ICLR 2026 Papers

Large Language Models (LLMs) increasingly rely on Chain-of-Thought (CoT) reasoning to solve complex problems. Contrary to the common belief that longer CoTs always improve performance, we demonstrate that **longer is not always better**. Across both real-world LLMs and theoretical models, task accuracy follows an inverted U-shaped curve with respect to CoT length: performance rises initially but declines once reasoning chains become too long. Through controlled experiments, we uncover **scaling behaviors of the optimal CoT length**: it increases with task difficulty but decreases with model capability. This exposes a significant mismatch with current practice, where supervised training often reuses the same CoT data across models and tasks without adaptivity. We further show that Reinforcement Learning (RL) can mitigate this gap by dynamically calibrating CoT length, thereby improving accuracy and offering a new perspective on differences between supervised fine-tuning and RL training. To explain these phenomena, we introduce an error-accumulation analysis that characterizes how reasoning errors propagate across steps and derives the scaling behaviors of CoT length observed empirically. Building on these insights, we show that training with optimally sized CoTs and applying length-aware filtering during inference yields substantial improvements in performance. Taken together, these findings establish a principled explanation of the ''overthinking'' effect and yield practical guidelines for calibrating CoT length in accordance with task complexity and model capability.

The reasoning large language model (RLLM) has been proven competitive in solving complex reasoning tasks such as mathematics, coding, compared to traditional LLM. However, the serving performance and behavior of RLLM remains \textit{unexplored}, which may undermine the deployment and utilization of RLLM in real-world scenario. To close this gap, in this paper, we conduct a comprehensive study of RLLM service. We first perform a pilot study on comparing the serving performance between RLLM and traditional LLM and reveal that there are several distinct differences regarding serving behavior: (1) \textit{significant memory usage and fluctuations}; (2) \textit{straggler requests}; (3) \textit{adaptive running time}; (4) \textit{domain preference}. Then we further investigate whether existing inference optimization techniques are valid for RLLM. Our main takeaways are that model quantization methods and speculative decoding can improve service system efficiency with small compromise to RLLM accuracy, while prefix caching, KV cache quantization may even degrade accuracy or serving performance for small RLLM. Lastly, we conduct evaluation under real world workload modeled by Gamma distribution to verify our findings. Empirical results for real world workload evaluation across different dataset are \textit{aligned} with our main findings regarding RLLM serving. We hope our work can provide the research community and industry with insights to advance RLLM inference serving.

Vision-language models (VLMs) have demonstrated strong performance in image geolocation, a capability further sharpened by frontier multimodal large reasoning models (MLRMs). This poses a significant privacy risk, as these widely accessible models can be exploited to infer sensitive locations from casually shared photos, often at street-level precision, potentially surpassing the level of detail the sharer consented or intended to disclose. While recent work has proposed applying a blanket restriction on geolocation disclosure to combat this risk, these measures fail to distinguish valid geolocation uses from malicious behavior. Instead, VLMs should maintain contextual integrity by reasoning about elements within an image to determine the appropriate level of information disclosure, balancing privacy and utility. To evaluate how well models respect contextual integrity, we introduce VLM-GEOPRIVACY, a benchmark that challenges VLMs to interpret latent social norms and contextual cues in real-world images and determine the appropriate level of location disclosure. Our evaluation of 14 leading VLMs shows that, despite their ability to precisely geolocate images, the models are poorly aligned with human privacy expectations. They often over-disclose in sensitive contexts and are vulnerable to prompt-based attacks. Our results call for new design principles in multimodal systems to incorporate context-conditioned privacy reasoning.

Recent reasoning Large Language Models (LLMs) demonstrate remarkable problem-solving abilities but often generate long thinking traces whose utility is unclear. We conduct a systematic analysis across models and datasets and discover a U-shaped entropy pattern: high entropy on simple problems despite high accuracy, low entropy on medium difficulty, and high entropy on hard problems reflecting uncertainty. The 22--25\% entropy reduction from simple to optimal regions reveals a fundamental inefficiency—an \emph{overthinking} phenomenon on easy instances. Building on these insights, we introduce \textbf{DiffAdapt}, a lightweight, deployment-ready framework that predicts problem difficulty from hidden states and selects among Easy/Normal/Hard reasoning strategies to allocate computation adaptively. DiffAdapt requires no retraining of the base LLM and is compatible with common inference optimizations. Across five models and eight benchmarks, DiffAdapt achieves comparable or improved accuracy while reducing token usage by up to 22.4\%, establishing a practical path toward compute-efficient reasoning.

