ICLR 2026 论文列表

Large reasoning models (LRMs) excel at a long chain of reasoning but often fail to faithfully follow instructions regarding output format, constraints, or specific requirements. We investigate whether this gap can be closed by integrating an instruction-tuned model (ITM) into an LRM. Analyzing their differences in parameter space, namely task vectors, we find that their principal subspaces are nearly orthogonal across key modules, suggesting a lightweight merging with minimal interference. However, we also demonstrate that naïve merges are fragile because they overlook the output format mismatch between LRMs (with explicit *thinking* and *response* segments) and ITMs (answers-only). We introduce **RAIN-Merging** (Reasoning-Aware Instruction-attention guided Null-space projection Merging), a gradient-free method that integrates instruction following while preserving thinking format and reasoning performance. First, with a small reasoning calibration set, we project the ITM task vector onto the null space of forward features at thinking special tokens, which preserves the LRM's structured reasoning mechanisms. Second, using a small instruction calibration set, we estimate instruction attention to derive module-specific scaling that amplifies instruction-relevant components and suppresses leakage. Across four instruction-following benchmarks and nine reasoning & general capability benchmarks, RAIN-Merging substantially improves instruction adherence while maintaining reasoning quality. The gains are consistent across model scales and architectures, translating to improved performance in agent settings.

In the past year, distillation has seen a renewed prominence in large language model (LLM) pretraining, exemplified by the Llama-3.2 and Gemma model families. While distillation has historically been shown to improve statistical modeling, its effects on new paradigms key to modern LLMs—such as test-time scaling and in-context learning—remain underexplored. In this work, we make three main contributions. First, we show that pretraining with distillation yields models that exhibit remarkably better test-time scaling. Second, we observe that this benefit comes with a trade-off: distillation impairs in-context learning capabilities, particularly the one modeled via induction heads. Third, to demystify these findings, we study distilled pretraining in a sandbox of a bigram model, which helps us isolate the common principal factor behind our observations. Finally, using these insights, we shed light on various design choices for pretraining that should help practitioners going forward.

Models are susceptible to adversarially out-of-distribution (OOD) data despite large training-compute investments into their robustification. Zaremba et al. (2025) make progress on this problem at test time, showing LLM reasoning improves satisfaction of model specifications designed to thwart attacks, resulting in a correlation between reasoning effort and robustness to jailbreaks. However, this benefit of test compute fades when attackers are given access to gradients or multimodal inputs. We address this gap, clarifying that inference-compute offers benefits even in such cases. Our approach argues that compositional generalization, through which OOD data is understandable via its in-distribution (ID) components, enables adherence to defensive specifications on adversarially OOD inputs. Namely, we posit the Robustness from Inference Compute Hypothesis (RICH): inference-compute defenses profit as the model's training data better reflects the attacked data’s components. We empirically support this hypothesis across vision language model and attack types, finding robustness gains from test-time compute if specification following on OOD data is unlocked by compositional generalization. For example, InternVL 3.5 gpt-oss 20B gains little robustness when its test compute is scaled, but such scaling adds significant robustness if we first robustify its vision encoder. This correlation of inference-compute's robustness benefit with base model robustness is the rich-get-richer dynamic of the RICH: attacked data components are more ID for robustified models, aiding compositional generalization to OOD data. Thus, we advise layering train-time and test-time defenses to obtain their synergistic benefit.

How cost-effectively can strong reasoning abilities be achieved in language models? Driven by this question, we present Tina, a family of tiny reasoning models achieved with high cost-efficiency. Tina shows that substantial reasoning performance can be developed using only minimal resources, by applying low-rank adaptation (LoRA) during reinforcement learning (RL), to an already tiny 1.5B parameter base model. This minimalist approach produces models that are competitive with, and sometimes surpass, SOTA RL reasoning models built upon the same base model. Crucially, this is achieved at a tiny fraction of the computational cost employed by existing models. In fact, the best Tina model achieves a >20\% reasoning performance increase and 43.33\% zero-shot Pass@1 accuracy on AIME24, at only \$9 USD cost (i.e., an estimated 260x reduction). Our work reveals the surprising effectiveness of efficient RL reasoning via LoRA. We validate this across multiple open-source reasoning datasets and various ablation settings starting with a single, fixed set of hyperparameters. Furthermore, we explore the hypothesis that this effectiveness and efficiency stem from LoRA rapidly adapting the model to the structural format of reasoning rewarded by RL, while largely preserving the base model's underlying knowledge. In service of accessibility and open research, we fully open-source all code, training logs, model weights, and checkpoints.

