Pengxin Zha, Yi Lu, Haojun Ma, Jun Zhu, Minghao Qiu, Xiaoyan Wang, Wei Shan, Maoqing Yao
Wheeled humanoid robots carry both arms on one actuated waist, usually without torque sensors, so the accuracy of both end-effectors depends on a single waist model. After a first stage that estimates the tracker and target extrinsics, we identify the waist and both arms in one problem from the target positions on both flanges, measured by an optical tracker. The kinematic model is complete, with waist-base, parallel-axis, and branch parameters; arm compliances are driven by measured torques and waist compliances by gravity torques of the robot's mass model; and Gaussian priors with evidence-maximizing scales stabilize the estimate. Independent and sequential single-chain calibration are restrictions of this problem, and its waist information is the sum of the single-chain Schur complements. On two robots with different waists, each without and with a payload, calibration reduces the mean errors on held-out poses from 11.3-31.1 mm to 0.50-1.11 mm. For every training-set size from 20 poses to all, the method has 17-63% lower mean errors than adaptations of two published whole-body calibration methods, mainly through the waist compliance. Independent calibration yields two waist models; using either for both arms raises the other arm's error to up to 3.21 mm.
In practice, incentive providers (i.e., principals) often cannot observe the reward realizations of incentivized agents, which is in contrast to many principal-agent models that have been previously studied. This information asymmetry challenges the principal to consistently estimate the agent's unknown rewards by solely watching the agent's decisions, which becomes even more challenging when the agent has to learn its own rewards. This complex setting is observed in various real-life scenarios ranging from renewable energy storage contracts to personalized healthcare incentives. Hence, it offers not only interesting theoretical questions but also wide practical relevance. This paper explores a repeated adverse selection game between a self-interested learning agent and a learning principal. The agent tackles a multi-armed bandit (MAB) problem to maximize their expected reward plus incentive. On top of the agent's learning, the principal trains a parallel algorithm and faces a trade-off between consistently estimating the agent's unknown rewards and maximizing their own utility by offering adaptive incentives to lead the agent. For a non-parametric model, we introduce an estimator whose only input is the history of principal's incentives and agent's choices. We unite this estimator with a proposed data-driven incentive policy within a MAB framework. Without restricting the type of the agent's algorithm, we prove finite-sample consistency of the estimator and a rigorous regret bound for the principal by considering the sequential externality imposed by the agent. Lastly, our theoretical results are reinforced by simulations justifying applicability of our framework to green energy aggregator contracts.
Motivated by a number of real-world applications from domains like healthcare and sustainable transportation, in this paper we study a scenario of repeated principal-agent games within a multi-armed bandit (MAB) framework, where: the principal gives a different incentive for each bandit arm, the agent picks a bandit arm to maximize its own expected reward plus incentive, and the principal observes which arm is chosen and receives a reward (different than that of the agent) for the chosen arm. Designing policies for the principal is challenging because the principal cannot directly observe the reward that the agent receives for their chosen actions, and so the principal cannot directly learn the expected reward using existing estimation techniques. As a result, the problem of designing policies for this scenario, as well as similar ones, remains mostly unexplored. In this paper, we construct a policy that achieves a low regret (i.e., square-root regret up to a log factor) in this scenario for the case where the agent has perfect-knowledge about its own expected rewards for each bandit arm. We design our policy by first constructing an estimator for the agent's expected reward for each bandit arm. Since our estimator uses as data the sequence of incentives offered and subsequently chosen arms, the principal's estimation can be regarded as an analogy of online inverse optimization in MAB's. Next, we construct a policy that we prove achieves a low regret by deriving finite-sample concentration bounds for our estimator. We conclude with numerical simulations demonstrating the applicability of our framework to personalized adherence-improving incentive design for long-term statin treatments.
Yutao Wu, Xiao Liu, Ming Wen, Hanxun Huang, Jun Wang, Xingjun Ma
Web agents increasingly browse, shop, and manage accounts on behalf of their users. Prior work has shown that they can be misled by adversarial content, including prompt injections and deceptive interfaces. Beyond such attacks, agents operating on the open web also encounter online fraud designed to deceive humans, including fraudulent online stores, fake investment platforms, and advance-fee schemes. How web agents respond to such fraud during task execution remains underexplored, in part because scam websites are short-lived, making systematic evaluation difficult. To address this gap, we introduce ScamWorld, an interactive environment that reconstructs reported scam websites across diverse scenarios, fraud mechanisms, and languages. Each reconstructed website reproduces documented fraud mechanisms, making relevant signs of fraud observable before the agent's final action and enabling controlled, repeatable evaluation. Through an extensive study of web agents powered by six frontier LLMs, we find that baseline deception rates range from 11% to 35% across models. Agents proceed more readily with payments than with requests for credentials or wallet information. A closer analysis of agent trajectories shows that some agents acknowledge suspicious conditions yet proceed, at times treating the user's instructions as authorisation to accept additional terms imposed by the website. ScamWorld provides a reusable testbed for systematically evaluating and stress-testing web agents against real-world online fraud.
