paperlist.ai

AI 顶会论文

AI 顶会论文库

检索 30+ 个顶会的论文，包括 NeurIPS、ICLR、ICML、CVPR、ACL、AAAI 等。输入关键词筛选，打开任意一篇查看摘要、作者和链接。

arXiv

Contemporary AI lacks the imagination to diverge or negate in science

Honglin Bao, Siyang Wu, Xiao Liu, Sida Li, Shiyun Cao, James A. Evans

Bold claims that AI will accelerate scientific discovery have raced ahead of evidence from working scientists, yet large-scale, scientist-in-the-loop evidence is scarce. Here we mount the largest evaluation to date, inviting authors of 121,640 recent preprints in biology, medicine, chemistry, and social science to judge large language model (LLM)-generated ideas derived from their own papers. 6,749 representative scientists returned 25,139 rating sets on novelty, feasibility, probability of being true, and favorability of adoption. Three patterns emerge. First, non-reasoning LLMs collapse into a narrow "hivemind" of similar ideas while reasoning models explore a wider hypothesis space, but no model spontaneously proposes null hypotheses, a move humans make more freely. Second, scientists reward ideas resembling their own and prize probability over novelty, though social scientists tolerate risk more than life scientists; senior social scientists are the harshest critics, and their skepticism is earned, as LLMs falter most in pluralistic fields demanding context-aware interpretation and evolving theories. Third, automated evaluators, including LLM-as-a-judge and state-of-the-art (SOTA) models, agree weakly with expert judgment. Retrieval augmentation and scientist persona prompting yield marginal gains. A Qwen3-14B reward model we post-trained on human ratings captures nuances of taste, beats SOTA models by up to 27%, and closes the gap to the consistency of human peer reviewers. An analysis of 39 million papers from 2010 to 2025 links survey findings to macro-level patterns: following ChatGPT's release, null claims are sharply suppressed and ideas contract. Agent-based simulations further suggest that saturated fields should especially prize human uniqueness. For all the hype, today's AI for science remains a collaborator whose imagination and judgment benefit from human grounding.

arXiv

dots.tts Technical Report

Shi Lian, Changtao Li, Bohan Li, Hankun Wang, Da Zheng, Junfeng Tian, Yufeng Ma, colin zhang, Kai Yu

We present dots$.$tts, a 2B-parameter continuous autoregressive text-to-speech (TTS) foundation model that models speech in a continuous latent space. Compared with existing continuous autoregressive models, our key innovations are threefold. First, we train an AudioVAE with multiple objectives to build a semantically structured and prediction-friendly continuous speech space. Second, we use full-history conditioning in the flow-matching head to preserve long-range consistency and reduce drift during generation. Third, we apply reward-free self-corrective post-training to the flow-matching head to further improve robustness and acoustic quality. After being trained on a large-scale multilingual corpus, dots$.$tts achieves the best average performance on Seed-TTS-Eval, with WERs of 0.94%/1.30%/6.60% and SIM scores of 81.0/77.1/79.5 on the zh/en/zh-hard test sets, respectively. Across other benchmarks, dots$.$tts also consistently demonstrates open-source state-of-the-art performance, exhibiting strong generation stability, voice cloning ability, and emotional expressiveness. For efficient inference, we further apply CFG-aware MeanFlow distillation, enabling low-latency speech generation with first-packet latencies of 85/54 ms in output streaming and dual-streaming modes, respectively. To facilitate reproducible research and practical deployment, we release the training and inference code, together with the pretrained, post-trained, and MeanFlow-distilled checkpoints, under the Apache 2.0 license.

arXiv

Understanding and Mitigating Premature Confidence for Better LLM Reasoning

Jingchu Gai, Guanning Zeng, Christina Baek, Chen Wu, J Zico Kolter, Andrej Risteski, Aditi Raghunathan

