AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents

Jingya Wang, Yuyang Gao, Liuzhenghao Lv, Yonghong Tian, Yuyang Liu

We introduce LabEvolver, a training-free framework that equips safe and grounded wet-lab agents with episodic memory from execution experience. LabEvolver couples a state-grounded inner trial loop for adaptive perception, online planning, and safety validation with an outer evolution loop that distills completed trajectories into reusable skill, strategy, and safety experience. On robotic solution-preparation tasks, LabEvolver demonstrates real-world feasibility, reducing pH-regulation completion time and safety-gate intercepts by 48.2% and 60.0%, respectively. On ALFWorld, it further improves cumulative success rate within 20 steps from 76.2% with ReAct to 91.4% over 500 continual tasks, showing generality beyond wet-lab settings. These results support learn-by-doing experience evolution as a feasible path toward closed-loop automated scientific discovery. The project page is available at https://andygao6186.github.io/LabEvolver/.

arXiv

Reason-Mediated Behavioral Models for Auditing LLM Social Simulators

Atharva Pandey, Gautam Jajoo

Large language models are increasingly used as social simulators, including as synthetic survey respondents. Most evaluations ask whether simulated outcomes resemble human outcomes. We argue that this is necessary but too weak: a simulator can match the final answer while using the wrong rationale-derived reason pattern. We study this problem through a 94-person sunscreen concept test in which each respondent evaluated three product concepts and wrote open-ended rationales. We map those rationales into signed reason states $Z$, where positive signs support adoption and negative signs block it. This gives a practical audit: holding respondent descriptors $D$, category context $K$, and concept treatment $X$ fixed, do human rationale-derived reasons help predict behavior $Y$, and can an LLM simulate the same reason state without seeing the human rationale or outcome? Human rationale-derived reasons substantially improve held-out prediction of purchase intent. LLM-simulated reasons are more brittle: they often sound plausible, but frequently echo the concept board rather than recover the respondent's acceptance or rejection path. The paper contributes an evaluation framework for social simulators. Reason states do not identify natural causal effects by themselves, but they provide an interpretable test of whether a simulator's stated reasons align with human evidence.

arXiv

From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement

Qinsi Wang, Jing Shi, Huazheng Wang, Kun Wan, Yiran Wu, Bo Liu, Qingyun Wu, Hai Helen Li, Yiran Chen, Handong Zhao, Wentian Zhao

Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as mathematics and coding, where correctness can be deterministically verifiable. Open-ended tasks instead often rely on human preferences, reward models, or LLM-based judges, introducing evaluation bias, judge capability bottlenecks, and additional inference costs. Drawing on the principle of self-supervised learning, which constructs pretext tasks to derive supervision from the data itself, we propose Reinforcement Learning with Self-Verifiable Rewards (RLSVR), a task-transformation-based training paradigm for extending RLVR to open-ended tasks. RLSVR transforms open-ended tasks into verifiable proxy environments whose internal rules and interaction outcomes automatically generate reward signals. We instantiate RLSVR with SpyRL, a Self-PlaY Reinforcement Learning method inspired by social deduction game Who Is the Spy?. Agents receive asymmetric information, complete the same target task, and vote to identify a designated spy. Because the spy identity is predetermined, voting outcomes provide fully verifiable rewards, while successful identification remains closely related to output quality. Experiments on text summarization, creative writing, and mathematical reasoning show that SpyRL outperforms existing self-improvement methods on non-verifiable tasks and yields consistent gains on verifiable reasoning tasks. These results demonstrate that task transformation can extend scalable RLVR-based self-improvement beyond inherently verifiable domains. Models and code have been released at https://github.com/wangqinsi1/RLSVR/tree/SpyRL.

arXiv

Mission-Level Runtime Assurance for LLM-Assisted ISR Swarms over a Verification-Aware Fabric

Nikolaos Kekatos, Panagiotis Katsaros, Alexios Lekidis, Theodoros Nestoridis, Tom Nianios