While reinforcement learning from verifiable rewards (RLVR) typically has relied on a single binary verification signal, symbolic proof assistants in formal reasoning offer rich, fine-grained structured feedback. This gap between structured processes and unstructured rewards highlights the importance of feedback that is both dense and sound. In this work, we demonstrate that the Lean proof assistant itself can serve as a symbolic process oracle, supplying both outcome-level and fine-grained tactic-level verified feedback during training. Proof attempts are parsed into tactic sequences, and Lean's elaboration marks both locally sound steps and the earliest failing step, yielding dense, verifier-grounded credit signals rooted in type theory. We incorporate these structured rewards into a GRPO-style reinforcement learning objective with first-error propagation and first-token credit methods that balances outcome- and process-level advantages. Experiments with STP-Lean and DeepSeek-Prover-V1.5 show that tactic-level supervision outperforms outcome-only baselines in most settings, delivering improvements on benchmarks such as MiniF2F and ProofNet. Beyond empirical gains, our study highlights a broader perspective: symbolic proof assistants are not only verifiers at evaluation time, but can also act as process-level reward oracles during training. This opens a path toward reinforcement learning frameworks that combine the scalability of language models with the reliability of symbolic verification for formal reasoning.

This paper introduces LLM2Fx-Tools, a multimodal tool-calling framework that generates executable sequences of audio effects (Fx-chain) for music post-production. LLM2Fx-Tools uses a large language model (LLM) to understand audio inputs, select audio effects types, determine their order, and estimate parameters, guided by chain-of-thought (CoT) planning. We also present LP-Fx, a new instruction-following dataset with structured CoT annotations and tool calls for audio effects modules. Experiments show that LLM2Fx-Tools can infer an Fx-chain and its parameters from pairs of unprocessed and processed audio, enabled by autoregressive sequence modeling, tool calling, and CoT reasoning. We further validate the system in a style transfer setting, where audio effects information is transferred from a reference source and applied to new content. Finally, LLM-as-a-judge evaluation demonstrates that our approach generates appropriate CoT reasoning and responses for music production queries. To our knowledge, this is the first work to apply LLM-based tool calling to audio effects modules, enabling interpretable and controllable music production.

Recent advances in alignment techniques such as Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Direct Preference Optimization (DPO) have improved the safety of large language models (LLMs). However, these LLMs remain vulnerable to jailbreak attacks that disguise harmful intent through indirect or deceptive phrasing. Using causal intervention, we empirically demonstrate that this vulnerability stems from shallow alignment mechanisms that lack deep reasoning, often rejecting harmful prompts without truly understanding why they are harmful. To mitigate this vulnerability, we propose enhancing alignment through reasoning-aware post-training. We construct and release a novel Chain-of-Thought (CoT) fine-tuning dataset that includes both utility-oriented and safety-critical prompts with step-by-step rationales. Fine-tuning on this dataset encourages models to produce principled refusals grounded in reasoning, outperforming standard SFT baselines. Furthermore, inspired by failure patterns in CoT fine-tuning, we introduce **Alignment-Weighted DPO**, which targets the most problematic parts of an output by assigning different preference weights to the reasoning and final-answer segments. This produces finer-grained, targeted updates than vanilla DPO and improves robustness to diverse jailbreak strategies. Extensive experiments across multiple safety and utility benchmarks show that our method consistently improves alignment robustness while maintaining overall model utility.

Humans can imagine and manipulate visual images mentally, a capability known as \textit{spatial visualization}. While many multi-modal benchmarks assess reasoning on visible visual information, the ability to infer unseen relationships through spatial visualization remains insufficiently evaluated as a spatial skill. This reliance on publicly sourced problems from IQ tests or math competitions risks data contamination and compromises assessment reliability. To this end, we introduce \textbf{\textit{SpatialViz-Bench}}, a comprehensive multi-modal benchmark for \textit{spatial visualization} with \emph{12} tasks across \emph{4} sub-abilities, comprising \emph{1,180} programmatically generated problems, a scalable framework that allows for expansion to ensure fair and continuously reliable evaluations. Our evaluation of \emph{27} Multi-modal Large Language Models (MLLMs) reveals wide performance variations, demonstrates the benchmark's strong discriminative power, and uncovers counter-intuitive findings: Chain-of-Thought (CoT) prompting paradoxically degrades accuracy on open-source models. Through statistical and qualitative analysis of error types, SpatialViz-Bench demonstrates that state-of-the-art MLLMs exhibit deficiencies in \textit{spatial visualization} tasks, thereby addressing a significant lacuna in the field.