Long chains of thought (CoT) from current language models frequently contain logical gaps and unjustified leaps, limiting the gains from additional test-time compute. Improving reasoning quality directly would require process reward models, but the step-level annotations needed to train them are expensive and scarce. We find such a signal in how the model's confidence evolves during reasoning: premature confidence, the tendency to commit to an answer early and use the remaining tokens to rationalize it, strongly predicts flawed reasoning across tasks and model scales. We exploit this in progressive confidence shaping, a reinforcement learning objective that trains models to update their confidence as they reason rather than commit early -- rewarding gradual confidence growth and penalizing early commitment, with no external labels or reward models. The method improves accuracy and reasoning quality from 1.5B to 8B parameters across arithmetic (Countdown), math (DAPO, AIME), and science (ScienceQA): on Countdown, accuracy improves 3.2x (+42.0pp) and flawed reasoning drops 48pp; on AIME, Pass@64 improves 6.6pp. Consistent with this mechanism, the method also improves faithfulness: on a safety benchmark, our models more transparently surface misleading content in their reasoning traces rather than concealing it. Controlled experiments reveal that the problem and its remedy scale together: premature confidence grows with model size and task difficulty, and so do the gains from addressing it.

arXiv

AgentPSO: Evolving Agent Reasoning Skill via Multi-agent Particle Swarm Optimization

Hyunmin Hwang, Jaemin Kim, Choonghan Kim, Hangeol Chang, Jong Chul Ye

Multi-agent reasoning has shown promise for improving the problem-solving ability of large language models by allowing multiple agents to explore diverse reasoning paths. However, most existing multi-agent methods rely on inference-time debate or aggregation, which can be vulnerable to incorrect peer influence and biased consensus. Moreover, the agents' reasoning procedures remain static, as they do not accumulate reusable behaviors across tasks. In this paper, we introduce \textbf{AgentPSO}, a particle-swarm-inspired framework for evolving multi-agent reasoning skills. AgentPSO treats each agent as a particle-like reasoner whose state is a natural-language skill and whose velocity is a semantic update direction, iteratively guiding agents toward higher-performing skill configurations. Across training iterations, each agent updates its skill by combining its previous velocity, personal-best skill, global-best skill, and a self-reflective direction derived from peer reasoning trajectories. This enables agents to learn reusable reasoning behaviors by drawing on their own experience and on the strongest skills found by the population, without updating the parameters of the backbone language model. Experiments on mathematical and general reasoning benchmarks show that AgentPSO improves over static single-agent skills and test-time-only multi-agent reasoning baselines. The evolved skills further transfer across benchmarks and to another backbone model, suggesting that AgentPSO captures reusable reasoning procedures rather than merely optimizing benchmark-specific prompts. Code is publicly available at https://github.com/HYUNMIN-HWANG/AgentPSO/.

arXiv

SALLIE: Generation-Free Hidden-State Detection of Jailbreaks and Prompt Injections Across Text and Vision

Guy Azov, Ofer Rivlin, Guy Shtar

Large Language Models (LLMs) and Vision-Language Models (VLMs) are vulnerable to jailbreaks and prompt injections delivered through text or images. Existing defenses often narrow threat coverage or add inference cost through input transformations, repeated model queries, generated responses, or threat-specific detectors. We present SALLIE (Single-Pass Activation Lookup for Layerwise Input Evaluation), a white-box, generation-free detector with a common architecture for textual and visual jailbreaks and prompt injections. Using one backbone forward pass, SALLIE extracts last-token residual-stream representations, applies layer-wise k-nearest-neighbor probes, and averages their scores across a contiguous layer range, without modifying the backbone or generating a response. To distinguish reusable detector configurations from deployment-specific tuning, we compare shared, threshold-only, and fully calibrated regimes across backbones and modalities. We evaluate SALLIE on compact open-weight VLMs (Gemma-3-4b-it, Phi-3.5-vision-instruct, SmolVLM2-2.2B-Instruct) using heterogeneous held-out test data spanning both threat families and modalities. Fully calibrated SALLIE-Phi achieves an aggregate F1 of 0.87 and higher aggregate balanced accuracy than RCS-KCD on all three backbones under a matched-calibration protocol. On Phi-3.5 visual inputs, threshold-only calibration yields an F1 of 0.99 with zero observed false positives, versus 0.54 for zero-shot Gemini-2.5-Flash-Lite. On Gemma-4-E2B-it, a backbone held out from representation-hyperparameter selection, the same procedure raises F1 from 0.62 to 0.79 on text and 0.68 to 0.98 on visual inputs. Representation hyperparameters can thus transfer across some backbones, while operating points remain backbone-, modality-, and distribution-dependent. We further report FPR/FNR tradeoffs, latency, and representation ablations.

arXiv

Goedel-Code-Prover: Hierarchical Proof Search for Open State-of-the-Art Code Verification

