicable or result in $\tilde{O}(ε^{-2})$ slow rates. We take the first step towards addressing this problem via a streamlined analysis of forward-KL-regularized offline CBs. We give the first $\tilde{O}(ε^{-1})$ upper bounds in tabular and general function approximation settings, both under notions of \emph{single-policy concentrability}. In particular, our convex-analytical pipeline unifies these settings by exploiting the pessimism principle in a novel way and completely bypasses the proof routines in previous works based on the mean value theorem, which might be of independent interest. Moreover, we provide rate-optimal lower bounds, manifesting the tightness of our upper bounds in terms of statistical rates. Our lower bounds also demonstrate that the forward-KL-regularized sample complexity recovers the unregularized slow rate in the low-regularization regime, similarly to the reverse-KL regularization.

cs.LG · cs.AI · cs.IT · math.ST · stat.ML

版本更新 · v5arXiv:2607.22877

Towards Trustworthy Physical Intelligence: From Theory to Practice Across Life Cycle

Yang Wang, Hongxuan Liu, Xinghui Xu, Arjun Menon, Xiaoran Cai, Yunyu He, Alex Tarvo, Jingzong Zhou 等 43 位作者

原文摘要

Physical intelligence refers to intelligence systems that understand, reason about and act in accordance with the physical world and its underlying laws, dynamics, and constraints. Unlike conventional AI systems, physical intelligence interacts continuously with uncertain physical environments, and its actions produce consequences that are physically irreversible. As existing trustworthy AI frameworks have been developed primarily for digital AI systems, they do not fully capture the distinctive challenges of Physical Intelligence, such as physical safety, cyber-physical security, and physical manufacturing process. To address this gap, our paper develop a foundation for understanding, operationalizing, and governing trustworthy physical intelligence. First, we characterize the fundamental capabilities of physical intelligence. Second, we investigate the role of physical laws and principles in AI and analyze the associated challenges of physical intelligence. Third, we provide a unified structure of the end-to-end physical intelligence life cycle across five core stages and introduce the Trustworthy Physical Intelligence Operationalization (T-PAIO) framework. Fourth, we formulate the Trustworthy Physical Intelligence (T-PAI) Principles as a theoretical framework that synthesizes key trustworthiness dimensions and establishes a conceptual foundation for the governance of physical intelligence systems.

cs.AI · cs.HC · cs.RO

新论文 · v1arXiv:2610.40627

ShallowDefend:利用浅层安全对齐构建多轮对话防御器

ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender

Qiyao Ma, Xiaodong Liu, Weiwei Yang, Mingqian Feng, Chengquan Guo, Jialin Song, Carolyn Buractaon, Zhe Zhao 等 12 位作者

AI 速读 基于摘要

作者提出 Assess→Decide→Respond 框架,让防御方在多轮对话的关键轮次先显式做出可解析的安全决策,再据此生成回复,以利用对齐的“浅层性”进行防御。

cs.CL

查看原文摘要

Safety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often. This brittleness has a mechanistic root: Alignment is shallow, with safety behavior concentrated in a model's first few output tokens, and existing defenses inherit the weakness, supervising the surface form of the response without ever giving the model an explicit safety decision to learn. We take the opposite view of shallowness: if a short span determines the trajectory of an entire response, that span is the highest-leverage place to intervene, and the defender can occupy it first. We propose Assess $\to$ Decide $\to$ Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it, with the commitment trained by reinforcement learning at the payload turn. Because the decision is a token span, the reward can be a rule: a self-verified reward checks the committed label against the conversation's known nature, requiring no judge and no annotation, while the same format also supports an LLM-as-judge reward over a graded decision space with safe redirection. Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness. A per-token KL analysis and interventional probes confirm why this works: training concentrates its effect on the brief deliberation that sets a response's trajectory, and the decision committed there is a causal control point. The very shallowness attackers exploit becomes the mechanism of defense.

版本更新 · v2arXiv:2510.18131

BlueCodeAgent: A Blue Teaming Agent Powered by Automated Red Teaming for CodeGen AI

Chengquan Guo, Yuzhou Nie, Chulin Xie, Zinan Lin, Wenbo Guo, Bo Li

原文摘要

Existing research on CodeGen AI security mainly focuses on red teaming, which aims to uncover vulnerabilities and risks in AI-generated code. However, progress on the blue teaming side remains limited, as effective defenses require a deep security analysis of given tasks and edge cases. To fill in this gap, we propose BlueCodeAgent, an end-to-end blue teaming agent powered by automated red teaming. Our red teaming component generates diverse risky instances, providing effective edge cases and guidance for the subsequent blue teaming process. Our blue teaming agent then conducts multi-level defense, leveraging these red teaming examples to detect previously seen and unseen risk scenarios through constitution summarization and dynamic code analysis. Our evaluation across four representative code-related tasks (bias instruction detection, malicious instruction detection, vulnerable code detection, and prompt injection detection) shows that BlueCodeAgent achieves significant gains over diverse baselines. In particular, for vulnerability detection tasks, BlueCodeAgent integrates dynamic analysis to effectively reduce false positives, a challenging problem as base models tend to be over-conservative. Overall, with GPT-4o as the base model, BlueCodeAgent achieves an average F1 score improvement of 14.7% across four tasks compared to directly prompting the model, attributed to its ability to summarize actionable constitutions and perform dynamic analysis. Our code and data are available at https://github.com/1mocat/BlueCodeAgent.

