ICLR 2026 Papers

Despite significant evolution of CUDA programming and domain-specific libraries, effectively utilizing GPUs with massively parallel engines remains difficult. Large language models (LLMs) show strong potential in generating optimized CUDA code from sequential code. However, using LLMs in practice faces two major challenges: cloud-based APIs pose risks of code leakage, and local deployment is often computationally expensive and inefficient. These drawbacks have spurred interest in small language models (SLMs), which are more lightweight and privacy-friendly. Encouragingly, recent studies show that SLMs can achieve performance comparable to LLMs on specific tasks. While SLMs can match LLMs on domain-specific tasks, their limited reasoning abilities lead to suboptimal performance in complex CUDA generation according to our experiments. To bridge this gap, we propose ReGraphT, a training-free, retrieval-augmented generation framework that transfers LLM-level reasoning to smaller models. ReGraphT organizes CUDA optimization trajectories into a structured reasoning graph, modeling the combined CUDA optimizations as state transitions, and leverages Monte Carlo Graph Search (MCGS) for efficient exploration. We also present a CUDA-specific benchmark with difficulty tiers defined by reasoning complexity to evaluate models more comprehensively. Experiments show that ReGraphT outperforms HPC-specific fine-tuned models and other retrieval-augmented approaches, achieving an average 2.33× speedup on CUDAEval and ParEval. When paired with DeepSeek-Coder-V2-Lite-Instruct and Qwen2.5-Coder-7B-Instruct, ReGraphT enables SLMs to approach LLM-level performance without the associated privacy risks or excessive computing overhead.

Reinforcement learning (RL) finetuning has become a key technique for enhancing the reasoning abilities of large language models (LLMs). However, its effectiveness critically depends on the selection of training data. Recent advances underscore the importance of online prompt selection methods, which typically concentrate training on partially solved or moderately challenging examples under the current policy, thereby yielding more effective model updates. While significantly accelerating RL finetuning in terms of training steps, they also incur substantial computational overhead by requiring extensive LLM rollouts over large candidate batches to identify informative samples, an expense that can outweigh the finetuning process itself. To address this challenge, this work proposes Dynamics-Predictive Sampling (DPS), which online predicts and selects informative prompts by inferring their learning dynamics prior to costly rollouts. Specifically, we introduce a new perspective by modeling each prompt's solving progress during RL finetuning as a dynamical system, where the extent of solving is represented as the state and the transition is characterized by a hidden Markov model. Using historical rollout reward signals, we perform online Bayesian inference to estimate evolving state distributions, and the inference outcome provides a predictive prior for efficient prompt selection without rollout-intensive filtering. Empirical results across diverse reasoning tasks, including mathematics, planning, and visual geometry, demonstrate that DPS substantially reduces redundant rollouts, accelerates the training process, and achieves superior reasoning performance.

Test-time compute allocation in large reasoning models (LRMs) is widely used and has applications in mathematical problem solving, code synthesis, and planning. Recent work has addressed this problem by scaling self-consistency and parallel thinking, adding generic thinking tokens and prompting models to re-read the question before answering. Unfortunately, these approaches either inject task-agnostic tokens or mandate heuristics that do not explain---and often ignore---the \emph{spontaneous} repetition that many LRMs exhibit at the head of their internal chains. In contrast, we analyze and harness the model's tendency to restate the question, which we term the \emph{Echo of Prompt (EOP)}, as a front-loaded, compute-shaping mechanism. We formalize its probabilistic cost by casting echo removal as rejection-based conditioning and defining the \emph{Echo Likelihood Gap} $\Delta\mathcal{L}$ as a computable proxy. This provides the missing theoretical link that links early repetition to likelihood gains and downstream accuracy. However, it does not by itself specify how to exploit EOP. Consequently, we develop \emph{Echo-Distilled SFT (ED-SFT)} to instill an ``echo-then-reason'' pattern through supervised finetuning, and \emph{Echoic Prompting (EP)} to re-ground the model mid-trace without training. While promising, quantifying benefits beyond verbosity is non-trivial. Therefore, we conduct length and suffix-controlled likelihood analyses together with layer-wise attention studies, showing that EOP increases answer to answer-prefix attention in middle layers, consistent with an \emph{attention refocusing} mechanism. We evaluate under identical decoding settings and compute budgets on GSM8K, MathQA, Hendrycks-MATH, AIME24, and MATH-500 under identical decoding settings and budgets, and find consistent gains over baselines.