Zenan Li, Ziran Yang, Deyuan He, Haoyu Zhao, Andrew Zhao, Shange Tang, Kaiyu Yang, Aarti Gupta, Zhendong Su, Chi Jin

Large language models (LLMs) can generate plausible code but offer limited guarantees of correctness. Formally verifying that implementations satisfy specifications requires constructing machine-checkable proofs, a task that remains beyond current automation. We propose a hierarchical proof search framework for automated code verification in Lean~4 that decomposes complex verification goals into structurally simpler subgoals before attempting tactic-level proving. Central to our approach is a principled decomposition score that combines constructive justification with structural effectiveness. The same score serves as both the training reward and the inference-time ranking criterion, aligning optimization and deployment. We train Goedel Code Prover, a single unified policy for both decomposition and completion, through supervised initialization followed by hybrid reinforcement learning, where a continuous decomposition reward supports planning exploration while supervised replay stabilizes proof generation. On three Lean-based code verification benchmarks comprising 427 tasks, our 8B-parameter model achieves a 62.0% prove success rate, a 2.6 improvement over the strongest baseline under the reported inference settings. We further observe consistent inference-time scaling: success rates improve monotonically with search iterations and sampling budget, while whole-proof baselines plateau within the evaluated budgets.

arXiv

AutoRefine: Compiling Trajectories into Validated Typed Agent Artifacts

Libin Qiu, Zhirong Gao, Junfu Chen, Yuhang Ye, Liangyu Li, Weizhi Huang, Xiaobo Xue, Wenkai Qiu, Shuo Tang

Large language model agents repeatedly encounter related tasks, yet systems that learn from trajectories commit every lesson to one predefined artifact form. A local constraint, a reusable procedure, and a delegated objective require different amounts of runtime ownership, so one form either under-specifies the correction or wraps it in execution machinery it does not need. We present AutoRefine, which treats trajectory learning as typed artifact compilation. It contrasts failed against successful trajectories to derive a type-neutral, evidence-linked intervention specification, then compiles that specification into the first Rule, Skill, or bounded Subagent that closes it under a runtime-relative ownership order: the selected schema must own every specified observation, state variable, dependent decision, and completion condition. Validation is stated in the same terms: a type-specific contract gate tests whether the generated object realizes its declared boundary, and a replay gate admits it only when it improves the correction cases linked to its source failures without regression on preservation cases. With GPT-5.6-terra as the shared backbone, AutoRefine records the highest success on ALFWorld, ScienceWorld, TravelPlanner, and SpreadsheetBench, and ties the best result on SkillCraft; on TravelPlanner it reaches 80.56% success against 50.0% for the strongest baseline. Removing boundary closure or replay validation costs 15.00 and 16.11 percentage points, the two largest losses among our construction and admission policies. In a longitudinal TravelPlanner stream, the repository holds 89--91% held-out success after 60 learning tasks with no net loss on previously solved tasks, and frozen repositories improve all 25 evaluated source--target pairs, more within a domain (14.20 points on average) than across domains (6.99).

arXiv

The Theory of Strategic Evolution: Games with Endogenous Players and the Seven Laws of Strategic Replicators

Kevin Vallier

Von Neumann founded both game theory and the theory of self-reproducing automata, but the two programs never merged. Rational players do not control their replication, and replicators do not choose strategically. Contemporary AI systems expose this gap: they optimize objectives, yet the population of AI systems is not fixed but expands and contracts based on performance. When capital can spawn capital, we need a theory that captures both rationality and replication. The Theory of Strategic Evolution analyzes strategic replicators: entities that optimize under resource constraints and spawn copies of themselves. The framework is organized around Seven Laws: 1. Strategic Selection: Mean fitness serves as a Lyapunov function; dominated types are eliminated. 2. ESDI Characterization: Equilibria exist, are generically finite, and satisfy Nash-KKT-LP equivalence. 3. H-$γ$ Stability: Multi-level systems are stable iff the spectral radius $ρ(Γ) < 1$. 4. G$\infty$ Closure: A unique maximal class of safe modifications exists and is closed under composition. 5. Constitutional Duality: Shadow prices implement any frontier allocation; welfare theorems hold. 6. Alignment Impossibility: Full reachability destroys Lyapunov structure; alignment requires bounded modification. 7. Hopf Transition: At critical coupling, systems undergo supercritical bifurcation producing limit cycles. Applications range from AI deployment dynamics to institutional design. The framework shows why "personality engineering" fails under selection pressure and identifies constitutional constraints necessary for stable alignment. The Lean 4 formalization verifies every result with zero custom axioms. The framework generates empirical predictions that remain untested. The paper and its repository are therefore a prototype kernel: an instrument for studying strategic replicators, not a set of results about the world.