cs.SE

版本更新 · v3arXiv:2510.02609

RedCodeAgent: Automatic Red-teaming Agent against Diverse Code Agents

Chengquan Guo, Chulin Xie, Yu Yang, Zhaorun Chen, Zinan Lin, Xander Davies, Yarin Gal, Dawn Song 等 9 位作者

原文摘要

Code agents have gained widespread adoption due to their strong code generation capabilities and integration with code interpreters, enabling dynamic execution, debugging, and interactive programming capabilities. While these advancements have streamlined complex workflows, they have also introduced critical safety and security risks. Current static safety benchmarks and red-teaming tools are inadequate for identifying emerging real-world risky scenarios, as they fail to cover certain boundary conditions, such as the combined effects of different jailbreak tools. In this work, we propose RedCodeAgent, the first automated red-teaming agent designed to systematically uncover vulnerabilities in diverse code agents. With an adaptive memory module, RedCodeAgent can leverage existing jailbreak knowledge, dynamically select the most effective red-teaming tools and tool combinations in a tailored toolbox for a given input query, thus identifying vulnerabilities that might otherwise be overlooked. For reliable evaluation, we develop simulated sandbox environments to additionally evaluate the execution results of code agents, mitigating potential biases of LLM-based judges that only rely on static code. Through extensive evaluations across multiple state-of-the-art code agents, diverse risky scenarios, and various programming languages, RedCodeAgent consistently outperforms existing red-teaming methods, achieving higher attack success rates and lower rejection rates with high efficiency. We further validate RedCodeAgent on real-world code assistants, e.g., Cursor and Codeium, exposing previously unidentified security risks. By automating and optimizing red-teaming processes, RedCodeAgent enables scalable, adaptive, and effective safety assessments of code agents. Our code and data are available at https://github.com/1mocat/RedCodeAgent.

cs.SE

新论文 · v1arXiv:2610.38867

Where You Inject Matters: The Trust Region Effect in LLM Agents

Chengquan Guo, Xiaodong Liu, Weiwei Yang, Mingqian Feng, Jialin Song, Qiyao Ma, Carolyn Buractaon, Christopher White 等 10 位作者

原文摘要

LLM agents are increasingly adopted in long-horizon tasks and interact with diverse external environments. Despite their growing capabilities, these agents remain vulnerable to real-world attacks such as indirect prompt injections, which manipulate agents through malicious instructions embedded in external environments. While existing risk assessment work has primarily focused on the injection content, the effect of where an injection appears has not been systematically characterized. Studying injection placement requires agents to operate over longer-horizon tasks and interact with diverse environmental surfaces. Thus, we build three agentic testbeds spanning tool use, coding, and daily workflow domains with diverse injection surfaces. Holding the injection content fixed, we find that agent susceptibility varies across injection placements. We refer to this phenomenon as the trust region effect: agents implicitly assign different levels of trust to different regions of their environments. For example, an agent powered by GPT-5.4 follows an unsafe injection instruction placed in a repository documentation file (i.e., a README or TODO file) 2.6 times more often on average than the same instruction placed in a code or data file. Building upon this observation, we develop learning-based attackers that generate injection content and select better injection placements. Our approach demonstrates that both where an injection is placed and what content it contains can affect the strength of indirect prompt injection attacks. To the best of our knowledge, this work provides the first systematic characterization for the effectiveness of different injection placements, spanning both coarse-grained environmental surfaces and fine-grained positions. The uneven trust that agents implicitly place across their environments provides important insights for future attack and defense design.