The forward and reverse Kullback-Leibler (KL) divergences arise as limiting objectives in learning and inference, yet they induce markedly different inductive biases that cannot be explained by expectation terms alone. We introduce the Surprisal-Rényi Free Energy (SRFE), a normalized Rényi-Chernoff log-moment functional of the likelihood ratio. Although SRFE is algebraically related to Rényi divergence for fixed values of the order parameter, its endpoint-symmetric normalization yields a distinct free-energy viewpoint for analyzing how forward- and reverse KL behavior emerges near singular limits. We show that SRFE recovers the forward and reverse KL divergences as endpoint limits and derive local expansions around both regimes in which the variance of the log-likelihood ratio appears as a first-order correction away from the KL objective. These expansions reveal an explicit mean-variance tradeoff governing departures from KL-dominated behavior. We further establish a Gibbs-type variational characterization in terms of weighted KL projections and prove that SRFE controls large deviations of excess code length through Chernoff-type bounds, giving a Minimum Description Length interpretation of the functional. We evaluate SRFE on two synthetic settings (two-dimensional Gaussian-mixture variational inference and multi-class classification) and the long-tailed CIFAR-10 and CIFAR-100 (CIFAR-LT) datasets. On CIFAR-LT, retraining the best configuration of each loss over five seeds, SRFE matches cross-entropy accuracy on CIFAR-10-LT and gives mixed results on CIFAR-100-LT, improving few-shot accuracy at one of three imbalance factors. Together, these results position SRFE as a variance- and tail-sensitive, normalized Rényi-Chernoff free energy that clarifies the geometric, statistical, and large-deviation structure underlying the transition between forward- and reverse-KL limits.
Predicting stock returns requires modeling persistent company relations and time-varying market signals. Graph Neural Network (GNN) approaches combine static priors with dynamic relational graphs, but treat every extracted edge as a trustworthy binary fact, ignoring heterogeneous reliability across relational sources. We propose LENS (LLM-Extracted priors and News-driven Semantics), a static--dynamic graph framework for stock return prediction and ranking that models relational reliability as a first-class signal across the pipeline. On the static side, we extract a multi-relational company graph from a Large Language Model and aggregate it over multiple runs with confidence weighting to downweight low-consistency edges. On the dynamic side, we build a news-driven semantic graph with exponential-decay temporal aggregation to capture changing market context under uneven coverage. A context-dependent gated fusion module rebalances the two graphs at the per-stock, per-day level. On S&P 500, LENS improves AR@5 by 16.4% and SR@5 by 22.6% over the strongest baseline, and on NASDAQ 100 it achieves comparable improvements (15.5% and 20.9% respectively). Ablations isolate each component's contribution, and a human-validation study of sampled relations supports that the aggregation is grounded, providing evidence that reliability-aware LLM relations and news-based semantic graphs carry complementary information for graph-based financial prediction.
Fang Wu, Hongbin Lin, Bingchen Duan, Yuchen Wang, Li Erran Li, Yanjie Huang, Mingxuan Cao, Yijun Ma 等 27 位作者
Protein-protein interactions underpin nearly all biological processes. Yet, designing proteins with desired interactions remains challenging due to the structural complexity governing molecular recognition. This study views protein design as a search optimization problem over a combinatorial sequence space under expensive structural evaluation. Toward this goal, we introduce a novel tree-based search framework, ProteoCraft, that employs approximate structural gradients from diffusion-based structure predictors as effective proposal policies to specify where the search should explore. Complete no-gradient inference of folding models separately assigns values to the resulting sequences, determining what is good. ProteoCraft allocates this verification budget through a coarse-to-fine paradigm: multi-site edits first explore distinct sequence basins, then verified promising states undergo strict single-site refinement. Across fourteen de novo binder-design settings spanning eight target proteins, five immune-relevant TCR-facing targets, and four localized CDR redesign tasks, ProteoCraft improves design quality and high-confidence yield. These improvements persist under independent AF2-IG and Rosetta-based interface evaluation. Prospective experimental validation further yielded an 82.9% binding hit rate for PD-L1 miniproteins, with a best measured KD of 0.879 nM, and confirmed binding of designed IFNAR2 VHH. Our code is available at https://github.com/smiles724/ProteoCraft_v1.