Reinforcement Learning with verifiable rewards (RLVR) has emerged as a primary learning paradigm for enhancing the reasoning capabilities of multi-modal large language models (MLLMs). However, during RL training, the enormous state space of MLLM and sparse rewards often leads to entropy collapse, policy degradation, or over-exploitation of suboptimal behaviors. This necessitates an exploration strategy that maintains productive stochasticity while avoiding the drawbacks of uncontrolled random sampling, yielding inefficient exploration. In this paper, we propose CalibRL, a hybrid-policy RLVR framework that supports controllable exploration with expert guidance, enabled by two key mechanisms. First, a distribution-aware advantage weighting scales updates by group rareness to calibrate the distribution, therefore preserving exploration. Meanwhile, the asymmetric activation function (LeakyReLU) leverages the expert knowledge as a calibration baseline to moderate overconfident updates while preserving their corrective direction. CalibRL increases policy entropy in a guided manner and clarifies the target distribution by estimating the on-policy distribution through online sampling. Updates are driven by these informative behaviors, avoiding convergence to erroneous patterns. Importantly, these designs help alleviate the distributional mismatch between the model’s policy and expert trajectories, thereby achieving a more stable balance between exploration and exploitation. Extensive experiments across eight benchmarks, including both in-domain and out-of-domain settings, demonstrate consistent improvements, validating the effectiveness of our controllable hybrid-policy RLVR training. Code is available at https://github.com/zhh6425/CalibRL.

Recent advances in applying reinforcement learning (RL) to large language models (LLMs) have led to substantial progress. In particular, a series of remarkable yet often counterintuitive phenomena have been reported in LLMs, exhibiting patterns not typically observed in traditional RL settings. For example, notable claims include that a single training example can match the performance achieved with an entire dataset, that the reward signal does not need to be very accurate, and that training solely with negative samples can match or even surpass sophisticated reward-based methods. However, the precise conditions under which these observations hold—and, critically, when they fail—remain unclear. In this work, we identify a key factor that differentiates RL observations: whether the pretrained model already exhibits strong *Model-Task Alignment*, as measured by pass@k accuracy on the evaluated task. Through a systematic and comprehensive examination of a series of counterintuitive claims, supported by rigorous experimental validation across different model architectures and task domains, our findings show that while standard RL training remains consistently robust across settings, many of these counterintuitive results arise only when the model and task already exhibit strong model-task alignment. In contrast, these techniques fail to drive substantial learning in more challenging regimes, where standard RL methods remain effective.

Puzzlehunts are a genre of complex, multi-step puzzles lacking well-defined problem definitions. In contrast to conventional reasoning benchmarks consisting of tasks with clear instructions and constrained environments, puzzlehunts requires discovering the underlying problem structure from multimodal evidence and iterative reasoning, mirroring real-world domains such as scientific discovery, exploratory data analysis, or investigative problem-solving. Despite progress in foundation models, their performance on open-ended settings remains largely untested. We introduce PuzzleWorld, a comprehensive benchmark of 667 puzzlehunt-style problems designed to assess step-by-step, open-ended, and creative multimodal reasoning. Each puzzle is annotated with the final solution, detailed reasoning traces, and cognitive skill labels, enabling holistic benchmarking and fine-grained diagnostic analysis. Most state-of-the-art models achieve only 1-4\% final answer accuracy. On PuzzleWorld, the best model solves only 14\% of puzzles and reaches 40\% stepwise accuracy, matching human puzzle novices but falling significantly behind puzzle enthusiasts. To demonstrate the value of our reasoning annotations, we show that fine-tuning a small model on reasoning traces boosts stepwise accuracy from 4\% to 11\%, which translates to improvements in downstream visual reasoning tasks. Our detailed error analysis reveals that current models exhibit myopic reasoning, are bottlenecked by the limitations of language-based inference, and lack sketching capabilities crucial for visual and spatial reasoning. We will publicly release PuzzleWorld to support future work on building more general, open-ended, and creative reasoning systems.

Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful paradigm for enhancing the reasoning ability of Large Language Models (LLMs). Yet current RLVR methods often explore poorly, leading to premature convergence and entropy collapse. Moreover, they tend to produce poorly calibrated policies that remain confident in their generations regardless of correctness. To address this challenge, we introduce **Curiosity-Driven Exploration (CDE)**, a framework that leverages the model's intrinsic sense of curiosity to guide exploration. We formalize curiosity with signals from both the actor and the critic: for the actor, we use perplexity over its generated response, and for the critic, we use the variance of value estimates from a multi-head critic architecture. Both signals serve as an exploration bonus within the RLVR framework to guide the model. Our theoretical analysis shows that the actor-wise bonus inherently penalizes overconfident errors and promotes diversity among correct responses; moreover, we connect the critic-wise bonus to the well-established count-based exploration bonus in RL. Empirically, our method achieves an approximate **+3** point improvement over standard RLVR using GRPO/PPO on AIME benchmarks.

The ultimate goal of expressive speech-to-speech translation (S2ST) is to accurately translate spoken content while preserving the speaker identity and emotional style. However, progress in this field is largely hindered by three key challenges: the scarcity of paired speech data that retains expressive styles, the complexity of multi-stage processing pipelines, and the limited transfer of translation capabilities from large language models (LLMs). In this work, we address these challenges by introducing UniSS, a novel single-stage framework for expressive S2ST. Our approach features carefully designed speech semantic and style modeling, enabling seamless integration with existing text-based LLM frameworks to develop a unified text-speech language model. To transfer translation capabilities from text to speech, we propose a cross-modal chain-of-thought prompting process that progressively aligns audio semantics with text and ensures style preservation in the decoded results. Furthermore, we construct and release a large-scale, high-quality expressive S2ST dataset, UniST, comprising 44.8k hours of data. Experimental results show that UniSS significantly outperforms previous methods in translation fidelity and speech quality while preserving voice, emotion, and duration consistency. Our work establishes a simpler and more effective paradigm for building the next generation of expressive S2ST systems. Audio samples are available at https://uniss-s2st.github.io/.

Chain-of-Thought (CoT) reasoning has proven effective in enhancing Large Language Models (LLMs) on complex tasks by decomposing problems into step-wise solutions. However, extending CoT to multi-modal settings remains challenging, as it requires modeling transitions of visual states alongside textual reasoning. Existing approaches often underperform due to limited capacity to model visual transitions or fragmented architectures. To overcome this limitation, we introduce Uni-CoT, a Unified Chain-of-Thought framework that captures structured visual transitions and seamlessly aligns them with textual logic, enabling coherent multimodal reasoning. To mitigate the computational and training challenges inherent to multi-modal reasoning, Uni-CoT introduces a two-level reasoning paradigm: a macro-level CoT for high-level planning and a micro-level CoT for localized subtask execution. This hierarchical design reduces computational overhead while maintaining coherence. Additionally, Uni-CoT incorporates a structured training paradigm with auxiliary tasks to stabilize optimization and improve generalization. Experiments on reasoning-driven image generation and understanding benchmarks demonstrate that Uni-CoT achieves state-of-the-art performance and remarkable generalization, underscoring its potential for complex multi-modal reasoning.

Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive models for faster inference via parallel token generation. We provide a rigorous foundation for this advantage by formalizing a model of parallel sampling and showing that DLMs augmented with polynomial-length chain-of-thought (CoT) can simulate any parallel sampling algorithm using an optimal number of sequential steps. Consequently, whenever a target distribution can be generated using a small number of sequential steps, a DLM can be used to generate the distribution using the same number of optimal sequential steps. However, without the ability to modify previously revealed tokens, DLMs with CoT can still incur large intermediate footprints. We prove that enabling remasking (converting unmasked tokens to masks or revision (converting unmasked tokens to other unmasked tokens) together with CoT further allows DLMs to simulate any parallel sampling algorithm with optimal space complexity. We further justify the advantage of revision by establishing a strict expressivity gap: DLMs with revision or remasking are strictly more powerful than those without. Our results not only provide a theoretical justification for the promise of DLMs as the most efficient sampler, but also advocate for why revisions should be enabled in DLMs.

