ICLR 2026 Papers

Recent advances in large language models (LLMs) have enabled agents that automate scientific discovery by interpreting data, generating analysis pipelines, and executing them with computational tools. However, existing benchmarks remain largely limited to unimodal datasets and slice-level tasks, overlooking the fact that real discovery requires multimodal integration, modeling, and hypothesis-driven reasoning. To address this gap, we introduce MoSciBench, the first benchmark for multimodal scientific discovery, constructed from peer-reviewed studies through a principled four-stage pipeline. MoSciBench spans six scientific domains, seven data modalities, and five categories of discovery questions, yielding 88 individual, end-to-end, data-driven tasks. Each task is designed as a cross-modal hypothesis verification workflow, requiring agents to align and integrate heterogeneous datasets before modeling and reasoning. We further evaluate four representative agent frameworks across multiple LLM families. Results show that multimodal discovery is substantially harder than unimodal tasks: even the strongest agents achieve only 48.94\% accuracy, with over 60\% of failures due to cross-modal alignment. Lightweight workflow scaffolding consistently improves performance, reducing alignment errors by 5–10\% and raising accuracy by 5.7\% on average. Our benchmark and evaluation framework thus establish a rigorous testbed for advancing LLM agents toward realistic, multimodal scientific discovery.

Dual Modular Redundancy (DMR) is a highly effective mechanism for detecting silent data corruption (SDC)—a critical reliability concern in large language model (LLM) training—by executing each operation twice. However, its high computation overhead has prevented practical deployment at scale. In this paper, we present SpareTrain, an LLM training system that achieves complete DMR with minimal overhead by repurposing the activation checkpointing mechanism and exploiting idle GPU time. Evaluations on up to 32 H200 GPUs show that SpareTrain improves throughput by 12–35\% over naive DMR, corresponding to only 3–14\% overhead compared to unprotected training, while maintaining full DMR error detection capabilities.

Large Language Models (LLMs) have emerged as a promising approach for binary decompilation. However, the existing LLM-based decompilers still are somewhat limited in effectively presenting a program's source-level structure with its original identifiers. To mitigate this, we introduce SK2Decompile, a novel two-phase approach to decompile from the skeleton (semantic structure) to the skin (identifier) of programs. Specifically, we first apply a Structure Recovery model to translate a program's binary code to an Intermediate Representation (IR) as deriving the program's "skeleton", i.e., preserving control flow and data structures while obfuscating all identifiers with generic placeholders. We also apply reinforcement learning to reward the model for producing program structures that adhere to the syntactic and semantic rules expected by compilers. Second, we apply an Identifier Naming model to produce meaningful identifiers which reflect actual program semantics as deriving the program's "skin". We train the Identifier Naming model with a separate reinforcement learning objective that rewards the semantic similarity between its predictions and the reference code. Such a two-phase decompilation process facilitates advancing the correctness and readability of decompilation independently. Our evaluations indicate that SK2Decompile, significantly outperforms the SOTA baselines, achieving 21.6% average re-executability rate gain over GPT-5-mini on the HumanEval dataset and 29.4% average R2I improvement over Idioms on the GitHub2025 benchmark.

The Model Context Protocol (MCP) standardizes how large language model (LLM) agents discover, describe, and call external tools. While MCP unlocks broad interoperability, it also enlarges the attack surface by making tools first-class, composable objects with natural-language metadata, and standardized I/O. We present MSB (MCP Security Benchmark), the first end-to-end evaluation suite that systematically measures how well LLM agents resist MCP-specific attacks throughout the full tool-use pipeline: task planning, tool invocation, and response handling. MSB contributes: (1) a taxonomy of 12 attacks including name-collision, preference manipulation, prompt injections embedded in tool descriptions, out-of-scope parameter requests, user-impersonating responses, false-error escalation, tool-transfer, retrieval injection, and mixed attacks; (2) an evaluation harness that executes attacks by running real tools (both benign and malicious) via MCP rather than simulation; and (3) a robustness metric that quantifies the trade-off between security and performance: Net Resilient Performance (NRP). We evaluate nine popular LLM agents across 10 domains and 405 tools, producing 2,000 attack instances. Results reveal the effectiveness of attacks against each stage of MCP. Models with stronger performance are more vulnerable to attacks due to their outstanding tool calling and instruction following capabilities. MSB provides a practical baseline for researchers and practitioners to study, compare, and harden MCP agents.