In Polyamorous Scheduling Problem, we are given an edge-weighted graph and must find a periodic schedule of matchings in this graph which minimizes the maximal weighted waiting time between consecutive occurrences of the same edge. This NP-hard problem generalises Bamboo Garden Trimming and is motivated by the need to find schedules of pairwise meetings in a complex social group. We present a $4 G^*$ algorithm for Polyamorous Scheduling, improving the previously known bound of $3 + \sqrt{5} \approx 5.236$. Our algorithm is inspired by the Deadline-Driven Heuristic which is optimal for Bamboo Garden Trimming (BGT).
Recursive self-improvement (RSI) in production agents involves transforming accumulated experience into reliable model updates while adapting to evolving business policies. We present CARE (Customer-service Agent Retrospective and Evolve), a grounded RSI framework that operationalizes this process in customer service. Its goal is to replace a frontier proprietary incumbent with an in-house mid-sized model that continues learning after deployment. After a one-time cold start on governed production traces, the current model serves production traffic alongside the incumbent, and its own failures drive the next update. A 24/7 inspection process establishes failures through evidence-blinded, independently reviewed judgments. Attribution separates model errors from defects in Skills, tools, or the harness, routing component failures to repair. Model-attributable trajectories are truncated at the first erroneous decision and paired with verified repairs. On-policy self-distillation (OPSD) then updates the model using the same model as teacher, with the diagnosis and repair supplied as privileged information unavailable to the student. To support continued evolution as business requirements change, training mixes failures from different Skill/SOP versions, each judged against the policy in force when it occurred. Candidates run an executable simulation test compiled from the latest SOP and then frozen, followed by offline inspection of the resulting traces. Promotion requires a combined simulation-and-inspection score at least at parity with the deployed version, confirmation through an online comparison with the incumbent, and human approval.
Xueqi Ma, Yanbei Jiang, Xingjun Ma, James Bailey, Sarah M. Erfani
Aligning large language models (LLMs) with multiple, potentially conflicting human values remains challenging, as existing multi-objective alignment methods often require costly fine-tuning and provide limited insight into the internal mechanisms underlying value-related behaviors. In this work, we investigate multi-value alignment from a mechanistic interpretability perspective by identifying value-sensitive attention heads (value heads) in LLMs. Using a probing-based value localization framework, we discover sparse sets of attention heads associated with diverse dimensions, spanning both human values (e.g., harmlessness, honesty) and stylistic preferences (e.g., coherence, complexity). Analyses across multiple LLM families reveal functional specialization and structured interactions among value heads, including conflicts between certain values. Targeted head ablations further induce substantial, value-specific behavioral changes, providing causal evidence of their functional importance. Building on these findings, we propose an inference-time activation steering strategy that enables controllable alignment with individual values or combinations of values without updating the underlying model parameters, supported by theoretical analysis. Experiments across multiple value alignment benchmarks demonstrate effective single- and multi-value control while largely preserving non-target model performance. Our findings connect mechanistic interpretability with flexible multi-value alignment, offering a foundation for more interpretable and pluralistic LLM behavior. The code is available at https://github.com/sihuo-design/ValueHeads.
Retrieval-augmented generation (RAG) extends large language models with external evidence, yet relevant evidence does not ensure that a generator can use it. Indiscriminate retrieval can degrade task performance while adding retrieval and long-context inference costs. Existing methods neither determine before large-scale execution whether a generator, task distribution, and retriever configuration is worth deploying nor diagnose multiple candidate retrievers under a shared budget. We introduce EviCast, a framework that forecasts retrieval value through interventional evidence-state decomposition. EviCast measures a generator's response to controlled evidence states once and combines this shared profile with the offline evidence-state distribution of each candidate retriever, yielding the end-to-end net utility and its 95% lower confidence bound without executing every RAG configuration. A state-refinement analysis characterizes when its binary and topology-aware instantiations are exact and when shared calibration is statistically more efficient than per-retriever calibration. In frozen evaluations, EviCast correctly decides all 14 HotpotQA retrieval configurations with a forecast error of 2.75 F1 points while reducing diagnostic generation calls by 80%, and it attains the lowest forecast error among seven predictors in the low-budget regime. Its topology-aware instantiation, EviCast-T, captures the complementarity between terminal and bridge evidence, correctly deciding all 16 configurations on MuSiQue and HotpotQA and 23 of 24 after prospective transfer to a third dataset. These results show that retrieval deployment depends on both evidence availability and the generator's ability to convert evidence structure into stable net utility. Code is available at https://github.com/YiboWang2002/EviCast.