Spoken Language Models (SLMs) are designed to take speech inputs and produce spoken responses. However, current SLMs lack the ability to perform an internal, unspoken thinking process before responding. In contrast, humans typically engage in complex mental reasoning internally, enabling them to communicate ideas clearly and concisely. Thus, integrating an unspoken thought process into SLMs is highly desirable. While naively generating a complete chain-of-thought (CoT) reasoning before starting to talk can enable thinking for SLMs, this induces additional latency for the speech response, as the CoT reasoning can be arbitrarily long. To solve this issue, we propose STITCH, a novel generation method that alternates between the generation of unspoken reasoning chunks and spoken response chunks. Since the audio duration of a chunk of spoken response is much longer than the time to generate the tokens in a chunk of spoken response, we use the remaining free time to generate the unspoken reasoning tokens. When a chunk of audio is played to the user, the model continues to generate the next unspoken reasoning chunk, achieving simultaneous thinking and talking. Remarkably, STITCH matches the latency of baselines that cannot generate unspoken CoT by design while outperforming those baselines by 15% on math reasoning datasets; STITCH also performs equally well on non-reasoning datasets as those baseline models. Some animations and demonstrations are on the project page: https://d223302.github.io/STITCH.

Deepfake detection remains a formidable challenge due to the evolving nature of fake content in real-world scenarios. However, existing benchmarks suffer from severe discrepancies from industrial practice, typically featuring homogeneous training sources and low-quality testing images, which hinder the practical usage of current detectors. To mitigate this gap, we introduce **HydraFake**, a dataset that contains diversified deepfake techniques and in-the-wild forgeries, along with rigorous training and evaluation protocol, covering unseen model architectures, emerging forgery techniques and novel data domains. Building on this resource, we propose **Veritas**, a multi-modal large language model (MLLM) based deepfake detector. Different from vanilla chain-of-thought (CoT), we introduce *pattern-aware reasoning* that involves critical patterns such as "planning" and "self-reflection" to emulate human forensic process. We further propose a two-stage training pipeline to seamlessly internalize such deepfake reasoning capacities into current MLLMs. Experiments on HydraFake dataset reveal that although previous detectors show great generalization on cross-model scenarios, they fall short on unseen forgeries and data domains. Our Veritas achieves significant gains across different out-of-domain (OOD) scenarios, and is capable of delivering transparent and faithful detection outputs.

Recent advances in 3D Large Language Models (3D-LLMs) show strong potential in understanding and interacting with 3D environments, yet their training data typically lack explicit reasoning processes, limiting complex spatial reasoning and task planning. To address this, we annotate SCoT, a million-scale Chain-of-Thought dataset spanning three levels: a) Spatial Perception (what is there), recognizing object properties, relations, and scene attributes; b) Spatial Analysis (what does it mean), inferring rationality, functionalities, and physical implications; c) Spatial Planning (what should I do), integrating perception and reasoning for actionable strategies. Unlike prior datasets supervising only answers, SCoT annotates intermediate reasoning grounded in scene cues, specifically for analysis and planning tasks. Results show that CoT supervision greatly benefits complex analysis and planning but induces hallucinations and accuracy drops in simple perception. These findings highlight both the necessity and the nuanced challenges of scene-grounded reasoning for advancing 3D intelligence.

Multi-modal large reasoning models (MLRMs) pose significant privacy risks by inferring precise geographic locations from personal images through hierarchical chain-of-thought reasoning. Existing privacy protection techniques, primarily designed for perception-based models, prove ineffective against MLRMs' sophisticated multi-step reasoning processes that analyze environmental cues. We introduce **ReasonBreak**, a novel adversarial framework specifically designed to disrupt hierarchical reasoning in MLRMs through concept-aware perturbations. Our approach is founded on the key insight that effective disruption of geographic reasoning requires perturbations aligned with conceptual hierarchies rather than uniform noise. ReasonBreak strategically targets critical conceptual dependencies within reasoning chains, generating perturbations that invalidate specific inference steps and cascade through subsequent reasoning stages. To facilitate this approach, we contribute **GeoPrivacy-6K**, a comprehensive dataset comprising 6,341 ultra-high-resolution images ($\geq$2K) with hierarchical concept annotations. Extensive evaluation across seven state-of-the-art MLRMs (including GPT-o3, GPT-5, Gemini 2.5 Pro) demonstrates ReasonBreak's superior effectiveness, achieving a 14.4\% improvement in tract-level protection (33.8\% vs 19.4\%) and nearly doubling block-level protection (33.5\% vs 16.8\%). This work establishes a new paradigm for privacy protection against reasoning-based threats.