arXiv

Multi-Modal Scene Graph with Kolmogorov-Arnold Experts for Audio-Visual Question Answering

Zijian Fu, Changsheng Lv, Xianlin Zhang, Mengshi Qi, Huadong Ma

In this paper, we propose a novel Multi-Modal Scene Graph with Kolmogorov-Arnold Expert Network for Audio-Visual Question Answering (SHRIKE). The task aims to mimic human reasoning by extracting and fusing information from audio-visual scenes, with the main challenge being the identification of question-relevant cues from complex audio-visual content. Existing methods fail to capture the structural information within videos and suffer from insufficient fine-grained modeling of multi-modal features. To address these issues, we are the first to introduce a new multi-modal scene graph that explicitly models objects and their relationships as a visually grounded, structured representation of the audio-visual scene, yielding 461,292 relation triplets over 9,288 musical performance videos. Furthermore, we design a Kolmogorov-Arnold Network (KAN)-based Mixture of Experts (MoE) to enhance the expressive power of the temporal integration stage. This enables more fine-grained modeling of cross-modal interactions within the question-aware fused audio-visual representation, capturing richer and more nuanced patterns and improving temporal reasoning performance. We evaluate the model on the established MUSIC-AVQA and MUSIC-AVQA v2 benchmarks, where it achieves state-of-the-art performance, reaching 78.14% average accuracy on MUSIC-AVQA, surpassing the previous best method QA-TIGER and ranking first under all four configurations of MUSIC-AVQA v2.0. Code and model checkpoints will be publicly released at https://github.com/feel12348/SHRIKE.

arXiv

When Grammar Guides the Attack: Uncovering Control-Plane Vulnerabilities in LLMs with Structured Output

Shuoming Zhang, Jiacheng Zhao, Hanyuan Dong, Ruiyuan Xu, zhicheng li, Yangyu Zhang, Shuaijiang Li, YUAN WEN, Chunwei Xia, Zheng Wang, Xiaobing Feng, Huimin Cui

Content Warning: This paper may contain unsafe or harmful content generated by LLMs that may be offensive to readers. Large Language Models (LLMs) increasingly serve as tooling platforms through structured output APIs, but the grammar-guided decoding that powers this feature opens a critical control-plane attack surface orthogonal to traditional data-plane vulnerabilities. We introduce Constrained Decoding Attack (CDA), a new jailbreak class that targets the LLM control plane. CDA is best characterized as a control-to-semantic pipeline: (1) schema-enforced logit masking injects a malicious prefix into the generation trajectory, and (2) the model itself completes the harmful intent. Unlike data-plane jailbreaks that rely on bypassing alignment with visible inputs, CDA acts on the decoding process itself, so internal safety alignment alone cannot stop it. We instantiate CDA with EnumAttack, which hides malicious content in enum fields, and the more evasive DictAttack, which decouples the payload across a benign prompt and a dictionary-based grammar. Across 13 proprietary/open-weight models and five standard benchmarks, DictAttack achieves 94.3--99.5% Attack Success Rate (ASR) on flagship models including gpt-5, gemini-2.5-pro, deepseek-r1, and gpt-oss-120b. While basic grammar auditing mitigates EnumAttack, DictAttack still sustains 75.8% ASR against SOTA jailbreak guardrails, exposing a "semantic gap" that demands cross-plane defenses bridging the data and control planes. Project page and code are available at https://ict-cda.github.io/.

arXiv

Establishing Boundary KKT Convergence of Mirror Descent through Reparameterization