AI agents often rely on persistent memory to retain user preferences and operational rules across sessions. However, this capability introduces a new safety risk: malicious instructions from messages, documents, or Web content can be retained in memory and influence future tasks even after the original attack content is removed. Existing attack success metrics, however, fail to capture who is affected and how attacks translate into downstream harm during otherwise normal task execution. We introduce MemHarmBench, a victim-centric benchmark for evaluating cross-session prompt injection in memory-persistent AI agents. MemHarmBench considers three categories of potential victims: users, workspace/project owners, and external stakeholders. Our evaluation first exposes an agent to malicious instructions and tests whether they are retained in memory. We then remove the attack content and initiate a new session with a legitimate task while preserving the agent's memory. We jointly evaluate whether the agent follows the attacker's intent, completes the legitimate task, and exhibits observable execution anomalies. Across multiple agent systems, we find that cross-session attacks can remain behaviorally stealthy: agents often complete tasks normally and exhibit no detectable anomalies while still making critical decisions aligned with the attacker's intent. By connecting malicious memory retention to downstream decisions and affected parties, MemHarmBench provides a systematic framework for exposing safety risks that conventional attack- and task-success metrics may overlook. We publicly release MemHarmBench at https://github.com/Kirito14IT/MemHarmBench.
Electroencephalography (EEG) provides a non-invasive means for neurological and mental health assessment, yet existing EEG models are typically developed for individual tasks and are difficult to unify across heterogeneous signal configurations and assessment objectives. Large language models (LLMs) offer a common interface for multiple tasks, but adapting heterogeneous EEG signals to a shared LLM raises a key challenge: how to preserve task-specific signal characteristics while enabling effective language-side parameter sharing. To address this problem, we propose EEG-HealthLLM, a unified and parameter-efficient framework for heterogeneous EEG health assessment. We introduce an Asymmetric Hierarchical Adaptation Strategy: task-specific EEG adapters preserve heterogeneous spatio-temporal dynamics, while a unified shared LoRA is injected into the LLM backbone (e.g., Phi-3) to capture shared high-level assessment semantics across tasks. We further establish EEG-HealthBench by reorganizing five existing EEG resources into paradigms covering epileptic seizure detection, major depressive disorder assessment, anxiety risk assessment, depressive symptom risk assessment, and emotional state assessment under a unified subject-independent protocol. Extensive experiments show that EEG-HealthLLM achieves an average balanced accuracy (BAC) of 69.58%, ranking first among the evaluated methods in both average BAC and macro-F1. Compared with independent single-task LoRA adaptation, it improves average BAC from 67.70% to 69.58% while reducing the aggregate language-side adapter footprint by 80.0%. Further analyses reveal task-dependent cross-task interactions and trade-offs, supporting the proposed hierarchical adaptation strategy. The implementation is publicly available at https://github.com/Marui-88/EEG-HealthLLM.
Huaiyuan Peng, Jisheng Dang, Wencan Zhang, Jingwen Zhao, Hong Peng, Wenjun Ma, Bin Hu, Qi Tian 等 9 位作者
Preference optimization for multimodal segmentation faces a granularity mismatch: dense masks contain spatially heterogeneous preference signals, whereas image-level objectives assign one shared nonlinear weight to all spatial margins. We introduce TopoDPO, a topology-aware dense preference alignment framework for structure-sensitive aerial referring segmentation. Topo-Prior mines informative preference pairs using connected-component discrepancy and macro-corner recall, while the dense objective applies the preference penalty to each local margin before spatial aggregation against a frozen supervised reference. On Topo-RS-Hard, a 4590-instance structural stress set, TopoDPO improves intersection over union (IoU) from 50.91% to 51.24% over image-level direct preference optimization (DPO) and reduces Betti-0 error from 1.34 to 0.90; paired image-cluster bootstrap intervals against image-level DPO exclude zero for all five reported metrics. Matched supervised controls and a complete factorial analysis separate the effects of topology-aware pair mining, localized correction, and dense reference-relative preference weighting. An unfiltered stratification over all 39,244 validation records further shows the clearest structural gains on samples with moderate or severe pre-alignment component-count errors. The code is publicly available at: https://github.com/phy709/TopoDPO.