Swarms of LLM-assisted autonomous robots are increasingly proposed for cooperative intelligence, surveillance, and reconnaissance (ISR) in contested environments. A growing class of their assurance failures arises not within any single platform but across the swarm: individually-compliant actions compose into a mission-level violation: a prohibited objective split across platforms to evade per-platform lim- its, or a collective budget quietly exceeded. Per-platform guardrails miss these by construction, and contested communications let the violation hide behind lost or delayed evidence. We present a three-tier (platfor- m/squad/mission) compositional runtime-verification framework that de- composes a mission policy into per-agent and cross-agent aspects, aggre- gates per-platform verdicts over a verification-aware messaging fabric, and fuses them with an evidence-aware, two-axis (security x complete- ness) algebra whose provenance names the platforms that jointly trig- gered a violation. Because the fabric makes evidence loss and silence observable, unsupported negative verdicts are downgraded to an explicit unknown rather than reported as mission-wide all-clears. On a simulated ISR mission, an indirect prompt injection that causes real LLM planners to split a prohibited collection task across four platforms is invisible to every per-platform monitor yet detected compositionally with full prove- nance; under an injected fault campaign a best-effort central monitor emits silent false all-clears while the verification-aware fabric emits none

arXiv

DynaResize: Runtime GPU Reallocation for Disaggregated LLM Post-Training

Hanlin Du, Zhiyuan Yan, Yungang Bao, Sa wang

RL-based LLM post-training increasingly disaggregates Rollout and Training across separate GPU resources, but static GPU partitioning suffers from severe pipeline bubbles under long-tail rollout latency. We present DynaResize, a runtime GPU reallocation system that dynamically switches GPUs between Rollout and Training to balance stage execution times without changing RL semantics. DynaResize decomposes resizing into fine-grained operations and removes non-startup-critical work from the critical path through communicator reuse, bounded state staging, and hysteresis-based resizing. Experimental results show that DynaResize can improve end-to-end throughput by 66.5% and reduce total execution time by 33% over the optimal static configuration, while hiding 27% of role-switching overhead.

arXiv

Quality Action Assurance: Multimodal Verification of Examiner Claims in VR OSCEs

Harry Rogers, Sally Shiels, Ashley Tomlinson, James Thomas, James Aylward, Nathan Gauge, Helen Higham, Alison Noble

Objective Structured Clinical Examinations (OSCEs) are the gold standard for assessing clinical competence, yet scoring remains vulnerable to examiner subjectivity, fatigue, and cognitive bias. Standard examiner validation via inter-rater statistics lacks explanatory power regarding the source of errors, as it neither analyzes examiner reasoning nor verifies examiner claims against actual events. Thus, we introduce Quality Action Assurance (QAA), a multimodal framework that verifies examiner claims in Virtual Reality (VR) pediatric OSCEs by comparing actions claimed by examiners against a reference record of events constructed from video, VR logs, and actor annotations. QAA combines a constrained temporal action alignment model, which performs action localization and actor source attribution, with a large language model that extracts examiner claims and checks them against the record. Across a 5-fold cross-validation, QAA achieves 99.2\% $\pm$ 0.7\% Actor F1 and 93.4\% $\pm$ 1.9\% W@16 for temporal alignment. Overall, QAA detects examiner errors with 69.9\% precision and 76.7\% recall; in retrospective evaluation, correcting the detected errors raises the share of factually correct transcripts from 39.2\% to 79.2\%, supporting fairer OSCE quality assessment.

arXiv

NeurOWL: An LLM-Based Neural-symbolic Framework for Incomplete OWL Ontology Reasoning