Reinforcement Learning, particularly through policy gradient methods, has played a central role in enabling reasoning capabilities of Large Language Models. However, the optimization stability of policy gradients in this setting remains understudied. As a result, existing implementations often resort to conservative hyperparameter choices to ensure stability, which requires more training samples and increases computational costs. Hence, developing models for reliably tracking the underlying optimization dynamics and leveraging them into training enables more sample-efficient regimes and further unleashes scalable post-training. We address this gap by formalizing the stochastic optimization problem of policy gradients with explicit consideration of second-order geometry. We propose a tractable computational framework that tracks and leverages curvature information during policy updates. We further employ this framework to design interventions in the optimization process through data selection. The resultant algorithm, Curvature-Aware Policy Optimization (CAPO), identifies samples that contribute to unstable updates and masks them out. Theoretically, we establish monotonic improvement guarantees under realistic assumptions. On standard math reasoning benchmarks, we empirically show that CAPO ensures stable updates under aggressive learning regimes where baselines catastrophically fail. With minimal intervention (rejecting fewer than 8% of tokens), CAPO achieves up to 30$\times$ improvement in sample efficiency over standard GRPO for LLM reasoning.

Modern data parallel (DP) training favors collective communication over parameter servers (PS) for its simplicity and efficiency under balanced workloads. However, the balanced workload assumption no longer holds in large language model (LLM) post-training due to the large variance in sequence lengths. Under imbalanced workloads, collective communication creates synchronization barriers, leading to under-utilization of devices with smaller workloads. This change in training dynamics calls for a revisit of the PS paradigm for its robustness to such imbalance. We propose **On-Demand Communication (ODC)**, which adapts PS into Fully Sharded Data Parallel (FSDP) by replacing collective all-gather and reduce-scatter with direct point-to-point communication. Compared to FSDP, ODC reduces the synchronization barrier from once per layer to once per minibatch and decouples the workload on each device so that faster workers are not stalled. It also enables simpler and more effective load balancing at the minibatch level. Across diverse LLM post-training tasks, ODC consistently improves device utilization and training throughput, achieving up to a 36\% speedup over standard FSDP. These results demonstrate that ODC is a superior fit for the prevalent imbalanced workloads in LLM post-training. Our implementation of ODC and integration with FSDP is open-sourced at https://github.com/sail-sg/odc.

Deep Research Agents (DRAs) are emerging as one of the most practical classes of LLM-based agents. Given an open-ended research task, they find, analyze, and synthesize large numbers of online sources to produce a comprehensive report at the level of a research analyst. This can compress hours of manual desk research into minutes. However, a comprehensive benchmark for systematically evaluating the capabilities of these agents remains absent. To bridge this gap, we introduce DeepResearch Bench, a benchmark consisting of 100 PhD-level research tasks, each meticulously crafted by domain experts across 22 distinct fields. To evaluate DRAs comprehensively, we propose two complementary and fully automated methodologies. The first is a reference-based method with adaptive criteria to assess the quality of generated research reports. The second evaluates a DRA’s information‑retrieval and collection capabilities by assessing its effective citation count and overall citation accuracy. By conducting extensive human consistency experiments, we demonstrate that our evaluation methods are highly aligned with expert judges and faithfully reflect human judgments of quality differences among DRA-generated content. We are open-sourcing DeepResearch Bench and key components of these frameworks to accelerate the development of practical LLM-based agents.