Multimodal foundation models, such as GPT-4o, have recently made remarkable progress, but it is not clear where exactly these models stand in terms of understanding vision. In this paper, we benchmark the performance of popular multimodal foundation models (GPT-4o, o4-mini, Gemini 1.5 Pro and Gemini 2.0 Flash, Claude 3.5 Sonnet, Qwen2-VL, Llama 3.2) on standard computer vision tasks (semantic segmentation, object detection, image classification, depth and surface normal prediction) and using established datasets (e.g., COCO, ImageNet and its variants, etc). The main challenges to performing this are: 1) most models are trained to output text and cannot natively express versatile domains, such as segments or 3D geometry, and 2) many leading models are proprietary and accessible only at an API level, i.e., there is no weight access to adapt them. We address these challenges by translating standard vision tasks into equivalent text-promptable and API-compatible tasks via prompt chaining to create a standardized benchmarking framework. We observe that 1) the models are not close to the state-of-the-art specialist models at any tasks, and 2) they perform semantic tasks notably better than geometric ones. However, 3) they are respectable generalists; this is remarkable as they are presumably trained on primarily image-text-based tasks. 4) While the prompt-chaining techniques affect performance, better models exhibit less sensitivity to prompt variations. 5) GPT-4o performs the best among non-reasoning models, securing the top position in 4 out of 6 tasks and 6) reasoning models, e.g. o3, show improvements in geometric tasks.

While chain-of-thought prompting enables powerful multi-step reasoning in Large Language Models (LLMs), the stochastic nature of the generation process undermines its reliability. In this work, we first analyze thousands of reasoning paths to identify Reasoning Drift, a key failure mode where models get locked into flawed reasoning patterns. We reveal that the manifestation of drift is a complex interplay between universal functional tendencies and unique, model-specific signatures. Based on the diagnosis, we propose Reasoning-Aware Activation Steering, a novel inference-time intervention method to gently nudge the model's activations away from pathological patterns. We pre-compute a library of vectors from contrastive functional transitions and apply them dynamically. Experiments show that our method effectively mitigates the drift problem and boosts accuracy. Additionally, it generalizes to out-of-distribution tasks, demonstrating a deeper capture of valid reasoning principles.

Vision language models (VLMs) can flexibly address various vision tasks through text interactions. Although successful in semantic understanding, state-of-the-art VLMs including GPT-5 still struggle in understanding 3D from 2D inputs. On the other hand, expert pure vision models achieve super-human accuracy in metric depth estimation, a key 3D understanding task. However, they require task-specific architectures and losses. Such difference motivates us to ask: Can VLMs reach expert-level accuracy without architecture or loss change? We take per-pixel metric depth estimation as the representative task and show that the answer is yes! Surprisingly, comprehensive analysis shows that text-based supervised-finetuning with sparse labels is sufficient for VLMs to unlock strong 3D understanding, no dense prediction head or complex regression/regularization loss is needed. The bottleneck lies in pixel reference and cross-dataset camera ambiguity, which we address through visual prompting and intrinsic-conditioned augmentation. With much smaller models, our method DepthLM surpasses the accuracy of most advanced VLMs by over 2x, making VLMs for the first time comparable with pure vision models. The simplicity of DepthLM also enables a single VLM to cover various 3D tasks beyond metric depth. Code and model are available at https://github.com/facebookresearch/DepthLM_Official.

Balancing exploration and exploitation remains a central challenge in reinforcement learning with verifiable rewards (RLVR) for large language models (LLMs). Current RLVR methods often overemphasize exploitation, leading to entropy collapse, reduced exploratory capacity, and ultimately limited performance gains. Although techniques that add randomness increase policy stochasticity, they frequently fail to escape dominant behavioral modes. The resulting sample-and-reward dynamics amplify these modes, eroding exploration and leading to entropy collapse. We introduce Exploration-Enhanced Policy Optimization (EEPO), a novel framework that promotes exploration through two-stage rollouts with adaptive unlearning. In the first stage, the model generates half of the trajectories; it then undergoes a lightweight, temporary unlearning step to suppress these sampled responses, forcing the second stage to explore different regions of the output space. This sample-then-forget mechanism actively steers the policy away from dominant modes and encourages mode-seeking exploration. Across five reasoning benchmarks, EEPO consistently outperforms baselines, achieving average gains of 24.3% on Qwen2.5-3B, 33.0% on Llama3.2-3B-Instruct, and 10.4% on Qwen3-8B-Base.