Hui Yang, Jiaoyan Chen, YIPING SONG, Renate Schmidt, Wen Zhang

OWL ontologies provide a formal knowledge representation framework that enables semantic reasoning, and have been widely adopted across domains such as healthcare and bioinformatics. In practice, however, real-world ontologies are often incomplete, which pose challenges for reasoning. In this work, we focus on a fundamental subsumption reasoning problem: given an incomplete ontology and a candidate (non-entailed) subsumption, determine whether the subsumption is semantically plausible and, if so, providing a logically sound explanation containing potential missing axioms. This task unifies subsumption verification with ontology abduction, and generalizes the latter by removing the need for a predefined candidate set of missing axioms. To address this subsumption reasoning problem, we propose NeurOWL, an end-to-end neuro-symbolic framework that jointly performs verification and abduction, leveraging both formally defined semantics and textual semantics through Large Language Models and ontology embeddings. We evaluate NeurOWL on real-world ontologies across multiple domains, demonstrating strong and robust performance across different domains.

arXiv

AuEmoChat: Authentic Emotion Understanding and Rendering for Conversational Speech Synthesis

Zhenqi Jia, Yuan Zhao, Aruukhan, Rui Liu, Haizhou Li

Conversational Speech Synthesis (CSS) aims to synthesize speech with human-like emotional expression and contextual consistency in user-agent interactions. Existing CSS methods struggle to render authentic human emotions due to limited predefined emotion label spaces (e.g., seven emotion categories), while redundant multimodal tokens in multi-turn dialogue history interfere with context understanding. To address these issues, we propose AuEmoChat, a CSS framework for authentic emotion understanding and rendering. First, we develop AuEmoCodec, which learns a discrete authentic emotion token space from large-scale emotional speech via finite scalar quantization, enabling a more authentic emotion representation than limited basic emotion categories. Furthermore, we propose AuEmoToMe, an authentic-emotion-guided token merging algorithm that merges redundant tokens in multimodal dialogue history while preserving emotion-relevant context. We integrate it into an autoregressive text-speech model to predict the target authentic emotion token and speech tokens. Finally, we propose Authentic Emotion Flow Matching, which renders speech by jointly conditioning on merged dialogue context, target authentic emotion, and acoustic priors. Extensive experiments on the NCSSD-EmCap dataset demonstrate that AuEmoChat outperforms state-of-the-art CSS baselines and generates more expressive and authentic emotional speech. The code and speech demos will be available at: https://github.com/AI-S2-Lab/AuEmoChat.

arXiv

The Capability Convergence Hypothesis: Capability from Access Structure, Not Scale

Wenhui Chen, Jianlin Chen, Ziyao Lin, Chi-Man Vong

The Platonic Representation Hypothesis (PRH) holds that as models scale, representations of heterogeneous networks converge toward a shared model of reality. We propose its sequel and boundary, the Capability Convergence Hypothesis (CCH): under a fixed per-token inference budget, representational convergence does not entail capability convergence. Capability instead converges toward a class, the access-complete hybrid: any architecture holding both a compressive O(1)-state channel and a scalable verbatim-index channel. We anchor it on a witness task, the Newton's-apple problem in an infinite stream, and name three resource walls: a Shannon wall barring any o(Nb)-state architecture, a horizon wall barring any fixed window, and a circuit wall barring fixed-depth attention-only composition (conditional on TC0 != NC1). Under an explicit separability assumption a hybrid crosses all three by paying each wall's price, so capability is strictly super-additive under composition. We separate what we prove from what we conjecture: the access-completeness principle rests on information-theoretic lower bounds and pre-registered experiments, while the field-level convergence trend is an economics-motivated conjecture. We report the first pre-registered small-scale tests under criteria frozen before the data: the predicted scissors gap is measured (exact-retrieval error 0.994 vs. 0.000 once a 64-scalar state gains one global-attention layer), the state-tracking bifurcation lands at the registered boundary, and a conjunction witness shows an irreducibly two-channel solution; one prediction failed with its direction reversed and is reported as such. Representational convergence is given freely by scale; capability convergence must be purchased by access structure.