The common approach to communicate a large language model's (LLM) uncertainty is to add a percentage number or a hedging word to its response. But is this all we can do? Instead of generating a single answer and then hedging it, an LLM that is fully transparent to the user needs to be able to reflect on its internal belief distribution and output a summary of all options it deems possible, and how likely they are. To test whether LLMs possess this capability, we develop the SelfReflect metric, an information-theoretic distance between a given summary and a distribution over answers. In interventional and human studies, we find that SelfReflect indicates even slight deviations, yielding a fine measure of faithfulness between a summary string and an LLM's actual internal distribution over answers. With SelfReflect, we make a resounding negative observation: modern LLMs are, across the board, incapable of revealing what they are uncertain about, neither through reasoning, nor chains-of-thoughts, nor explicit finetuning. However, we do find that LLMs are able to generate faithful summaries of their uncertainties if we help them by sampling multiple outputs and feeding them back into the context. This simple approach shines a light at the universal way of communicating LLM uncertainties whose future development the SelfReflect score enables. To support the development of this universal form of LLM uncertainties, we publish the code that implements our metric for arbitrary LLMs under https://github.com/apple/ml-selfreflect .

When using an LLM through an API provider, users expect the served model to remain consistent over time, a property crucial for the reliability of downstream applications and the reproducibility of research. Existing audit methods are too costly to apply at regular time intervals to the wide range of available LLM APIs. This means that model updates are left largely unmonitored in practice. In this work, we show that while LLM log probabilities (logprobs) are usually non-deterministic, they can still be used as the basis for cost-effective continuous monitoring of LLM APIs. We apply a simple statistical test based on the average value of each token logprob, requesting only a single token of output. This is enough to detect changes as small as one step of fine-tuning, making this approach more sensitive than existing methods while being 1,000x cheaper. We introduce the TinyChange benchmark as a way to measure the sensitivity of audit methods in the context of small, realistic model changes.

Large Language Models (LLMs) as judges and LLM-based data synthesis have emerged as two fundamental LLM-driven data annotation methods in model development. While their combination significantly enhances the efficiency of model training and evaluation, little attention has been given to the potential contamination brought by this new model development paradigm. In this work, we expose preference leakage, a contamination problem in LLM-as-a-judge caused by the relatedness between the synthetic data generators and LLM-based evaluators. To study this issue, we first define three common relatednesses between the data generator LLM and the judge LLM: being the same model, having an inheritance relationship, and belonging to the same model family. Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests that preference leakage is a pervasive and real-world problem that is harder to detect compared to previously identified biases in LLM-as-a-judge scenarios. All of these findings imply that preference leakage is a widespread and challenging problem in the area of LLM-as-a-judge.

Classical intuitions cast minimizing reverse KL as "mode seeking" and forward KL as "mass covering". In KL-regularized reinforcement learning, however, the regularizer determines _both_ the target distribution's shape _and_ the divergence being implicitly minimized, making its role more nuanced than simply inducing generic mode-seeking or mass-covering behaviour. Specifically, the target distribution is defined jointly by the reward function, the reference model, the type of regularizer, and the regularization strength. We show that under common settings—such as low regularization strength and equal verifiable rewards—both forward and reverse KL regularization tend to specify target distributions whose mass concentrates on a single high-reward region. Thus, the objective itself _by construction_ induces diversity collapse, regardless of the policy optimization algorithm used. Building on this perspective, we introduce a simple and scalable modification that rescales rewards to induce target distributions assigning substantial probability across _all_ high-reward regions. This yields a principled objective that maintains high solution quality while achieving broad reward-mode coverage. Empirically, this approach improves post-training diversity and performance for Large Language Models and Chemical Language Models, and is effective with either forward or reverse KL regularization, while using either naively fails.

We introduce MCP-Bench, a benchmark for evaluating large language models (LLMs) on realistic, multi-step tasks that demand tool use, cross-tool coordination, precise parameter control, and planning/reasoning for solving tasks. Built on the Model Context Protocol (MCP), MCP-Bench connects LLMs to 28 representative live MCP servers spanning 250 tools across domains such as finance, traveling, scientific computing, and academic search. Unlike prior API-based benchmarks, each MCP server provides a set of complementary tools designed to work together, enabling the construction of authentic, multi-step tasks with rich input–output coupling. Also, tasks in MCP-Bench test agents’ ability to retrieve relevant tools from fuzzy instructions without explicit tool names, plan multi-hop execution trajectories for complex objectives, ground responses in intermediate tool outputs, and orchestrate cross-domain workflows—capabilities not adequately evaluated by existing benchmarks that rely on explicit tool specifications, shallow few-step workflows, and isolated domain operations. We propose a multi-faceted evaluation framework covering tool-level schema understanding and usage, trajectorylevel planning and task completion. Experiments on 20 advanced LLMs reveal persistent challenges in MCP-Bench.