Training large language models (LLMs) for complex reasoning via Reinforcement Learning with Verifiable Rewards (RLVR) is effective but limited by reliance on costly, domain-specific supervision. We explore Reinforcement Learning from Internal Feedback (RLIF), a framework that enables LLMs to learn from intrinsic signals without external rewards or labeled data. We propose Intuitor, an RLIF method that uses a model's own confidence—termed self-certainty—as its sole reward signal. Intuitor replaces external rewards in Group Relative Policy Optimization (GRPO) with self-certainty scores, enabling fully unsupervised learning. Experiments demonstrate that Intuitor matches GRPO's performance on mathematical benchmarks and achieves competitive, sometimes better, generalization to out-of-domain tasks like code generation, without requiring gold solutions or test cases. Our findings show that intrinsic model signals can drive effective learning across domains, offering a scalable alternative to RLVR for autonomous AI systems where verifiable rewards are unavailable. Code is available in the supplementary materials.

While current Multimodal Large Language Models (MLLMs) have demonstrated proficiency in reasoning tasks such as mathematics and logic, their capacity for long-chain reflective reasoning, a prerequisite for solving complex real-world problems, remains largely underexplored. In this work, we first conduct an extensive empirical investigation to evaluate this capability. Leveraging a carefully designed data synthesis engine, we construct MM-HELIX, a multimodal benchmark consisting 1260 samples of 42 challenging synthetic tasks that require iterative thinking and backtracking. Empirical results on this benchmark reveal that existing MLLMs exhibit significant performance deficits in long-chain reflective reasoning. To address this limitation, we generate post-training data and further explore learning paradigms for exploiting such data. We first develop the Step-Elicited Response Generation pipeline to create MM-HELIX-100K, a large-scale dataset of 100k high-quality, reflective reasoning traces for instruction-tuning stage. Given that standard Reinforcement Learning fails on complex tasks due to sparse reward signals and catastrophic forgetting after Supervised Fine-Tuning, we propose Adaptive Hybrid Policy Optimization (AHPO), a novel training strategy that dynamically unifies offline supervision and online optimization into a single stage. This strategy enables the model to learn from expert data when rewards are sparse and conduct independent exploration once proficient. When applied to the Qwen2.5-VL-7B baseline, our method achieves a +18.6\% accuracy improvement on MM-HELIX benchmark and demonstrates strong generalization with a +5.7\% average performance gain on general mathematic and logic tasks. Our work demonstrate that reflective reasoning in MLLMs can be effectively learned and generalized, paving the way for developing more capable MLLMs.

Large Language Models (LLMs) have recently improved mathematical reasoning through Reinforcement Learning with Verifiable Reward (RLVR). However, existing RLVR algorithms require large query budgets, making annotation costly. We investigate whether fewer but more informative queries can yield similar or superior performance, introducing active learning (AL) into RLVR. We identify that classic AL sampling strategies fail to outperform random selection in this setting, due to ignoring \textbf{objective uncertainty} when only selecting by subjective uncertainty. This work proposes an \textbf{uncertainty consistency} metric to evaluate how well subjective uncertainty aligns with objective uncertainty. In the offline setting, this alignment is measured using the Point-Biserial Correlation Coefficient (PBC). For online training, because of limited sampling and dynamically shifting output distributions, PBC estimation is difficult. Therefore, we introduce a new online variant, computed from normalized advantage and subjective uncertainty. Theoretically, we prove that the online variant is strictly negatively correlated with offline PBC and supports better sample selection. Experiments show our method consistently outperforms random and classic AL baselines, achieving full-dataset performance while training on only 30\% of the data, effectively reducing the cost of RLVR for reasoning tasks.\footnote{The code is available at \hyperref[https://github.com/yihao-123/uncertainty-consistency]{https://github.com/yihao-123/uncertainty-consistency}. }

