论文检索

Ayan Banerjee, Shomrik Barua Banerjee, Sandeep Gupta

Human decision-making in safety-critical domains is governed by abstract policies that intentionally omit exhaustive preconditions/ triggers and contingencies. Executing such underspecified policies reliably in open-world settings remains a fundamental challenge for large language models (LLMs). We introduce NITI, Neural Bridging for Incremental Execution and Trigger Inference from Underspecified Human Policies, a neuro-symbolic framework that treats LLMs not as autonomous planners, but as execution-time concretizers of human intent. NITI incrementally executes abstract policies via verifier-grounded interfaces, infers implicit applicability conditions, repairs execution through neural bridging when assumptions fail, and halts safely under state inconsistency. We evaluate NITI on two structurally distinct embodied domains: a new benchmark of World Cubing Championship 2×2 Rubik’s Cube scrambles (n=50) and a safety-critical automated insulin dosing task. Across multiple frontier LLMs, NITI enables reliable long-horizon execution without task-specific training or search with minimal contextualization infence overhead, substantially outperforming one-shot and chain-of-thought baselines. Our results show that compositional, verifier-grounded execution is essential for safe human–AI collaboration in open-world decision-making. Code and benchmark available here - https://github.com/ImpactLabASU/ACLNITI

FU Yuqing, Yimin Deng, Wanyu Wang, Yuhao Wang, Yejing Wang, Hongshi Liu, Yiqi Wang, Xiao Han, Maolin Wang, Guoshuai Zhao 等

Multi-hop Question Answering (MHQA) aims to answer questions that require multi-step reasoning. The complexity of user queries, coupled with potential knowledge deficiencies in Large Language Models (LLMs), gives rise to two pivotal challenges that underpin the performance on this task: the correct identification of the reasoning path and the accurate retrieval of essential knowledge. Existing approaches primarily rely on prompt-based methods to generate reasoning paths, which are further combined with traditional sparse or dense retrieval to produce the final answer. However, the generation of reasoning paths commonly lacks effective control over the generative process, thus leading the reasoning astray. Meanwhile, the retrieval methods over-rely on knowledge matching or similarity scores rather than evaluating the practical utility of the information, resulting in retrieving homogeneous or non-useful information. Therefore, we propose a Structured Entity-Aware Retrieval with Chain-of-Reasoning Navigator framework named SEARCH-R. Specifically, SEARCH-R trains an end-to-end reasoning path navigator, which is able to provide a powerful sub-question decomposer by fine-tuning the Llama3.1-8B model. Moreover, a novel dependency tree-based retrieval is designed to evaluate the informational contribution of the document quantitatively. Extensive experiments on three challenging multi-hop datasets validate the effectiveness of the proposed framework. The code and dataset are available at: https://github.com/Applied-Machine-Learning-Lab/ACL2026_SEARCH-R.

Amit Agarwal, Hitesh Laxmichand Patel, Meizhu Liu, Jyotika Singh, Karan Dua, Hansa Meghwani, Matthew Rowe, M. Avendi, Yassi Abbasi, Tao Sheng 等

Reference-free image–to–text evaluators are now standard for scoring image–caption alignment, yet it is unclear whether they respect semantic invariances. We present an invariance probe on five popular evaluators (CLIPScore, PAC-S, UMIC, FLEUR, and a deterministic LLM judge) under semantics-preserving perturbations along three axes: spatial (flips, context-preserving repositioning, light rotations), object (scale, category), and socio-linguistic framing (cultural/economic adjectives with neutral and length-matched controls). Across curated slices of three detection datasets and three caption evaluation suites, we find consistent non-semantic sensitivities: benign spatial edits and simple phrasing changes shift scores by (≈)6–9% on average, and for systems separated by just 0.7% these shifts can cause ranking flips in upto (∼)37% of cases, particularly under spatial changes. A small human study also supports this finding and confirms that annotators generally judge perturbed pairs as equally correct, so these shifts reflect metric behavior rather than semantic change. We further propose invariance-calibrated scoring, a post-hoc adjustment that roughly halves median absolute sensitivity while retaining correlation with learned caption evaluators.