cs.AI · cs.CR · cs.LG

版本更新 · v2arXiv:2510.03199

Best-of-Majority: Minimax-Optimal Strategy for Pass@$k$ Inference Scaling

Qiwei Di, Kaixuan Ji, Xuheng Li, Heyang Zhao, Quanquan Gu

原文摘要

LLM inference often generates a batch of candidates for a prompt and selects one via strategies like majority voting or Best-of- N (BoN). For difficult tasks, this single-shot selection often underperforms. Consequently, evaluations commonly report Pass@$k$: the agent may submit up to $k$ responses, and only the best of them is used when computing regret. Motivated by this, we study inference scaling in the more general Pass@$k$ inference setting, and prove that neither majority voting nor BoN exhibits the desirable scaling with $k$ and the sampling budget $N$. Combining the advantages of majority voting and BoN, we propose a new inference strategy called Best-of-Majority (BoM), with a pivotal step that restricts the candidates to the responses with high frequency in the $N$ samples before selecting the top-$k$ rewards. We prove that when the sampling budget is $N=\tildeΩ(C^*)$, the regret of BoM is $O(ε_{\mathrm{opt}}+\sqrt{ε_{\mathrm{RM}}^2C^*/k})$, where $C^*$ is the coverage coefficient, $ε_{\mathrm{RM}}$ is the estimation error of the reward model, and $ε_{\mathrm{opt}}$ is the estimation error of reward at the optimal response. We further establish a matching lower bound, certifying that our algorithm is minimax optimal. Beyond optimality, BoM has a key advantage: unlike majority voting and BoN, its performance does not degrade when increasing $N$. Experimental results of inference on math problems show BoM outperforming both majority voting and BoN.

cs.LG · stat.ML

版本更新 · v2arXiv:2603.02429

Dimension-Independent Convergence of Underdamped Langevin Monte Carlo in KL Divergence

Shiyuan Zhang, Qiwei Di, Xuheng Li, Quanquan Gu

原文摘要

Underdamped Langevin dynamics (ULD) is a widely-used sampler for Gibbs distributions $π\propto e^{-V}$, and is often empirically effective in high dimensions. However, existing non-asymptotic convergence guarantees for discretized ULD typically scale polynomially with the ambient dimension $d$, leading to vacuous bounds when $d$ is large. The main known dimension-free result concerns the randomized midpoint discretization in Wasserstein-2 distance (Liu et al.,2023), while dimension-independent guarantees for ULD discretizations in KL divergence have remained open. We close this gap by proving the first dimension-free KL divergence bounds for discretized ULD. Our analysis refines the KL local error framework (Altschuler et al., 2025) to a dimension-free setting and yields bounds that depend on $\mathrm{tr}(\mathbf{H})$, where $\mathbf{H}$ upper bounds the Hessian of $V$, rather than on $d$. As a consequence, we obtain improved iteration complexity for underdamped Langevin Monte Carlo relative to overdamped Langevin methods in regimes where $\mathrm{tr}(\mathbf{H})\ll d$.

cs.LG · math.OC · stat.ML

新论文 · v1arXiv:2610.38243

Nearly Minimax Optimal Regret for Adversarial Linear Bandits with Strong Corruptions

Qiwei Di, Quanquan Gu

原文摘要

We study adversarial linear bandits with strong corruption, in which an adaptive adversary chooses the reward vectors and a strong adversary corrupts each observed reward after seeing the learner's action, subject to a known total corruption budget $C$. Strong corruption breaks the importance-weighted reward estimates used by adversarial bandit algorithms: by corrupting only the rounds in which a rarely sampled action is drawn, the adversary amplifies each unit of corruption by the inverse of its sampling probability. The best known bound for this problem, $\tilde O(\sqrt{d^3T} + dC)$ by Liu et al. (2024), is a factor $\sqrt{d}$ above the corruption-free rate in its leading term, and whether $\tilde O(d\sqrt{T} + dC)$ regret is achievable was left open. We propose CRW-FTRL, which runs follow-the-regularized-leader with a radius-weighted entropy regularizer and estimates rewards with a normalized kernel estimator whose sensitivity to corruption is governed by per-action radii instead of inverse sampling probabilities. The radii are nondecreasing and are chosen jointly with the sampling distribution through a monotone log-determinant barrier problem, so that a negative linear term in the regularizer cancels both the corruption bias and the estimation error of every comparator. For $K$ actions in $\mathbb{R}^d$, CRW-FTRL achieves $\tilde O(\sqrt{dT\log K} + dC)$ regret with high probability against an adaptive adversary, which matches the best known corruption-free rate and the $Ω(dC)$ lower bound up to logarithmic factors. Applied to a discretization of a general compact action set, it yields $\tilde O(d\sqrt{T} + dC)$ regret, which settles the statistical rate of corrupted adversarial linear bandits under strong corruption.

cs.LG · stat.ML

版本更新 · v2arXiv:2610.35743

SEMA-Auto: Sighted and Efficient Multi-Turn Attack with Autonomy

Mingqian Feng, Xiaodong Liu, Weiwei Yang, Chenliang Xu, Jialin Song, Qiyao Ma, Chengquan Guo, Carolyn Buractaon 等 11 位作者