arXiv

BeatEdit: Symbolic Music Generation as Explicit Editing

Haoyu Gu, Lekai Qian, Haowu Zhou, Qi Liu, Shuai Wang

Music creation is fundamentally a process of revision. Yet symbolic music generation remains dominated by paradigms that produce complete sequences from scratch, with limited support for selective modification. Edit-based methods have proven effective for text transformation tasks, but remain largely unexplored for symbolic music. We trace this absence to the representational level: conventional event-based music encodings lack the structural properties required by explicit music editing. In contrast, the BEAT encoding, a beat-grid-anchored representation originally designed for autoregressive generation, possesses structural properties amenable to editing. We propose BeatEdit, the first framework for symbolic music generation based on explicit edit operations, recasting generation as producing new content by editing a draft rather than synthesizing from scratch. BeatEdit comprises three complementary mechanisms along an axis of increasing edit density: per-token sequence tagging for error correction, iterative refinement for accompaniment editing, and tag-then-fill for segment completion. All these mechanisms share a single encoding and pre-trained backbone, achieving higher precision and perceptual quality than autoregressive and diffusion methods across all three tasks, while remaining efficient, with single-pass inference completing in under 100 ms. Cross-encoding evaluation further reveals that encoding design substantially influences editing effectiveness, with notable encoding-method interaction effects. Code is available at https://github.com/Haoyu-Gu/BeatEdit-code

arXiv

Solution Space Path Planning: A Real-Time Human-Centered Path Planning Algorithm for En-Route Air Traffic Control

Yiyuan Zou, Wenying Lyu, Clark Borst

As technology advances, various algorithms have been proposed for air traffic management, yet their operational adoption in tactical control remains limited. This gap motivates a human-centered design emphasizing algorithmic interpretability, controller-relevant operational constraints, and real-time computation. Inspired by the interpretability and flexibility of solution-space displays, as well as by the decision logic controllers naturally apply when enforcing operational constraints, this study extends the solution-space concept to path planning and develops a fast conflict-free path-planning algorithm for en-route Air Traffic Control (ATC), termed Solution Space Path Planning (SSPP). The algorithm integrates three intent-based conflict detection methods---distance-based, time-interval-based, and zone-based---within the solution-space framework to identify conflict-free paths in computationally efficient ways. SSPP is developed using both vertex-based and edge-based search nodes, resulting in two variants---SSPPV and SSPPE, respectively. Empirical results show that SSPPV paired with zone-based conflict detection performs best, computing paths in 3.69 ms on average in the Dutch Delta sector using a 5 nmi grid. SSPPV remains approximately 3.77 times faster than SSPPE while offering competitive effectiveness, making it suitable for time-critical operations and interactive 'what-if' probing in real time. An extension to SSPPV and SSPPE further examines the trade-off between delay minimization and separation requirements, demonstrating the flexibility of SSPP in revising optimization objectives. This study not only proposes a novel path-planning algorithm but also shows how such algorithms can be designed to align with human use and operational requirements, supporting their integration into future ATC systems.

arXiv

LabGuard: Grounding Natural-Language Laboratory Rules into Runtime Guards for Embodied Laboratory Agents

Jingpu Yang, Fengxian Ji, Zhengzhao Lai, Zhexuan Cui, Guangxian Ouyang, Qian Jiang, Fan Zhang, Min Peng, Qianqian Xie, Preslav Nakov, Zhuohan Xie

Scientific embodied agents are increasingly capable of carrying out laboratory procedures, but executing these procedures safely in dynamic laboratory environments remains challenging. Current safety approaches often overlook the intermediate step of transforming laboratory natural language, including safety rules, manuals, protocols, and standard operating procedures, into machine-checkable runtime constraints. We introduce LabGuard (Laboratory Guard), a language-to-execution safety suite that grounds natural-language laboratory rules into executable specifications and deploys them as runtime guards. LabGuard includes three core components: LabGuard-IR, which defines a typed executable representation; LabGuard-Bench, which provides 812 supervised annotations expanded from 203 seed laboratory rules; and LabGuard-Grounder, which maps natural-language laboratory rules into LabGuard-IR. The resulting IR instances are handled by the LabGuard Pipeline, which compiles them into runtime monitors and applies them at the controller boundary. Experiments show that LabGuard generalizes to unseen laboratory-rule sources, achieves 79.4 task-scope F1, and reduces unsafe events from 39.5% to 23.8% after monitor compilation. In LabUtopia, its runtime monitors integrate with ACT, keeping interventions below 0.5% while preserving task success.