Adrien Petralia, Alexandre Gavaudan, Artjom Joosen, Etienne Goffinet, Richard Kurle, Yassine El Ouahidi, Yassir Bendou, Abdelaziz Bounhar 等 356 位作者
We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.
Zhenye Yang, Jinpeng Chen, Wenbo Fu, Jun Ma, Huan Li, Qimin Zhou, Tao Wang, Li Sun 等 9 位作者
Multimodal Recommender Systems (MMRS) leverage rich item features to mitigate data sparsity, yet similarity-driven graph construction and propagation can favor semantically similar substitutes over functionally related complements. When behavioral associations connect semantically distant items, complementary signals can remain geometrically isolated, yielding fragmented "Similarity Islands" and semantic drift when standard propagation is applied. To overcome these limitations, we propose MAGE (Manifold-Augmented Graph Embedding), a framework for personalized multimodal ranking that decouples complementarity modeling into offline geometric warping and online topological propagation. The offline phase enhances item representations by extracting complement-aware residual directions via a Soft Conceptor mechanism and injecting them under a local gap-preserving constraint, incorporating co-purchase evidence while controlling the magnitude of geometric perturbation. Synergistically, the online phase employs an asymmetric dual-stream GNN that treats visual features as stable anchors and augmented textual features as flexible probes, fused through a self-calibrated cross-modal mechanism to reconcile shared semantics and modality-specific residual cues. Extensive experiments on four Amazon datasets show that MAGE yields consistent improvements over strong multimodal recommendation baselines. Together with the supporting analyses in the appendix, these results suggest that geometry-aware augmentation can alleviate the semantic mismatch that arises in complementary recommendation settings. Our code is available at https://github.com/Alex-bupt/mage.
On a chain of $N$ states, any local learning rule with bounded communication range needs $Θ(N)$ rounds before the loss can influence the earliest parameters, and this latency is usually attributed to depth. We show that it is a property of the communication geometry rather than of depth. For $N=m^d$ states placed at constant density in $d$ dimensions, a light-cone argument gives an $Ω(N^{1/d})$ lower bound on the number of bounded-range rounds that any local algorithm needs before its gradient estimates can be sign-correct at every parameter. We construct a unit-range, degree-$2d$ directed acyclic graph that contains a Hamiltonian computation of length $N-1$ and in which every state lies within directed distance $d(m-1)=Θ(N^{1/d})$ of the output, matching the lower bound. In the realizable-circuit model, depth and credit depth are therefore independent parameters: represented depth $Θ(N)$ and causal credit latency $Θ(N^{1/d})$ coexist, with a separation of order $N^{1-1/d}$. SpatialPC realizes this separation within predictive coding. Credit is propagated by adjoint prediction dynamics (APD), a local update whose $K$-round state on a settled acyclic graph is the truncated reverse-path sum $a^K=\sum_{k<K}(J^\top)^k b$, with error exactly $(J^\top)^K a^\star$. Graph geometry fixes when credit can first arrive at a state; the Jacobian-weighted path tail fixes how many further rounds accurate credit requires. A sparse two-dimensional Moore compiler retains the $N-1$ backbone with $\tfrac32N-2$ unit-range edges and has measured worst-node reachability $\sqrt N+2$ up to $N=65{,}536$, where a 65,535-edge backbone is reached in 258 local hops. With $T=2\sqrt N$ rounds, SpatialPC keeps finite-round task gradients aligned with settled gradients (cosine $0.93$ on MNIST and $0.97$ on Fashion-MNIST at $N=256$) and reaches $97.6\%$ and $88.1\%$ test accuracy after 20 epochs.