In recent years, the research focus of large language models (LLMs) and agents has shifted increasingly from demonstrating novel capabilities to complex reasoning and tackling challenging tasks. However, existing evaluations focus mainly on math/code contests or general tasks, while existing multi-domain academic benchmarks lack sufficient reasoning depth, leaving the field without a rigorous benchmark for high-level reasoning. To fill this gap, we introduce the ACADREASON benchmark, designed to evaluate the ability of LLMs and agents to acquire and reason over academic knowledge. It consists of 50 expert-annotated academic problems across five high-reasoning domains, including computer science, economics, law, mathematics, and philosophy. All questions are sourced from top-tier publications in recent years and undergo rigorous annotation and quality control to ensure they are both challenging and answerable. We conduct systematic evaluations over 10 mainstream LLMs and agents. The results show that most LLMs scored below 20 points, with even the cutting-edge GPT-5 achieving only 16 points. While agents achieved higher scores, none exceeded 40 points. This demonstrates the current capability gap between LLMs and agents in super-intelligent academic research tasks and highlights the challenges of ACADREASON. The code and data for the ACADREASON benchmark are available at https://anonymous.4open.science/r/Acadreason-Benchmark-1BD3/.

As Large Language Models (LLMs) are increasingly applied in high-stakes domains, their ability to reason strategically under uncertainty becomes critical. Poker provides a rigorous testbed, requiring not only strong actions but also principled, game-theoretic reasoning. In this paper, we conduct a systematic study of LLMs in multiple realistic poker tasks, evaluating both gameplay outcomes and reasoning traces. Our analysis reveals LLMs fail to compete against traditional algorithms and identifies three recurring flaws: reliance on heuristics, factual misunderstandings, and a “knowing–doing” gap where actions diverge from reasoning. An initial attempt with behavior cloning and step-level reinforcement learning improves reasoning style but remains insufficient for accurate game-theoretic play. Motivated by these limitations, we propose ToolPoker, a tool-integrated reasoning framework that combines external solvers for GTO-consistent actions with more precise professional-style explanations. Experiments demonstrate that ToolPoker achieves state-of-the-art gameplay while producing reasoning traces that closely reflect game-theoretic principles.

Temporal reasoning over long, multi-session dialogues is a critical capability for conversational agents. As dialogue histories grow in length and accumulate noise, existing long-context models struggle to accurately identify temporally pertinent information, significantly impairing reasoning performance. To address this, we introduce **Memory-T1**, a framework that learns a time-aware memory selection policy using reinforcement learning (RL). It employs a coarse-to-fine strategy, first pruning the dialogue history into a candidate set with temporal and retriever filters, followed by an RL agent that selects the precise evidence. The RL training is guided by a multi-level reward function optimizing (i) accuracy, (ii) evidence grounding, and (iii) temporal consistency. This temporal consistency reward provides a dense signal by evaluating alignment at both the session-level (range proximity) and the utterance-level (evidence density), enabling the agent to resolve subtle chronological ambiguities. On the Time-Dialog benchmark, Memory-T1 boosts a 7B model to an overall score of 67.0\%, establishing a new state-of-the-art performance for open-source models and outperforming a 14B baseline by 10.2\%. Ablation studies show temporal consistency and evidence grounding rewards jointly contributing to a 15.0\% performance gain.Moreover, Memory-T1 maintains robustness up to 128k tokens, where baseline models collapse, proving effectiveness against noise in extensive dialogue histories.