We study how generative world models trained on video games can go beyond mere reproduction of gameplay visuals to learning game mechanics—the modular rules that causally govern gameplay. We introduce a formalization of the concept of game mechanics that operationalizes mechanic-learning as a causal counterfactual inference task and uses the causal consistency principle to address the challenge of generating gameplay with world models that do not violate game rules. We present Multiverse Mechanica, a playable video game testbed that implements a set of ground truth game mechanics based on our causal formalism. The game natively emits training data, where each training example is paired with a set of causal DAGs that encode causality, consistency, and counterfactual dependence specific to the mechanic that is in play—these provide additional artifacts that could be leveraged in mechanic-learning experiments. We provide a proof-of-concept that demonstrates fine-tuning a pre-trained model that targets mechanic learning. Multiverse Mechanica is a testbed that provides a reproducible, low-cost path for studying and comparing methods that aim to learn game mechanics—not just pixels.

The foundation of reproducible science lies in protocols that are precise, logically ordered, and executable. The autonomous generation of these protocols through natural language queries could greatly improve the efficiency of the reproduction process. However, current leading large language models (LLMs) often generate incomplete or inconsistent protocols, limiting their utility. To address this limitation, we first introduce SciRecipe, a large-scale dataset of over 12K structured protocols spanning 27 biological subfields and encompassing both comprehension and problem-solving tasks. To further improve protocol generation, we propose the "Sketch-and-Fill" paradigm, which separates analysis, structuring, and expression to ensure each step is explicit and verifiable. Complementing this, the structured component-based reward mechanism evaluates step granularity, action order, and semantic fidelity, aligning model optimization with experimental reliability. Building on these components, we develop Thoth, trained through a staged Knowledge-to-Action process that progresses from knowledge acquisition to operational reasoning and ultimately to robust, executable protocol generation. Across multiple benchmarks, Thoth consistently surpasses both proprietary and open-source LLMs, achieving significant improvements in step alignment, logical sequencing, and semantic accuracy. Our approach paves the way for reliable scientific assistants that bridge knowledge with experimental execution.

Videos, with their unique temporal dimension, demand precise grounded understanding, where answers are directly linked to visual, interpretable evidence. Despite significant breakthroughs in text-based reasoning with large language models, multi-modal reasoning - especially for videos - remains limited. In this work, we fill this gap by introducing VideoMind, a novel video-language agent for temporal-grounded video reasoning. Our method involves two key innovations: (1) We identify four essential capabilities for grounded video reasoning and propose a role-based agentic workflow, comprising a planner to coordinate roles, a grounder for temporal event localization, a verifier to assess event candidates, and an answerer for question answering. (2) To efficiently integrate these roles during inference, we propose a novel Chain-of-LoRA mechanism, where a unified base model with multiple LoRA adapters is leveraged to enable seamless role switching, balancing efficiency and flexibility. Extensive experiments on 15 benchmarks across Grounded VideoQA, Video Temporal Grounding, and General VideoQA tasks demonstrate the effectiveness of the proposed scheme in advancing video agent, test-time scaling, and long-form video reasoning. Code, models, datasets, and demos are available at https://videomind.github.io/.

Large language models (LLMs) have demonstrated remarkable success across diverse domains through in-context learning or fine-tuning. However, adapting LLMs to Knowledge Base Question Answering (KBQA) remains challenging, as KBQA necessitates multi-step reasoning over large-scale structured knowledge bases. Directly prompting LLMs with entire knowledge bases incurs prohibitive computational costs, while existing methods provide limited guidance on effectively fine-tuning LLMs for such complex reasoning tasks. In this work, we propose Chain-of-Decomposition (\texttt{CoD}), a novel framework that decomposes KBQA into three modular steps: (1) an LLM-free retrieval module to extract query-relevant subgraphs from the knowledge base, (2) a parameter-free reformulation step that transforms retrieved contexts into structured reasoning paths, and (3) a lightweight LLM-based reasoning module trained to evaluate the logical validity of each path. By isolating computation-heavy retrieval and rule-based reformulation from LLM reasoning, \texttt{CoD} reduces task complexity and enables efficient fine-tuning focused solely on the final verification step. Comprehensive experiments demonstrate that Llama-2 7B, fine-tuned with the proposed \texttt{CoD} surpasses strong baselines, including GPT-4 augmented with retrieved knowledge, achieving state-of-the-art performance on WebQSP and MetaQA benchmarks.