Hao Li, Yiqun Zhang, Zhaoyan Guo, Chenxu Wang, Shengji Tang, Qiaosheng Zhang, Yang Chen, Biqing Qi, Peng Ye, Lei Bai 等

Large language model (LLM) routing assigns each query to the most suitable model from an ensemble. We introduce LLMRouterBench, a large-scale benchmark and unified framework for LLM routing. It comprises over 400K instances from 21 datasets and 33 models. Moreover, it provides comprehensive metrics for both performance-oriented and performance-cost trade-off routing, and integrates 10 representative routing baselines. Using LLMRouterBench, we systematically re-evaluate the field. While confirming strong model complementarity—the central premise of LLM routing—we find that many routing methods exhibit similar performance under unified evaluation, and several recent approaches, including commercial routers, fail to reliably outperform a simple baseline. Meanwhile, a substantial gap remains to the Oracle, driven primarily by persistent model-recall failures. We further show that backbone embedding models have limited impact, that larger ensembles exhibit diminishing returns compared to careful model curation, and that the benchmark also enables latency-aware analysis. All code and data are available at https://github.com/ynulihao/LLMRouterBench.

Juyong Jiang, Jiasi Shen, Sunghun Kim, Kang Min Yoo, Jeonghoon Kim, Sungju Kim

While Large Language Models (LLMs) have revolutionized code generation, standard “System 1” approaches that generate solutions in a single forward pass often hit a performance ceiling on complex algorithmic tasks. Existing iterative refinement strategies attempt to bridge this gap at inference time, yet they predominantly rely on external oracles, execution feedback, or computationally expensive prompt-response cycles. In this work, we propose ReflexiCoder, a novel reinforcement learning (RL) framework that internalizes the structured reasoning trajectory, encompassing initial generation, bug and optimization aware reflection, and self-correction, directly into the model’s weights. Unlike prior methods, ReflexiCoder shifts the paradigm from external-dependent refinement to an intrinsic, fully autonomous self-reflection and self-correction capabilities at inference time. We utilize an RL-only training paradigm with granular reward functions to optimize the entire reflection-correction trajectory, teaching the model how to debug without reliance on ground-truth feedback or execution engines at inference time. Extensive experiments across seven benchmarks demonstrate that our ReflexiCoder-8B establishes a new state-of-the-art (SOTA) among leading open-source models in the 1.5B-14B range, achieving 94.51% (87.20%) on HumanEval (Plus), 81.80% (78.57%) on MBPP (Plus), 35.00% on BigCodeBench, 52.21% on LiveCodeBench, and 37.34% on CodeForces in a single-attempt setting, rivaling or surpassing proprietary models like GPT-5.1. Notably, our framework is significantly more token-efficient than base models, reducing inference-time compute overhead by approximately 40% through disciplined, efficient reasoning and reflection patterns. The source code and data are available at https://github.com/juyongjiang/ReflexiCoder.

Karthik Singaravadivelan, Anant Gupta, Zekun Wang, Christopher J. MacLellan

Topic modeling seeks to uncover latent semantic structure in text corpora with minimal supervision. Neural approaches achieve strong performance but require extensive tuning and struggle with lifelong learning due to catastrophic forgetting and fixed capacity, while classical probabilistic models lack flexibility and adaptability to streaming data. We introduce CobwebTM, a low-parameter lifelong hierarchical topic model based on incremental probabilistic concept formation. By adapting the Cobweb algorithm to continuous document embeddings, CobwebTM constructs semantic hierarchies online, enabling unsupervised topic discovery, dynamic topic creation, and hierarchical organization without predefining the number of topics. Across diverse datasets, CobwebTM achieves strong topic coherence, stable topics over time, and high-quality hierarchies, demonstrating that incremental symbolic concept formation combined with pretrained representations is an efficient approach to topic modeling.

Tejal Nair, Mahmud Wasif Nafee, Maiqi Jiang, Ashley Gao, Haipeng Chen, Yanfu Zhang