Unlearning in Large Language Models (LLMs) is crucial for protecting private data and removing harmful knowledge. Most existing approaches rely on fine-tuning to balance unlearning efficiency with general language capabilities. However, these methods typically require training or access to retain data, which is often unavailable in real world scenarios. Although these methods can perform well when both forget and retain data are available, few works have demonstrated equivalent capability in more practical, data-limited scenarios. To overcome these limitations, we propose Detect-Reasoning Augmented GeneratiON (DRAGON), a systematic, reasoning-based framework that utilizes in-context chain-of-thought (CoT) instructions to guard deployed LLMs before inference. Instead of modifying the base model, DRAGON leverages the inherent instruction-following ability of LLMs and introduces a lightweight detection module to identify forget-worthy prompts without any retain data. These are then routed through a dedicated CoT guard model to enforce safe and accurate in-context intervention. To robustly evaluate unlearning performance, we introduce novel metrics for unlearning performance and the continual unlearning setting. Extensive experiments across three representative unlearning tasks validate the effectiveness of DRAGON, demonstrating its strong unlearning capability, scalability, and applicability in practical scenarios.

Agent memory shapes how Large Language Model (LLM)-powered agents, akin to the human brain, progressively refine themselves through environment interactions. Existing paradigms remain constrained: parametric memory forcibly adjusts model parameters, and retrieval-based memory externalizes experience into structured databases, yet neither captures the fluid interweaving of reasoning and memory that underlies human cognition. To address this gap, we propose MemGen, a dynamic generative memory framework that equips agents with a human-esque cognitive faculty. It consists of a \textit{memory trigger}, which monitors the agent’s reasoning state to decide explicit memory invocation, and a \textit{memory weaver}, which takes the agent's current state as stimulus to construct a latent token sequence as machine-native memory to enrich its reasoning. In this way, MemGen enables agents to recall and augment latent memory throughout reasoning, producing a tightly interwoven cycle of memory and cognition. Extensive experiments across eight benchmarks show that MemGen surpasses leading external memory systems such as ExpeL and AWM by up to $38.22\\%$, exceeds GRPO by up to $13.44\\%$, and exhibits strong cross-domain generalization ability. More importantly, we find that without explicit supervision, MemGen spontaneously evolves distinct human-like memory faculties, including planning memory, procedural memory, and working memory, suggesting an emergent trajectory toward more naturalistic forms of machine cognition.

The widespread adoption of AI systems in the economy hinges on their ability to generate economic value that outweighs their inference costs. Evaluating this tradeoff requires metrics that account for both performance and costs. Building on production theory, we develop an economically grounded framework for evaluating language models by combining accuracy and inference cost. We formalize cost-of-pass, the expected monetary cost of generating a correct solution. We then define the frontier cost-of-pass as the minimum cost-of-pass achievable across available models or the human-expert, using the approximate cost of hiring an expert. Our analysis reveals distinct economic insights. First, lightweight models are most cost-effective for basic quantitative tasks, large models for knowledge-intensive ones, and reasoning models for complex quantitative problems, despite higher per-token costs. Second, tracking this frontier cost-of-pass over the past year reveals significant progress, particularly for complex quantitative tasks where the cost has roughly halved every few months. Third, to trace key innovations driving this progress, we examine counterfactual frontiers—estimates of cost-efficiency without specific model classes. We find that innovations in lightweight, large, and reasoning models have been essential for pushing the frontier in basic quantitative, knowledge-intensive, and complex quantitative tasks, respectively. Finally, we assess the cost-reductions from common inference-time techniques (majority voting and self-refinement), and a budget-aware technique (TALE-EP). We find that performance-oriented methods with marginal performance gains rarely justify the costs, while TALE-EP shows some promise. Overall, our findings underscore that complementary model-level innovations are the primary drivers of cost-efficiency, and our economic framework provides a principled tool for measuring this progress and guiding deployment.