Kuangyu Ding, Kim-Chuan Toh

Sequence convergence to a boundary Karush--Kuhn--Tucker (KKT) point has long remained unclear for nonconvex mirror descent with Legendre kernels. The difficulty arises from the blow-up of the gradient of the Legendre kernel at the boundary. Recent work~\cite{dingtoh2026nonkkt} shows that mirror descent can accumulate at non-KKT boundary points despite decreasing objective values, precluding a convergence guarantee to KKT points in general. Despite this negative result, mirror descent remains effective in many real applications. Motivated by this contrast, we address the boundary difficulty directly and establish KKT convergence of mirror descent for a broad class of structured nonconvex problems. We analyze mirror descent in reparameterized variables, where the Hessian metric is flattened and remains nondegenerate as the boundary is approached. Under extension and definability conditions jointly coupling the objective, the Legendre kernel, and the feasible region, the reparameterized sequence has finite length and converges, thereby recovering convergence to a KKT point of the original sequence. Our general framework applies to some concrete instances: Shannon entropy, Fermi--Dirac entropy, and power kernels on polyhedron.

arXiv

bioMoR: Biology-Guided Mixture-of-Recursions for Effective Genomic Learning

Koushik Howlader, Tirtho Roy, Md Tauhidul Islam, Wei Le

Transformer models for high-dimensional omics analysis process thousands of genes or pathways, although only a subset requires deep computation. Mixture-of-Recursions (MoR) improves efficiency through adaptive token-choice or expert-choice routing. We propose bioMoR, which, to the best of our knowledge, is the first framework to apply MoR to gene-level and pathway-level learning. Our contributions include identifying three locations for integrating structured biological knowledge within an MoR backbone: graph-based information sharing refines token embeddings, a structural bias guides self-attention toward biologically related tokens, and a graph-aware router uses neighborhood information to determine each token's recursion depth. These techniques are centered on our insight that additional knowledge of token interaction can effectively help models construct embeddings and select which tokens should be learned more deeply. Across eight benchmarks spanning diverse omics data types and evaluated under a unified five-fold cross-validation protocol, bioMoR improves average macro-F1 by 8.2 percentage points and balanced accuracy by 7.1 percentage points over the strongest biology-agnostic MoR baseline while using 75 percent fewer parameters and up to 58 percent fewer FLOPs than a non-recursive Transformer. The selected marker genes or pathways provide biological interpretability, while their token-specific recursion depths reveal how computation is allocated.

arXiv

End-to-End Fairness Optimization with Fair Decision-Focused Learning

Yu Wang, Violet Xinying Chen

Many real-world systems rely on predictive models to inform decisions, and fairness concerns arise in both the prediction and decision stages. We introduce end-to-end fairness optimization (E2EFO) as a unifying framework that integrates fairness across the prediction-to-decision pipeline. We focus on resource allocation with group-based fairness: the prediction task estimates allocation impacts while limiting accuracy disparity across groups, and the decision task distributes those impacts equitably by optimizing a group-based alpha-fairness measure. Within this framework, we propose fair decision-focused learning (FDFL), a training paradigm that jointly accounts for prediction accuracy, prediction fairness, and decision regret -- the loss in decision fairness due to imperfect predictions. FDFL trains the predictor by gradient descent, combining the objective gradients through multi-task learning techniques. The core computational challenge is the decision Jacobian with respect to the predictor parameters: we derive exact closed-form formulas for a tractable class of fair allocation and apply a differentiable optimization layer in the general case. We further establish a finite-sample generalization bound for the scalarized FDFL objective. Numerical experiments on a healthcare-based single resource allocation and a synthetic multiple resource allocation illustrate the value of jointly accounting for prediction fairness and decision fairness in prediction-informed decision-making.

arXiv

FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications

Krish Agarwal, Zhuoming Chen, Yanyuan Qin, Zhenyu Gu, Atri Rudra, Beidi Chen

Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and intra-model parallelism. Existing serving systems and auto-parallelism compilers commit to limited transformations and fixed workload assumptions, so achieving high performance on a new application requires hand-crafting an efficient implementation. We present FlashRT, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput. Using a new chain-of-program paradigm, FlashRT directs a generic coding agent through a multi-pass transformation process where an agent transforms the reference into an intermediate representation (IR) to capture data dependencies and persistent-state scopes, validates this IR via a sequential interpreter, and performs static analyses to identify candidate transformations. Then, the agent iteratively implements, verifies, and benchmarks each candidate under a measurement-gated optimization loop to produce effective deployments that span different hardware budgets. Across various applications, including video world models and multimodal LLMs, FlashRT converts reference implementations into highly efficient deployments, delivering up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs. On AMD MI355X GPUs, FlashRT matches the peak latency reduction while increasing peak throughput improvement to 3.6x, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization. In fact, for Qwen3-Omni text-to-audio inference, FlashRT reduces response latency by 65% compared to the expert vLLM-Omni implementation on AMD MI355X.