Although large language models (LLMs) excel at factual recall, they can still propagate stale or incorrect knowledge, making in-context knowledge editing a gradient-free remedy suitable for black-box APIs. These knowledge editors that use in-context learning typically rely on a single retriever and surface-similarity heuristics to build prompts. However, a key observation in this study is that retrievers can be complementary: semantic rankers may recover paraphrased evidence, while lexical or feature-based retrievers may preserve precise entities and cues. This creates two gaps in single-retriever editors: they (i) miss complementary evidence that different retrievers surface and (ii) cannot adapt when one retriever is clearly more reliable for a query. We introduce a Feature-Weighted Ensemble for In-context Knowledge Editing (FWE-IKE) that calibrates three heterogeneous rankers (LLM-, BERT-, and MLP-based), extracts simple confidence features from each ranker, predicts per-query mixture weights, and applies a conservative margin-based routing gate that selects a single expert when confident; otherwise we mix calibrated distributions with learned per-query weights. On the CounterFact benchmark, FWE-IKE attains 88.33% Edit-Success Rate, a +3.0 point gain over the best single retriever and approaching the oracle upper bound (91%). Case studies, an ablation study, and analyses show the method systematically recovers complementary wins (e.g., BERT-only, LLM-only, MLP-only slices). FWE-IKE improves edit accuracy without touching model weights and provides a practical path to more robust, confidence-aware retrieval for IKE.

Louis Estève, Christophe Servan, Thomas Lavergne, Agata Savary

Diversity has been gaining interest in the NLP community in recent years. At the same time, state-of-the-art transformer models such as ModernBERT use very large pre-training datasets, which are driven by size rather than by diversity. This summons to investigate theimpact of diversity on pre-training. We do so in this study, with the express intent of reducing pre-training dataset size, while retaining atleast comparable performance. We compare diversity-driven sampling algorithms, and we use the best one to pre-train several ModernBERT models on French with a fixed compute budget. We fine-tune and evaluate them on a variety of French benchmarks. We compare them with models pre-trained on randomly sampled data of commensurate size, with the same compute budget. We find that both random and diversity-driven sampling may reduce the pre-training dataset by up to 94% and the pre-training time by up to 73% while maintaining performance. Moreover, in some tasks, the inherent quality of models, estimated via head-only fine-tuning, is up to 10 points higher with diversity sampling than with random sampling.

Hongzhi Zhang, Yuanze Hu, Tinghai Zhang, Jia Fu, Tao Wang, Junwei Jing, Zhaoxin Fan, Wei Bi, Ruiming Tang, Han Li 等

The evolution of Large Language Models (LLMs) towards autonomous agents has catalyzed progress in Deep Research. While retrieval capabilities are well-benchmarked, the post-retrieval synthesis stage—where agents must digest massive amounts of context and consolidate fragmented evidence into coherent, long-form reports—remains under-evaluated due to the subjectivity of open-ended writing.To bridge this gap, we introduce DeepSynth-Eval, a benchmark designed to objectively evaluate information consolidation capabilities. We leverage high-quality survey papers as gold standards, reverse-engineer research requests, and construct Oracle Contexts from their bibliographies to isolate synthesis from retrieval noise. We propose a fine-grained evaluation protocol using General Checklists (for factual coverage) and Constraint Checklists (for structural organization), transforming subjective judgment into verifiable metrics. Experiments across 96 tasks reveal that synthesizing information from hundreds of references remains a significant challenge. Our results demonstrate that agentic "plan-then-write" workflows significantly outperform single-turn generation, effectively reducing hallucinations and improving adherence to complex structural constraints.

Sarmistha Das, Vaibhav Vishal, Shreyas Guha, Amaan Ali, Kitsuchart Pasupa, Sriparna Saha

In the contemporary epoch of multilingual education, learning idioms provides a fascinating gateway towards creativity, cultural values, historical context, and diverse perspectives inherent to various linguistic traditions. This paper showcases the navigation of retaining figurative and cultural semantics in low-resource Southeast Asian languages such as Hindi, Bengali, and Thai, where culturally rich idioms pose significant obstacles for computational modelling and cross-linguistic transfer due to their deep metaphorical complexity. To tackle such complexity, we present Varnika (वर्णिका) , a reconstructed multimodal idiom corpus comprising 3,533 multilingual idioms, enriched with seven idiomatic tones aligned with both textual and visual representations. Additionally, to infer informative idiomatic understanding, we introduce a Hybrid Mixture-of-Experts (HybridMoE) framework that embeds multiple idiomatic expert opinions while mitigating expert sparsity by integrating outputs from both selected and unselected experts through controlled hybridisation, further augmented with Idiomatic Property Signals via masked multimodal embeddings. To analyse the performance across multiple dimensions, we propose the IDIO-TONE and Idiomatic Validation Score, a three-stage evaluation pipeline measuring (i) literal translation fidelity, (ii) visual- semantic alignment, and (iii) idiomatic meaning retention. Empirical evaluations highlight that HybridMoE achieves 5–6% performance gains across advanced vision language models, demonstrating improved representation of figurative language and culturally embedded meaning in multilingual multimodal settings. Resources are available at (https://github.com/sarmistha-D/Hybrid_MOE).