Recent advances in vision-language-action (VLA) and world-action models (WAMs) have significantly improved robotic manipulation, yet deploying robots in the open world demands a higher level of autonomy beyond executing predefined tasks. Robot agents offer a promising paradigm for this transition, integrating task discovery, contextual reasoning, physical execution, and adaptive recovery into a closed interaction loop. However, existing benchmarks largely evaluate isolated capabilities, failing to capture the complexity of open-world robot autonomy. We introduce RobotAgentBench (RAB), a comprehensive benchmark for evaluating robot agents in realistic, open-world household environments. Through an agent-assisted, human-in-the-loop real-to-sim pipeline, RAB constructs 50 diverse household layouts with interactive objects and physically grounded tasks, encompassing 90 atomic tasks, 18 short-horizon tasks, 8 long-horizon tasks, and 4 open-ended scenarios. RAB provides a multi-level evaluation of robot agent capabilities: (1) task-level assessments systematically examine five complementary dimensions: task discovery, context reasoning, manipulation, failure recovery, and robustness; (2) scenario-level evaluation assesses how agents integrate these capabilities to autonomously fulfill high-level user needs. Furthermore, RAB enables systematic comparisons of standalone VLA/WAM models and diverse robot agent architectures under a unified planning-execution-verification framework. Extensive experiments reveal substantial gaps between manipulation and autonomy, demonstrate the advantages of agentic architectures and recursive self-improvement, and highlight the remaining challenges toward open-world robot autonomy.
Jinming Wu ByteDance Inc., Zihao Deng, Wei Li Bytedance, Yiding Liu Tiktok, Bo You ByteDance Inc., Bo Li, Zejun MA TikTok, Ziwei Liu Nanyang Technological University
Robust deployment of large multimodal models (LMMs) in real-world scenarios requires access to external knowledge sources, given the complexity and dynamic nature of real-world information. Existing approaches such as retrieval-augmented generation (RAG) and prompt engineered search agents rely on rigid pipelines, often leading to inefficient or excessive search behaviors. We present MMSearch-R1, the first end-to-end reinforcement learning framework that enables LMMs to perform on-demand, multi-turn search in real-world Internet environments. Our framework integrates both image and text search tools, allowing the model to reason about when and how to invoke them guided by an outcome-based reward with a search penalty. To support training, We collect a multimodal search VQA dataset through a semi-automated pipeline that covers diverse visual and textual knowledge needs and curate a search-balanced subset with both search-required and search-free samples, which proves essential for shaping efficient and on-demand search behavior. Extensive experiments on knowledge-intensive and info-seeking VQA tasks show that our model not only outperforms traditional RAG-based baselines of the same model size, but also matches the performance of a larger RAG-based model while reducing search calls by over 30%. We further analyze key empirical findings to offer actionable insights for advancing research in multimodal search.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhenyun Yin, Shujie Wang, Xuhong Wang Shanghai AI Laboratory, Xingjun Ma Fudan University, Yingchun Wang Shanghai Artificial Intelligence Laboratory
Large language models with search capabilities frequently exhibit miscalibrated confidence, producing incorrect answers with high certainty. We present Deliberative Searcher, a reasoning-primary framework that integrates search operations into chain-of-thought generation while maintaining explicit confidence calibration. Our method employs constrained reinforcement learning with adaptive Lagrangian multipliers to jointly optimize correctness and reliability. Experiments across five benchmarks demonstrate substantial improvements: our 7B model reduces average false-certain rates from 54% in baselines to 2%, while our 72B variant achieves competitive accuracy with closed-source models and reduces false-certain rates to 9%. The well-calibrated confidence scores also enable more efficient test-time compute: instead of standard majority voting, we use confidence-weighted aggregation and match the performance of 16-sample majority voting with only 4 samples, a 4\times reduction in inference compute. These results establish calibrated confidence as a foundation for both trustworthy outputs and adaptive test-time compute, demonstrating the value of the proposed constrained RL framework in search-augmented language models.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Hanbing Liu, Lang Cao University of Illinois at Urbana-Champaign, Yuanyi Ren, Mengyu Zhou Alibaba Group, Haoyu Dong, Xiaojun Ma Microsoft, Shi Han Microsoft, Dongmei Zhang Microsoft
Large language models (LLMs) show strong reasoning abilities but often produce unnecessarily long explanations that reduce efficiency. Although reinforcement learning (RL) has been used to improve reasoning, most methods focus on accuracy and rely on uniform length-based rewards that overlook the differing contributions of individual tokens, often harming correctness. We revisit length optimization in RL through the perspective of token significance. Observing that many chain-of-thought (CoT) tokens contribute little to the final answer, we introduce a significance-aware length reward that selectively penalizes insignificance tokens, reducing redundancy while preserving essential reasoning. We also propose a dynamic length reward that encourages more detailed reasoning early in training and gradually shifts toward conciseness as learning progresses. Integrating these components into standard policy optimization yields a framework that improves both reasoning efficiency and accuracy. Experiments across multiple benchmarks demonstrate substantial reductions in response length while preserving or improving correctness, highlighting the importance of modeling token significance for efficient LLM reasoning.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表