Chain-of-thought (CoT) prompting is a de-facto standard technique to elicit reasoning-like responses from large language models (LLMs), allowing them to spell out individual steps before giving a final answer. While the resemblance to human-like reasoning is undeniable, the driving forces underpinning the success of CoT reasoning still remain largely unclear. In this work, we perform an in-depth analysis of CoT traces originating from competition-level mathematics questions, with the aim of better understanding how, and which parts of CoT actually contribute to the final answer. To this end, we introduce the notion of a \textit{potential}, quantifying how much a given part of CoT increases the likelihood of a correct completion. Upon examination of reasoning traces through the lens of the potential, we identify surprising patterns including (1) its often strong non-monotonicity (due to reasoning tangents), (2) very sharp but sometimes tough to interpret spikes (reasoning insights and jumps) as well as (3) at times lucky guesses, where the model arrives at the correct answer without providing any relevant justifications before. While some of the behaviours of the potential are readily interpretable and align with human intuition (such as insights and tangents), others remain difficult to understand from a human perspective. To further quantify the reliance of LLMs on reasoning insights, we investigate the notion of CoT transferability, where we measure the potential of a weaker model under the partial CoT from another, stronger model. Indeed aligning with our previous results, we find that as little as 20% of partial CoT can ``unlock'' the performance of the weaker model on problems that were previously unsolvable for it, highlighting that a large part of the mechanics underpinning CoT are transferable.

Flowcharts are widely used to represent processes and relationships through intuitive visual representations. However, accurately interpreting these diagrams remains challenging due to their structural complexity and high visual diversity. Existing flowchart datasets often lack fine-grained control over key properties such as graph complexity and rendering style, limiting their utility for training and testing of multimodal large language models (MLLMs) on visual reasoning tasks. To address these limitations, we introduce FlowGen, a controllable synthesizer that generates flowcharts that have customizable structural features and supports multiple renderer backends. FlowGen enables fine-grained control over graph properties such as graph order and size, branched arrows, and nested subgraphs, facilitating systematic evaluation of MLLMs' capabilities. Extensive experiments on open-source and proprietary MLLMs show that training on FlowGen substantially improves flowchart parsing and question answering (QA), while also enhancing generalization to other public datasets. Furthermore, FlowGen provides challenging test datasets that expose consistent weaknesses in current MLLMs, particularly related to high structural complexity and varied rendering styles. Our code and data are publicly available at https://github.com/nju-websoft/FlowGen.

Long chain-of-thought (CoT) is an essential ingredient in effective usage of modern large language models, but our understanding of the reasoning strategies underlying these capabilities remains limited. While some prior works have attempted to categorize CoTs using predefined strategy types, such approaches are constrained by human intuition and fail to capture the full diversity of model behaviors. In this work, we introduce the CoT Encyclopedia, a bottom-up framework for analyzing and steering model reasoning. Our method automatically extracts diverse reasoning criteria from model-generated CoTs, embeds them into a semantic space, clusters them into representative categories, and derives contrastive rubrics to interpret reasoning behavior. Human evaluations show that this framework produces more interpretable and comprehensive analyses than existing methods. Moreover, we show that this understanding translates into measurable improvements on both problem-solving and safety benchmarks. We can predict which strategy a model is likely to use and guide it toward more effective alternatives. Finally, we show that training data format (e.g., free-form vs. multiple-choice) impacts reasoning far more than data domain, highlighting the importance of format-aware model design. In short, the CoT Encyclopedia turns reasoning from a black box into a controllable asset, enabling LLMs that think more clearly, perform more reliably, and act more safely.

Reinforcement learning (RL) based on the final answer's reward has driven recent progress in small language models (SLMs) on reasoning-heavy tasks such as math and code. However, applying the same techniques to retrieval-augmented generation (RAG) benchmarks like multi-hop QA has yielded limited gains—often trailing supervised or prompting-only baselines. Instead, we argue that a viable path for RL in multi-hop QA is to use test-time scaling judiciously, for optimizing both the final answer accuracy and the efficiency in reaching that answer. We propose FrugalRAG, a two-stage finetuning framework that adaptively _reduces_ the number of retrieval steps based on a question's difficulty. First, we train an SLM with supervised finetuning on a full-exploration policy that generates broad sub-queries. Then, we apply RL to adaptively prune search depth based on question difficulty, directly rewarding policies that balance correctness with frugality. Unlike prior approaches requiring 10× more data, our method achieves competitive performance with only ~1,000 examples. On HotPotQA and other multi-hop QA benchmarks, FrugalRAG attains state-of-the-art efficiency–accuracy tradeoffs, cutting retrieval cost nearly in half. Moreover, on the challenging BrowseCompPlus benchmark, it surpasses SLM-based and other baselines after training on only 200 examples. These results demonstrate the use of RL—not to increase reasoning steps but to reduce them—as an effective solution for scalable, efficient RAG.