arXiv

An Auto-Scaling Approach for Serverless Environments Based on a Multi-Expert Consensus Mechanism

Mobina Kashaniyan, Mehrdad Ashtiani, Amirhossein Ghassemi

Serverless computing provides automatic resource management and pay-per-use execution, but effective autoscaling remains challenging because of dynamic workloads, cold-start latency, and dependencies among functions. We present a dependency-aware autoscaling framework that integrates graph-based bottleneck identification, short-term workload forecasting, multi-model consensus, and cost-aware scaling control. Serverless applications are represented as directed dependency graphs, and structurally important functions are identified using weighted degree centrality. Resource demand is predicted using lightweight MLP, LSTM, and CNN models. Their outputs are combined through a performance-weighted probabilistic ensemble inspired by Bayesian model averaging. The controller further incorporates cold-start awareness and cost comparison to select among scale-up, scale-down, and hold actions. Experiments using real workload traces show that supervised forecasting substantially outperforms unsupervised clustering for autoscaling decision generation. The proposed ensemble achieves 99.88 percent prediction accuracy and reduces prediction error compared with representative hybrid forecasting methods. Evaluations across multiple cloud pricing models also demonstrate consistent infrastructure cost reductions while maintaining performance targets. The results show that combining dependency analysis, multi-expert forecasting, and cost-aware control provides a robust and practical solution for serverless autoscaling.

arXiv

Concept-Guided Spatial Regularization for World Models in Atari Pong

Yukuan Lu, zaishuo xia, Weyl Lu, Yubei Chen

World models are usually evaluated as components of model-based reinforcement learning (MBRL) systems, leaving their standalone reliability understudied. We reproduce five visual world-model agents in Atari Pong -- DreamerV3, DIAMOND, TWISTER, Simulus, and STORM -- and match their reported agent performance. We then freeze the learned world models and evaluate them in two ways. In a closed-loop rollout diagnostic, a policy trained separately from the corresponding MBRL agent interacts with each frozen model, and we inspect the generated visual trajectories for visual and dynamical errors. Across all five models, these rollouts contain clear failures, including ball disappearance, incorrect motion, and invalid ball-paddle interactions. Beyond visual trajectories, we further evaluate the frozen models with pixel-space zero-shot MBRL, a challenging setting in which a new policy is trained entirely inside each frozen world model and then evaluated in the real environment. Across all five models, these policies substantially underperform those produced by the corresponding original MBRL pipelines. For DreamerV3, mean return drops from $-5.5$ to $-20.9$, near the minimum of $-21$. We hypothesize that insufficient modeling of task-critical concepts, such as the ball in Pong, contributes to these failures and propose Concept-Guided Spatial Regularization (CGSReg), an auxiliary reconstruction loss on segmented concept regions. CGSReg improves both closed-loop rollouts and pixel-space zero-shot MBRL in DreamerV3, DIAMOND, and TWISTER, and improves zero-shot MBRL in Simulus; STORM shows no clear improvement.

arXiv

Proxy OPD: On-Policy Distillation with Transferable Relative Proxy Update

Daocheng Fu, Rong Wu, Yu Yang, Jianbiao Mei, Licheng Wen, Pinlong Cai, Xuemeng Yang, Yong Liu, Botian Shi, Yu Qiao

Post-training for large language models typically couples policy exploration with model optimization, hindering the reuse of high-reward behaviors from policy exploration. While on-policy distillation alleviates this by consolidating independently optimized experts, its reliance on matching absolute expert distributions can yield suboptimal supervision, especially when the target model possesses a different prior or already surpasses the expert's capabilities. To alleviate this, we introduce Proxy OPD (P-OPD), an asynchronous post-training framework that transfers reward-induced policy improvements rather than absolute policy distributions. P-OPD first optimizes a proxy policy via reward feedback. It then extracts the relative distributional changes between the proxy's initial and optimized states, transferring these directional updates through the target model's own on-policy trajectories while retaining the target policy as the reference. This decoupled formulation requires the proxy to provide merely a useful direction of improvement rather than superior absolute capability, enabling update signals from older or weaker proxies to remain highly effective. Systematic experiments on Qwen3-family models across mathematical reasoning and code generation demonstrate that P-OPD consistently enhances already strong target models. Furthermore, transfer intensity can be dynamically modulated through signal scaling, making the extracted update signals seamlessly reusable across diverse model variants and training configurations. These results establish relative policy updates as highly reusable, adjustable assets for scalable, reward-based post-training.