Chain-of-Thought (CoT) reasoning often fails to faithfully reflect a language model's underlying decision process. We address this by introducing a \emph{Markovian} language model framework with an autoencoder-style \emph{reasoning bottleneck}: it creates a text-based bottleneck where CoT serves as an intermediate representation, forcing the model to compress essential reasoning into interpretable text before making predictions, in the sense of learning short intermediate descriptions that make answers easy to compute from questions. We train this system with a GRPO-style policy gradient algorithm using parallel sampling, a frozen baseline CoT$'$, within-batch standardized advantages, and actor-reward (chain-rule) gradients. On QA tasks, Markovian training recovers most of the gains of a non-Markovian GRPO variant while forcing the model to answer from the CoT alone (e.g., GSM8K: 19.6\% $\to$ 57.1\%; ARC-Challenge: 36.1\% $\to$ 79.9\%; on average only $\approx$3-4 pp below a non-Markovian upper bound). Perturbation analyses across types and severities show that Markovian models incur systematically larger log-probability drops under CoT corruption than matched Non-Markovian baselines, indicating stronger causal reliance on the CoT. Cross-model evaluation confirms that learned CoTs generalize across architectures, suggesting they capture transferable reasoning patterns rather than model-specific artifacts.

While chain-of-thought (CoT) distillation from advanced large language models (LLMs) has proven effective in general reasoning tasks, it struggles in scientific domains where even advanced models often produce incorrect or superficial reasoning due to high complexity and specialized knowledge requirements. Directly distilling from such flawed outputs results in low-quality training data and limits the performance of smaller student models. To overcome this, we propose CoT-Evo, an evolutionary CoT distillation framework. It begins by constructing a diverse pool of reasoning trajectories from multiple LLM thinkers, enriches them with automatically retrieved domain knowledge, and iteratively refines the trajectories using novelty-driven selection, reflective recombination and mutation. The refinement is guided by a fitness function that evaluates answer correctness, coherence, and effective knowledge utilization. This results in a high-quality CoT dataset tailored for scientific reasoning. We employ this evolved dataset to fine-tune a compact model, which achieves state-of-the-art performance on scientific reasoning benchmarks. Our work establishes a scalable approach to synthesizing high-fidelity scientific reasoning data from diverse and fallible LLMs.

Current large language model (LLM) development treats task-solving and preference-alignment as separate challenges, optimizing first for objective correctness, then for alignment to aggregated human preferences. This paradigm fails in human-facing applications where solving a problem correctly is insufficient if the response mismatches the user’s needs. This challenge intensifies in just-in-time scenarios where no prior user interaction history exists due to cold-start conditions or privacy constraints. LLMs need to identify what they don’t know about user preferences, strategically elicit preference values through questioning, then adapt their reasoning processes and responses accordingly—a complicated chain of cognitive processes which we term personalized reasoning. We introduce PREFDISCO, an evaluation methodology that transforms static benchmarks into interactive personalization tasks using psychologically-grounded personas with sparse preferences. Our framework creates scenarios where identical questions require different reasoning chains depending on user context, as optimal explanation approaches vary by individual expertise and preferences while maintaining factual accuracy. Evaluation of 21 frontier models across 10 tasks reveals 29.0% of naive personalization attempts produce worse preference alignment than generic responses, yet generic responses also fail to serve individual user needs effectively. These findings suggest personalized reasoning requires dedicated development rather than emerging naturally. PREFDISCO establishes personalized reasoning as a measurable research frontier and reveals fundamental limitations in current LLMs’ interactive capabilities, providing a foundation for developing systems that can adapt to individual users in education, healthcare, and technical domains where personalization is critical.

Recent advances in large reasoning language models (LRMs) rely on test-time scaling, which extends long chain-of-thought (CoT) generation to solve complex tasks. However, overthinking in long CoT not only slows down the efficiency of problem solving, but also risks accuracy loss due to the extremely detailed or redundant reasoning steps. We propose a simple yet effective method that allows LLMs to self-truncate CoT sequences by early exit during generation. Instead of relying on fixed heuristics, the proposed method monitors model behavior at potential reasoning transition points and dynamically terminates the next reasoning chain's generation when the model exhibits high confidence in a trial answer. Our method requires no additional training and can be seamlessly integrated into existing o1-like reasoning LLMs. Experiments on 10 reasoning benchmarks (e.g., GSM8K, MATH-500, AMC, GPQA, AIME and LiveCodeBench) show that the proposed method is consistently effective on 11 cutting-edge reasoning LLMs of varying series and sizes, reducing the length of CoT sequences by an average of 19.1% to 80.1% while improving accuracy by 0.3% to 5.0%.