原文摘要

Existing jailbreak attackers fall into two classes: static planners fix their single- or multi-turn adversarial plans in advance, remaining simple but losing sight of the victim's replies, wasting turns and limiting transferability; interactive generators read each victim reply to generate the next-turn attack but require in-loop evaluators for success signals or hints, incurring substantial costs and inflated performance. Bridging this gap, we propose SEMA-Auto, a sighted and efficient multi-turn attacker with autonomy, interacting with the victim model and terminating the session by itself upon success. To stabilize its behaviors and boost its capabilities, SEMA-Auto is trained in two stages. Synthetic bootstrap tuning installs the notion of attack targets and conforming decisions by fine-tuning the base model with bootstrapped trajectories. Decision-supervised reinforcement learning then optimizes two objectives jointly yet separately, reinforcing the attack with higher jailbreak reward while supervising the decision at each turn against an absolute target. Our method achieves state-of-the-art (SOTA) attack success rates (ASR) across multiple datasets, victims, and judges, outperforming all static and interactive baselines, within almost the fewest turns, demonstrating both effectiveness and efficiency. For example, SEMA-Auto performs a $76.0\%$ ASR@1 within an average of $3.0$ turns against DeepSeek-V4-Flash on AdvBench, while the best static baseline is $54.8\%$ with fixed $7$ turns and the best interactive baseline is $43.8\%$ in $3.7$ turns. Against safety-hardened frontier models, GPT-oss-120B, GPT-5.4, Claude-Haiku-4.5, and GPT-6-Astra, our counter-trained attacker achieves an average $59.6\%$ ASR@1 on HarmBench, $13.8\%$ over prior SOTA. Our code is available at: https://github.com/microsoft/SEMA

cs.CL

新论文 · v1arXiv:2610.36849

When All Rollouts Fail: Policy-Compatible Joint Decoding for Hard-Problem Reasoning

Zehao Wang, Hanzhu Chen, Runquan Gui, Zhe Wang, Lin Yang

原文摘要

Reinforcement Learning with Verifiable Rewards (RLVR) derives its learning signal from binary rewards via on-policy exploration and becomes a powerful paradigm for improving the reasoning of large language models. Yet it faces a learning cliff on problems beyond model's current ability: when every rollout fails, the samples reveal nothing about a correct solution, and group-relative estimators (e.g., GRPO, RLOO) collapse the update to zero advantage. Expert trajectories offer a remedy, but they are incompatible with the current policy and often lie in the student's low-probability regions, so the surviving updates demand large parameter shifts and decay into surface-level imitation rather than transferable reasoning. We argue that a useful learning signal on hard problems must be both verifiably correct and sufficiently compatible with the current policy. We propose Policy-Compatible Reinforcement Learning (PICO), which performs distribution-constrained joint decoding: expert data guides decoding at the token level while preserving the student's relative token probabilities, and a dynamic top-k mechanism confines the sampling distribution under a finite divergence budget, yielding trajectories that are verified and policy-compatible by construction. These trajectories are optimized jointly with on-policy rollouts under a mixed-behavior objective that mitigates their distribution mismatch via importance sampling. Experiments across reasoning diverse datasets and agentic benchmarks demonstrate that PICO consistently outperforms standard RLVR and prior off-policy baselines.

cs.LG · cs.AI

版本更新 · v2arXiv:2605.02141

On the Optimal Sample Complexity of Offline Multi-Armed Bandits with KL Regularization

Kaixuan Ji, Qiwei Di, Heyang Zhao, Qingyue Zhao, Quanquan Gu

原文摘要

Kullback-Leibler (KL) regularization is widely used in offline decision-making and offers several benefits, motivating recent work on the sample complexity of offline learning with respect to KL-regularized performance metrics. Nevertheless, the exact sample complexity of KL-regularized offline learning remains largely from fully characterized. In this paper, we study this question in the setting of multi-armed bandits (MABs). We provide a sharp analysis of KL-PCB (Zhao et al., 2026), showing that it achieves a sample complexity of $\tilde{O}(ηSAC^{π^*}/ε)$ under large regularization $η= \tilde{O}(ε^{-1})$, and a sample complexity of $\tildeΩ(SAC^{π^*}/ε^2)$ under small regularization $η= \tildeΩ(ε^{-1})$, where $η$ is the regularization parameter, $S$ is the number of contexts, $A$ is the number of arms, $C^{π^*}$ policy coverage coefficient at the optimal policy $π^*$, $ε$ is the desired sub-optimality, and $\tilde{O}$ and $\tildeΩ$ hide all poly-logarithmic factors. We further provide a pair of sharper sample complexity lower bounds, which matches the upper bounds over the entire range of regularization strengths. Overall, our results provide a nearly complete characterization of offline multi-armed bandits with KL regularization.

cs.LG · cs.AI · math.ST · stat.ML