Weiwei Qi, Zefeng Wu, Tianhang Zheng, Zikang Zhang, Xiaojun Jia, Zhan Qin, Kui Ren

Ensuring Large Language Model (LLM) safety is crucial, yet the lack of a clear understanding about safety mechanisms hinders the development of precise and reliable methodologies for safety intervention across diverse tasks. To better understand and control LLM safety, we propose the Expected Safety Impact (ESI) framework for quantifying how different parameters affect LLM safety. Based on ESI, we reveal distinct safety-critical patterns across different LLM architectures: In dense LLMs, many safety-critical parameters are located in value matrices (V) and MLPs in middle layers, whereas in Mixture-of-Experts (MoE) models, they shift to late-layer MLPs. Leveraging ESI, we further introduce two targeted intervention paradigms for safety enhancement and preservation, i.e., Safety Enhancement Tuning (SET) and Safety Preserving Adaptation (SPA). SET can align unsafe LLMs by updating only a few safety-critical parameters, effectively enhancing safety while preserving original performance. SPA safeguards well-aligned LLMs during capability-oriented intervention (e.g., instruction tuning) by preventing disruption of safety-critical weights, allowing the LLM to acquire new abilities while maintaining safety capabilities. Extensive evaluations on different LLMs demonstrate that SET can reduce the attack success rates of unaligned LLMs by over 50% with only a 100-iteration update on 1% of model weights. SPA can limit the safety degradation of aligned LLMs within 1% after a 1,000-iteration instruction fine-tuning on different tasks. Our code is available at: https://github.com/ZJU-LLM-Safety/SafeWeights-ACL

Qirui Liu, Hao Chen, Weijie Shi, Jiajie Xu, Jia Zhu

Accurately identifying student misconceptions is crucial for personalized education but faces three challenges: (1) data scarcity with long-tail distribution, where authentic student reasoning is difficult to synthesize; (2) fuzzy boundaries between error categories with high annotation noise; (3) deployment paradox—large models overlook unconventional approaches due to pretraining bias and cannot be deployed on edge, while small models overfit to noise. Unlike traditional methods that increase diversity through large-scale data synthesis, we propose a two-stage knowledge distillation framework that mines high-value samples from existing data. The first stage performs standard distillation to transfer task capabilities. The second stage introduces a dual-layer marginal selection mechanism based on cognitive uncertainty, identifying four types of critical samples based on teacher model uncertainty and confidence differences. For different data subsets, we design difficulty-adaptive mechanism to balance hard/soft label contributions, enabling student models to inherit inter-class relationships from teacher soft labels while distinguishing ambiguous error types. Experiments show that with augmented training on only 10.30% of filtered samples, we achieve MAP@3 of 0.9585 (+17.8%) on the MAP-Charting dataset, and using only a 4B parameter model, we attain 84.38% accuracy on cross-topic tests of middle school algebra misconception benchmarks, significantly outperforming sota LLM (67.73%) and standard fine-tuned 72B models (81.25%). Our code is available at https://anonymous.4open.science/r/acl2026_map-5847/.

Jianming Chen, Yawen Wang, Junjie Wang, Xiaofei Xie, Shoubin Li, Qing Wang, Fanjiang Xu

Embodied agents in safety-critical applications such as Vision-Language Navigation (VLN) rely on multiple interdependent capabilities (e.g., perception, memory, planning, decision), making failures difficult to localize and attribute. Existing testing methods are largely system-level and provide limited insight into which capability deficiencies cause task failures. We propose a capability-oriented testing approach that enables failure detection and attribution by combining (1) adaptive test case generation via seed selection and mutation, (2) capability oracles for identifying capability-specific errors, and (3) a feedback mechanism that attributes failures to capabilities and guides further test generation. Experiments show that our method discovers more failure cases and more accurately pinpoints capability-level deficiencies than state-of-the-art baselines, providing more interpretable and actionable guidance for improving embodied agents.