In post-training for reasoning Large Language Models (LLMs), the current state of practice trains LLMs in two independent stages: Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR, shortened as "RL" below). In this work, we challenge whether high SFT scores translate to improved performance after RL. We provide extensive counter-examples where this is not true. We find high SFT scores can be biased toward simpler or more homogeneous data and are not reliably predictive of subsequent RL gains or scaled-up post-training effectiveness. In some cases, RL training on models with improved SFT performance could lead to substantially worse outcome compared to RL on the base model without SFT. We study alternative metrics and identify generalization loss on held-out reasoning examples and Pass@large k performance to provide strong proxies for the RL outcome. We trained hundreds of models up to 12B-parameter with SFT and RLVR via GRPO and ran extensive evaluations on 7 math benchmarks with up to 256 repetitions, spending $>$1M GPU hours. Experiments include models from Llama3, Mistral-Nemo, Qwen3 and multiple state-of-the-art SFT/RL datasets. Compared to directly predicting from pre-RL performance, prediction based on generalization loss and Pass@large k achieves substantial higher precision, improving $R^2$ coefficient and Spearman's rank correlation coefficient by up to 0.5 (2x). This provides strong utility for broad use cases. For example, in most experiments, we find SFT training on unique examples for a one epoch underperforms training on half examples for two epochs, either after SFT or SFT-then-RL; With the same SFT budget, training only on short examples may lead to better SFT performance, though, it often leads to worse outcome after RL compared to training on examples with varying lengths. This work develops an enhanced evaluation tool that will be open-sourced.

Systematic reviews (SR), in which experts summarize and analyze evidence across individual studies to provide insights on a specialized topic, are a cornerstone for evidence-based clinical decision-making, research, and policy. Given the exponential growth of scientific articles, there is growing interest in using large language models (LLMs) to automate SR generation. However, the ability of LLMs to critically assess evidence and reason across multiple documents to provide recommendations at the same proficiency as domain experts remains poorly characterized. We therefore ask: Can LLMs match the conclusions of systematic reviews written by clinical experts when given access to the same studies? To explore this question, we present MedEvidence, a benchmark pairing findings from 100 SRs with the studies they are based on. We benchmark 24 LLMs on MedEvidence, including reasoning, non-reasoning, medical specialist, and models across varying sizes (from 7B-700B). Through our systematic evaluation, we find that reasoning does not necessarily improve performance, larger models do not consistently yield greater gains, and knowledge-based fine-tuning degrades accuracy on MedEvidence. Instead, most models exhibit similar behavior: performance tends to degrade as token length increases, their responses show overconfidence, and, contrary to human experts, all models show a lack of scientific skepticism toward low-quality findings. These results suggest that more work is still required before LLMs can reliably match the observations from expert-conducted SRs, even though these systems are already deployed and being used by clinicians. We release our codebase and benchmark to the broader research community to further investigate LLM-based SR systems.

Recent advances in multi-modal large reasoning models (MLRMs) have shown significant ability to interpret complex visual content. While these models possess impressive reasoning capabilities, they also introduce novel and underexplored privacy risks. In this paper, we identify a novel category of privacy leakage in MLRMs: Adversaries can infer sensitive geolocation information, such as users' home addresses or neighborhoods, from user-generated images, including selfies captured in private settings. To formalize and evaluate these risks, we propose a three-level privacy risk framework that categorizes image based on contextual sensitivity and potential for geolocation inference. We further introduce DoxBench, a curated dataset of 500 real-world images reflecting diverse privacy scenarios divided into 6 categories. Our evaluation across 13 advanced MLRMs and MLLMs demonstrates that most of these models outperform non-expert humans in geolocation inference and can effectively leak location-related private information. This significantly lowers the barrier for adversaries to obtain users' sensitive geolocation information. We further analyze and identify two primary factors contributing to this vulnerability: (1) MLRMs exhibit strong geolocation reasoning capabilities by leveraging visual clues in combination with their internal world knowledge; and (2) MLRMs frequently rely on privacy-related visual clues for inference without any built-in mechanisms to suppress or avoid such usage. To better understand and demonstrate real-world attack feasibility, we propose GeoMiner, a collaborative attack framework that decomposes the prediction process into two stages consisting of clue extraction and reasoning to improve geolocation performance. Our findings highlight the urgent need to reassess inference-time privacy risks in MLRMs to better protect users' sensitive information.