arXiv

Latent Actions from Factorized Transition Effects under Agent Ambiguity

Heejeong Nam, Chandradithya S Jonnalagadda, Harshit Aggarwal, Eric Xu, Randall Balestriero

Latent Action Models (LAMs) learn action-like proxies from observation. However, in multi-object or distractor-rich scenes, observations contain not only agent motion but also distractors, camera dynamics, and background changes, making recovery of the underlying action intrinsically ambiguous without supervision. We argue that the appropriate unsupervised target is therefore not the true action itself, but a state-conditioned compositional summary of the transition effects present in the scene, enabling better action alignment and more effective utilization of action supervision. To this end, we propose a two-stage framework. We first pretrain Observed Transition Factorization (OTF) to discover reusable local transition primitives using a compositional codebook. We then aggregate these primitives into compact state-conditioned latent actions, instantiated as OTF-LAM-Pixel under the standard inverse-forward dynamics framework and OTF-LAM-Dino, a decoder-free variant operating in a frozen DINOv2 representation space. Experiments show that the learned transition primitives transfer across visual appearance and morphology, while the resulting latent actions exhibit substantially stronger action alignment than existing LAMs, make more effective use of downstream action supervision, and achieve competitive or superior downstream policy performance.

arXiv

Cognitive World Model for Progressive BDI/E Trajectory Evaluation of Conversational Agents

Minghui Ma, Bin Guo, Hao Wang, Han Wang, Mengqi Chen, Jingqi Liu, Yan Liu

As LLM-based conversational agents advance toward increasingly open-ended and interaction-intensive scenarios, task completion alone provides an incomplete assessment of their effectiveness. The evolution of users' internal states, including beliefs, desires, intentions, and emotions (BDI/E), serves as an intermediate signal connecting agent behaviors with interaction outcomes and reflects how conversational strategies shape users during multi-turn interactions. However, existing evaluation paradigms primarily focus on surface-level responses or final outcomes, providing limited insight into the underlying cognitive processes. This limitation makes it difficult to diagnose why agents succeed or fail and to optimize their interaction strategies. To address this challenge, we propose Cognitive World Model (CogWM), an LLM-based cognitive user model that jointly models users' BDI/E states and corresponding responses, enabling explicit cognitive trajectory tracking. Trained on 150K user-turn samples with Qwen3-14B, CogWM achieves superior performance over existing user simulation baselines in both response fidelity and cognitive state understanding. Interactions with six state-of-the-art LLMs demonstrate that CogWM enables progressive comparison of agents through cognitive trajectories, revealing distinct agent patterns and complementary relationships between cognitive evolution and behavioral outcomes.

arXiv

ReSum: Synergizing LLM Reasoning and Summarization with Reinforcement Learning

Xucong Wang, Ziyu Ma, Yong Wang, Shidong Yang, Hailang Huang, Renda Li, Pengkun Wang, Xiangxiang Chu