arXiv

EHR-MPC: Inference-Time Control for Sepsis Treatment with Generative Patient Digital Twins

Joshua Pickard, Wei Qi, Na Li, Ann Woolley, Lisa Cosimi, Roy Kishony, Deborah Hung

Sepsis is a leading cause of mortality, yet optimal treatment policies remain contested. Existing reinforcement learning (RL) approaches learn fixed strategies for sepsis treatment, limiting adaptability to changing clinical objectives during inference. We propose EHRMPC, a framework that decouples learning patient dynamics from optimizing treatment by training a patient digital twin in the form of a generative electronic health record (EHR) model. The digital twin predicts clinical trajectories under interventions and enables model predictive control (MPC) to optimize treatments via inference-time planning over simulations. We evaluate EHR-MPC on a multicenter ICU sepsis cohort spanning 8 hospitals in the Mass General Brigham health system using both off-policy importance sampling and on-policy simulation-based evaluation. Relative to RL baselines, EHR-MPC achieves comparable off-policy performance and improved simulation performance. Unlike RL, this work frames sepsis treatment optimization as inference-time control over learned patient dynamics, establishing a general framework for decision making with generative clinical models.

arXiv

RSPO: Reward-Swap Policy Optimization for Multi-Turn LLM Agents

Qiang Liu, Taian Guo, Ruizhi Qiao, Xing Sun

Reinforcement learning holds significant potential for training large language models (LLMs) to handle multi-turn interactive tasks. However, in long-horizon, multi-turn tasks characterized by sparse outcome rewards, directly training with outcome rewards often results in slow convergence due to the sparsity of signals and the lack of fine-grained feedback. Furthermore, the model may fail to learn successful trajectories that are not sampled during training, thereby limiting its performance. Conversely, while employing customized dense process rewards provides richer signals and accelerates convergence, these surrogate rewards may exhibit potential misalignment with the ground-truth outcome rewards. This inconsistency can bias the training direction and ultimately degrade the model's final performance. In this work, we propose Reward-Swap Policy Optimization (RSPO), a method designed to leverage the rich information from dense process rewards to facilitate training with outcome rewards. By utilizing a reward-swap mechanism, RSPO ensures the diversity of sampled trajectories while guaranteeing consistency between the optimization objective and the true outcome rewards, thereby elevating the performance ceiling of the model. We conduct extensive experiments on two challenging agent benchmarks, WebShop and ALFWorld. By applying our method to various reinforcement learning algorithms, including GRPO, PPO, and GiGPO, we demonstrate that RSPO achieves consistent performance improvements across different baselines and benchmarks.

arXiv

Minimum Block Width for Universal Approximation by Residual Neural Networks with Inner Width One

Qi Zhou, Xuan Zhou, Xiao-Song Yang

In this paper, we study the universal approximation property of residual neural networks. For input and output dimensions $d_x$ and $d_y$, and LeakyReLU, ReLU, ReLU-like activation functions, the upper and lower bounds of the minimum block width are established. To achieve $L^p$ approximation $(1\leq p <+\infty)$ on any compact set, we show that the exact minimum block width is $\max\{d_x,d_y\}$ when each residual branch has inner width 1. Furthermore, we show that residual neural networks with block width $\min\{d_x+d_y, \max\{2d_x+1,d_y\}\}$ can achieve uniform approximation on any compact set under the constraint that each residual branch has inner width 1. Besides, for any activation function family, we prove that there exist functions that cannot be approximated by residual neural networks with block width less than $\max\{d_x, d_y\}$, both in the $L^p$ sense and the uniform sense, regardless of inner width. Consequently, for LeakyReLU, ReLU, ReLU-like activation functions and $d_y\geq 2d_x+1$, the exact minimum block width for uniform approximation is $d_y$ when each residual branch has inner width 1.

继续浏览

更多 AI 顶会论文

30+ 个顶会，按会议、年份和主题浏览。

浏览全部会议

进入任一会议和年份，查看完整论文列表。

全部会议 →

搜索论文

按标题、作者或关键词，在 21 万+ 篇论文里搜索。

搜索论文 →