Reasoning language models such as DeepSeek-R1 produce long chain-of-thought traces during inference time which make them costly to deploy at scale. We show that using compression techniques such as neural network pruning produces greater performance loss than in typical language modeling tasks, and in some cases can make the model slower since they cause the model to produce more thinking tokens but with worse performance. We show that this is partly due to the fact that standard LLM pruning methods often focus on input reconstruction, whereas reasoning is a decode-dominated task. We introduce a simple, drop-in fix: during pruning we jointly reconstruct activations from the input and the model’s on-policy chain-of-thought traces. This “Reasoning-Aware Compression” (RAC) integrates seamlessly into existing pruning workflows such as SparseGPT, and boosts their performance significantly. Anonymized code can be found at: https://github.com/Anon-ICLR-RAC/ICLR-2025-Anonymous-Submission

Document reranking is a key component in information retrieval (IR), aimed at refining initial retrieval results to improve ranking quality for downstream tasks. Recent studies—motivated by large reasoning models (LRMs)—have begun incorporating explicit chain-of-thought (CoT) reasoning into LLM-based rerankers. However, the effectiveness of such reasoning for ranking tasks remains underexplored. In this work, we present the first systematic study of reasoning in reranking across both logits-based pointwise and listwise settings, under both supervised fine-tuning and reinforcement learning. Using diverse benchmarks, including reasoning-intensive datasets (BRIGHT) and standard IR benchmarks (BEIR), we find that reasoning-augmented rerankers consistently underperform their direct counterparts that predict rankings without CoT, despite substantially higher inference costs. Our analysis reveals three core limitations: (i) in pointwise rerankers, reasoning breaks calibration and biases models toward the positive class, raising TPR but lowering TNR, which inflates false positives and degrades ranking in negative-dominant pools; (ii) in listwise rerankers, explicit reasoning improves the fit during training but leads to higher variance and fails to improve performance in both in-domain and out-of-domain evaluations, even when reinforcement learning shortens rationales; and (iii) overall, directly fine-tuned rerankers remain more stable, effective, and robust. These findings challenge the assumption that explicit reasoning is universally beneficial for reranking. We conclude by highlighting future directions, including calibration-aware scoring for pointwise rerankers and the design of concise, targeted reasoning strategies to mitigate overfitting and overthinking in listwise rerankers.

Reasoning models leveraging long chains of thought employ various cognitive skills, such as verification of their answers, backtracking, retrying by an alternate method, and more. Previous work has shown that when a base language model exhibits these skills, training that model further with reinforcement learning (RL) can learn to leverage them. How can we get models to leverage skills that aren't exhibited by base models? Our work, SkillFactory, is a method for fine-tuning models to roughly learn these skills during a supervised fine-tuning (SFT) stage prior to RL. Our approach does not rely on distillation from a stronger model, but instead uses samples from the model itself, rearranged to provide training data in the format of those skills. These "silver" SFT traces may be imperfect, but are nevertheless effective for priming a model to acquire skills during RL. Our evaluation shows that (1) starting from SkillFactory SFT initialization helps a model to generalize to harder variants of a task post-RL, despite lower performance pre-RL; (2) cognitive skills are indeed used by the model; (3) RLed SkillFactory models are more robust to regression on out-of-domain tasks than RLed base models. Our work suggests that inductive biases learned prior to RL help models learn robust cognitive skill use.