Reinforcement Learning with Verifiable Rewards (RLVR) is a central technique for improving long-horizon reasoning in Large Language Models (LLMs). However, existing RLVR methods often encourage unnecessarily long reasoning rollouts, which can degrade reasoning coherence and exhaust the available context budget. Existing approaches to long-context organization often depend on external mechanisms to organize rollouts, rather than enabling the model to manage its own reasoning trajectory. To address this limitation, we propose ReSum, a novel RLVR framework that enables LLMs to compress and organize their reasoning trajectories through self-summarization. Our pilot studies show that self-summarization stabilizes generation by lowering token-level entropy, and that introducing a ``summarization'' phrase can substantially mitigate errors propagated from an incorrect rollout prefix. Motivated by these findings, ReSum adopts a summarization-aware adaptive rollout mechanism that contrastively evaluates whether self-summarization benefits the ongoing reasoning process. Specifically, when the model spontaneously triggers self-summarization, ReSum masks the summarization phrase to create a contrastive branch; for non-summarization positions, it instead randomly injects the phrase to create a matched branch. We further design a summarization-aware advantage to enable finer-grained comparison between contrastive rollout trajectories. Extensive experiments show that ReSum improves performance at an average of 4\% while reducing rollout length by 18.6\%.

arXiv

Role-Agent: Bootstrapping LLM Agents via Dual-Role Evolution

Xucong Wang, Ziyu Ma, Shidong Yang, Tongwen Huang, Pengkun Wang, Yong Wang, Xiangxiang Chu

Although Large Language Model (LLM) agents have demonstrated strong performance on complex tasks, their learning is often limited by inefficient interaction feedback and static training environments, which hinder broader generalization. To address these limitations, this paper introduces Role-Agent, \textcolor{black}{a framework} that harnesses a single LLM to function concurrently as both the agent and the environment, enabling a bootstrapped co-evolution. Role-Agent comprises two synergistic components: World-In-Agent (WIA) and Agent-In-World (AIW). In WIA, the LLM acts as the agent and predicts future states after each action; the alignment between predicted and actual states is then used as a process reward, encouraging environment-aware reasoning. In AIW, the LLM analyzes failure modes from failed trajectories and retrieves tasks with similar failure patterns, thereby reshaping the training data distribution for targeted practice. Experiments on multiple benchmarks show that Role-Agent consistently improves performance, yielding an average gain of over 4\% over strong baselines.

arXiv

A Multi-Agent System for Motor Design Optimization via an FEA-AI Hybrid Approach

Jinseong Han, Sunwoong Yang, Namwoo Kang

This study presents a large language model (LLM)-based multi-agent framework for interior permanent magnet synchronous motor (IPMSM) design optimization that mitigates limitations of conventional workflows: expertise-dependent problem setup and data preparation, the prohibitive computational cost of finite element analysis (FEA), and the unreliability of AI surrogates in unexplored regions. To this end, we first introduce a Design agent that formulates the optimization problem in natural language, leveraging retrieval-augmented generation to improve answer accuracy on motor design problems from below 50% to 67-80%. Furthermore, a Training agent autonomously repairs improperly defined design spaces by reasoning over solver failure history, raising the success ratio of the geometry sampling from 28% to 84% for AI training. Additionally, to resolve cost and reliability simultaneously, an Optimization agent employs an uncertainty-aware FEA-AI hybrid model: the AI surrogate is the primary evaluator, and FEA is selectively invoked where predictive uncertainty is high. Under the same FEA budget, this hybrid model achieves up to 44% lower iron loss in single-objective and 22.5% higher hypervolume in multi-objective optimization than conventional FEA-only search. Under the same evaluation budget, it reduces computation time by 52-55% while retaining 90-92% of FEA-only hypervolume. Conversely, AI-only search converges to false optima, leaving half its Pareto designs infeasible. Notably, a controller agent adaptively updates the uncertainty threshold that triggers FEA each round, eliminating manual tuning and achieving 5.8% lower single objective iron loss than with a fixed threshold. These results establish domain specialized LLM agents with uncertainty-aware hybrid evaluation as a reliable, scalable paradigm for simulation-driven design automation.

arXiv

MemForest: An Efficient Agent Memory System with Hierarchical Temporal Indexing

Han Chen, Zining Zhang, Wenqi Pei, Bingsheng He, Ming Wu, Jason Zeng, Michael Heinrich, Wei Wu, Hongbao Zhang