Jyotika Singh, Fang Tu, Miguel Ballesteros, Weiyi Sun, Sandip Ghoshal, Michelle Yuan, Yassine Benajiba, Sujith Ravi, Dan Roth

Large language models (LLMs) suffer significant performance degradation when user instructions and context are distributed over multiple conversational turns, yet multi-turn (MT) interactions dominate chat interfaces. The routine approach of appending full chat history to prompts rapidly exhausts context windows, leading to increased latency, higher computational costs, and diminishing returns as conversations extend. We introduce **MT-OSC**, a **O**ne-off **S**equential **C**ondensation framework that efficiently and automatically condenses chat history in the background without disrupting the user experience. MT-OSC employs a Condenser Agent that uses a few-shot inference-based Condenser and a lightweight Decider to selectively retain essential information, reducing token counts by up to 72% in 10-turn dialogues. Evaluated across 13 state-of-the-art LLMs and diverse multi-turn benchmarks, MT-OSC consistently narrows the multi-turn performance gap—yielding improved or preserved accuracy across datasets while remaining robust to distractors and irrelevant turns. Our results establish MT-OSC as a scalable solution for multi-turn chats, enabling richer context within constrained input spaces, reducing latency and operational cost, while balancing performance.

Weizhi Zhang, Xinyang Zhang, Chenwei Zhang, Liangwei Yang, Jingbo Shang, Zhepei Wei, Henry Peng Zou, Zijie Huang, Zhengyang Wang, Yifan Gao 等

Large Language Model (LLM) empowered agents have recently emerged as advanced paradigms that exhibit impressive capabilities in a wide range of domains and tasks. Despite their potential, current LLM agents often adopt a one-size-fits-all approach, lacking the flexibility to respond to users’ varying needs and preferences. This limitation motivates us to develop PersonaAgent, the first personalized LLM agent framework designed to address versatile personalization tasks. Specifically, PersonaAgent integrates two complementary components: a personalized memory module that includes episodic and semantic memory mechanisms; a personalized action module that enables the agent to perform tool actions tailored to the user. At the core, the persona (defined as unique system prompt for each user) functions as an intermediary: it leverages insights from personalized memory to control agent actions, while the outcomes of these actions in turn refine the memory. Based on the framework, we propose a test-time user-preference alignment strategy that simulate the latest n interactions to optimize the persona prompt, ensuring real-time user preference alignment through textual loss feedback between simulated and ground-truth responses. Experimental evaluations demonstrate that PersonaAgent significantly outperforms other baseline methods by not only personalizing the action space effectively but also scaling during test-time real-world applications. These results underscore the feasibility and potential of our approach in delivering tailored, dynamic user experiences.

Bashar Talafha, Amin Abu Alhassan, Muhammad Abdul-Mageed

Zero-shot ASR for Arabic remains challenging: while multilingual models perform well on Modern Standard Arabic (MSA), error rates rise sharply on dialectal and accented speech due to linguistic mismatch and scarce labeled data. We study \textit{context-aware} decoding as a lightweight test-time adaptation paradigm that conditions inference on external side information without parameter updates. For promptable encoder–decoder ASR (e.g., Whisper), we incorporate context through (i) decoder prompting with first-pass hypotheses and (ii) encoder/decoder prefixing with retrieved speech-text exemplars, complemented by simple prompt reordering and optional speaker-matched synthetic exemplars to improve robustness in informal and multi-speaker settings. To extend contextual adaptation beyond promptable architectures, we introduce \textit{proxy-guided $n$-best selection} for CTC ASR: given one or more external proxy hypotheses, we select from a model’s n-best list by minimizing text-level distance to the proxies, enabling contextual inference without direct prompting. Across ten Arabic conditions spanning MSA, accented MSA, and multiple dialects, the best-performing context-aware variants yield average relative WER reductions of 22.29% on MSA, 20.54% on accented MSA, and 9.15% on dialectal Arabic. For CTC ASR on our Common Voice MSA testbed, proxy-guided selection reduces WER by 15.6% relative and recovers a substantial fraction of oracle n-best gains, showing that external-context guidance can also benefit non-promptable ASR.