Memory is a fundamental component for long-context LLM agents, supporting persistent state across interactions through a continuous serve-and-update lifecycle. Despite substantial prior work, many stateful systems retain sequential autoregressive extraction or state-dependent maintenance on the write path, delaying when new evidence becomes queryable. To address these challenges, we present MemForest, a memory framework that reformulates agent memory as a write-efficient temporal data-management problem. MemForest breaks the sequential bottleneck via parallel extraction, decoupling memory construction into concurrent, independent operations. We further introduce MemTree, a hierarchical temporal index that organizes memory as time-ordered trees and replaces global rewrites with localized dirty-path refresh. Dirty summaries can be refreshed in parallel across nodes and trees. End-to-end work remains proportional to incoming content; the logarithmic bound applies only to structural insertion and level-dependent refresh depth in balanced trees. We evaluate MemForest on two long-context benchmarks, LongMemEval-S and LoCoMo. Experiments use Qwen3-4B, Qwen3-30B, and Gemma-4-12B-IT. With Qwen3-30B, MemForest reaches 81.8 percent pass at 1 on LongMemEval-S, while its input-normalized build rate is 6.0 times that of EverMemOS. On LoCoMo categories 1 to 4, it reaches 84.09 percent, within 0.13 percentage points of EverMemOS; on a matched conversation, its build rate is 9.5 times higher. These results show that MemForest reduces memory-freshness latency while retaining strong answer quality.

arXiv

PAIR: Prefix-Aware Internal Reward Model for Multi-Turn Agent Optimization

Wonjoong Kim, Yeonjun In, Sangwu Park, Dongha Lee, Chanyoung Park

A significant hurdle for current LLMs is the execution of complex, multi-stage tasks. Group Relative Policy Optimization (GRPO) has been emerging as a leading choice, but its reliance on sparse outcome rewards severely limits credit assignment across intermediate steps. Existing remedies such as running full rollouts to assign step-level advantages, calling external LLM judges at each step, or computing intrinsic rewards that require ground-truth answers at every evaluation introduce significant costs or practical constraints. We hypothesize that internal correctness probing over LLM hidden states can be repurposed as a step-level reward signal, potentially addressing all of these limitations at once. However, existing probing research assumes clean inputs, and we first show that this assumption breaks down in multi-step settings: hidden-state probes degrade severely under prefix contamination tracking coherence with the (possibly corrupted) prefix rather than grounded correctness, while attention-based features remain robust to contamination but underperform on clean prefixes. Building on this complementary relationship, we propose the Prefix-Aware Internal Reward (PAIR), a two-stage model with a frozen hidden-state probe estimating belief-consistency and a lightweight attention-based head correcting it toward grounded correctness. Experimental results show that PAIR achieves the highest AUROC on contaminated trajectories while operating at negligible inference cost, enabling dense step-level reward signals for GRPO training without external model calls, ground-truth dependencies, or full-trajectory rollouts.

arXiv

Dual-Dimensional Consistency: Balancing Budget and Quality in Adaptive Inference-Time Scaling

Rongman Xu, Yifei Li, Tianzhe Zhao, Yanrui Wu, Bo Li, Hang Yan

Large Language Models (LLMs) have demonstrated remarkable abilities in reasoning. However, maximizing their potential through inference-time scaling faces challenges in trade-off between sampling budget and reasoning quality. Current strategies remain inefficient as they typically treat sampling width and depth as orthogonal objectives, where width consensus methods risk reinforcing hallucinations, while depth pruning mechanisms prematurely truncate complex yet valid reasoning chains. Therefore, we propose Dual-Dimensional Consistency (DDC), a unified framework that bridges path quality with adaptive termination. By coupling Confidence-Weighted Bayesian protocol with a Trend-Aware Stratified Pruning, our method ensures that computational resources are concentrated on high quality reasoning paths, filtering hallucinations while accelerating consensus. Evaluations across five benchmarks demonstrate that this approach reduces token consumption by over 10 times while maintaining or exceeding the accuracy of strong baselines across various LLMs.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.