Muhammad Arslan Manzoor, Dilshod Azizov, Daniil Orel, Umer Siddique, Zain Muhammad Mujahid, Yufang Hou, Preslav Nakov

News outlets shape public opinion on a scale, which makes automated detection of political bias and factuality essential. Yet, the field still lacks unified resources, comprehensive evaluations in diverse approaches, and systematic analyzes of the representations and fusion strategies that matter the most, especially under label sparsity and dataset diversity. In addition, there is little empirical work that reports broad observation driven findings about what consistently works, what fails, and why. We address these gaps with four contributions: (i) MBFC-2025, a large-scale label set that covers ~2,600 outlets from Media Bias/Fact Check (MBFC); (ii) multi-view representations for ACL-2020 ~900 outlets and MBFC-2025, spanning Alexa graphs, hyperlink graphs, LLM-derived graphs, articles, and Wikipedia descriptions; (iii) systematic evaluation and analysis of embedding views and fusion strategies, including an RL-based fusion variant; and (iv) extensive experiments that achieve state-of-the-art results on ACL-2020 and establish strong benchmarks on MBFC-2025.

Yibo Yan, Shen Wang, Jiahao Huo, Jingheng Ye, Zhendong Chu, Xuming Hu, Philip S. Yu, Carla P Gomes, Bart Selman, Qingsong Wen

Scientific reasoning, the process through which humans apply logic, evidence, and critical thinking to explore and interpret scientific phenomena, is essential in advancing knowledge reasoning across diverse fields. However, despite significant progress, current scientific reasoning models still struggle with generalization across domains and often fall short of multimodal perception. Multimodal Large Language Models (MLLMs), which integrate text, images, and other modalities, present an exciting opportunity to overcome these limitations and enhance scientific reasoning. Therefore, **this position paper argues that MLLMs can significantly advance scientific reasoning across disciplines such as mathematics, physics, chemistry, and biology**. We highlight the current state of MLLM applications in scientific reasoning, noting their ability to integrate and reason over diverse data types. However, challenges such as multimodal alignment, data diversity, and reasoning depth remain obstacles to achieving their full potential. To address these challenges, we propose actionable suggestions in the near future. Overall, our work offers a novel perspective on MLLM integration with scientific reasoning, providing the LLM community with valuable insights for achieving Artificial General Intelligence (AGI).

Yang Lyu, Jin Cao, Yang Xiao, Zhe Sun, Ben Niu, Fenghua Li, Hui LI

Distributed LLM inference avoids sending raw inputs by transmitting intermediate hidden states, a practice widely assumed to preserve privacy. We challenge this assumption and demonstrate that intermediate representations alone are sufficient to leak sensitive user attributes. This setting poses a fundamental obstacle for existing attribute inference attacks, which typically rely on auxiliary embedding-attribute pairs. To characterize this previously underexplored privacy risk, we reformulate attribute inference as zero-shot matching over candidate attributes directly in the intermediate representation space, and introduce a purely intermediate-representation-based attribute inference attack, termed IR-AIA. To address two structural challenges that hinder attribute inference from intermediate representations, we propose SG-APCR to address layer-dependent anisotropy in intermediate embeddings and a sliding-window similarity matching strategy to handle subword-level semantic fragmentation. Experiments across three LLMs and three real-world datasets show that sensitive attributes can be reliably inferred using only intermediate representations, achieving Top-1 accuracy of up to 0.997 on CMS, 0.980 on Skytrax, and 0.986 on ECHR. These results reveal that intermediate states commonly considered safe to share can expose sensitive personal attributes on their own.

Sangjun Song, Minjae Oh, Seungkyu Lee, Sungmin Jo, Yohan Jo

Large Reasoning Models (LRMs) allocate substantial inference-time compute to Chain-of-Thought (CoT) reasoning, improving performance on mathematics, scientific QA, and tool usage. However, this introduces overthinking: LRMs often reach a correct intermediate solution, continue reasoning, and overwrite it with an incorrect answer. We first demonstrate that oracle stopping—where we inject lt;/think gt; at every sentence boundary and select the best stopping point in hindsight—improves average accuracy by 8% while reducing thinking tokens by 72%, exposing substantial overthinking. Motivated by this finding, we propose ThinkBrake, which monitors the log-probability margin between the top continuation token and lt;/think gt; at sentence boundaries, stopping reasoning when this margin narrows. ThinkBrake requires no training and achieves favorable accuracy–efficiency trade-offs across math, scientific QA, and tool usage benchmarks, reducing thinking token usage by up to 30%. Furthermore, we provide theoretical analysis showing that ThinkBrake is equivalent to test-time realignment with a reward bonus for the lt;/think gt; token.