Cancer patients are increasingly turning to large language models (LLMs) for medical information, making it critical to assess how well these models handle complex, personalized questions. However, current medical benchmarks focus on medical exams or consumer-searched questions and do not evaluate LLMs on real patient questions with patient details. In this paper, we first have three hematology-oncology physicians evaluate cancer-related questions drawn from real patients. While LLM responses are generally accurate, the models frequently fail to recognize or address false presuppositions} in the questions, posing risks to safe medical decision-making. To study this limitation systematically, we introduce Cancer-Myth, an expert-verified adversarial dataset of 585 cancer-related questions with false presuppositions. On this benchmark, no frontier LLM---including GPT-5, Gemini-2.5-Pro, and Claude-4-Sonnet---corrects these false presuppositions more than $43\%$ of the time. To study mitigation strategies, we further construct a 150-question Cancer-Myth-NFP set, in which physicians confirm the absence of false presuppositions. We find typical mitigation strategies, such as adding precautionary prompts with GEPA optimization, can raise accuracy on Cancer-Myth to $80\%$, but at the cost of misidentifying presuppositions in $41\%$ of Cancer-Myth-NFP questions and causing a $10\%$ relative performance drop on other medical benchmarks. These findings highlight a critical gap in the reliability of LLMs, show that prompting alone is not a reliable remedy for false presuppositions, and underscore the need for more robust safeguards in medical AI systems.

Neural network pruning is a promising technique to mitigate the excessive computational and memory requirements of large language models (LLMs). Despite its promise, however, progress in this area has diminished, as conventional methods are seemingly unable to surpass moderate sparsity levels (50-60\%) without severely degrading model accuracy. This work breaks through the current impasse, presenting a principled and effective method called $ \text{Elsa}$, which achieves extreme sparsity levels of up to 90\% while retaining high model fidelity. This is done by identifying several limitations in current practice, all of which can be traced back to their reliance on a surrogate objective formulation. $ \text{Elsa}$ tackles this issue directly and effectively via standard and well-established constrained optimization techniques based on ADMM. Our extensive experiments across a wide range of models and scales show that $ \text{Elsa}$ achieves substantial improvements over existing methods; e.g., it achieves 7.8$ \times$ less perplexity than the best existing method on LLaMA-2-7B at 90\% sparsity. Moreover, we show that $ \text{Elsa}$ remains stable even at extreme sparsity (e.g., 95\%), yielding up to $\times$3.98 inference speedup and $\times$7.80 memory compression over its dense counterpart. We also present $ \text{Elsa}_ {-L}$, a quantized variant that scales to extremely large models (27B), and establish its theoretical convergence guarantees. These results highlight meaningful progress in advancing the frontier of LLM sparsity, while promising that significant opportunities for further advancement may remain in directions that have so far attracted limited exploration.

Frontier model developers aim to train models continually to possess emergent, diverse capabilities. To extend capabilities, the current pre-training and post-training paradigm requires manually starting training runs with static datasets or reward functions every time. Addressing this limitation, our work pursues the insight that open-endedness (via the coevolution of models and tasks) can discover models with increasingly novel skills in a single run. We introduce a new model development framework that extends coevolution to large language model (LLM) discovery, open-ended \textit{Assessment Coevolving with Diverse Capabilities} (AC/DC). AC/DC evolves both LLMs via model merging and natural language tasks via synthetic data generation. AC/DC discovers growing archives of LLMs that surpass the capabilities of larger LLMs while taking up less GPU memory. In particular, our LLM populations achieve a broader Coverage of expertise than other curated models or baselines on downstream benchmarks, without \textit{any} explicit benchmark optimization. Furthermore, AC/DC improves Coverage over time, continually innovates on tasks and models, and improves performance in multi-agent best-of-N selection. Our findings highlight the potential of coevolution as a means of discovering broader sets of capabilities from base LLMs. Overall, AC/DC brings us one step closer to a profoundly new paradigm of LLM development, where continual improvements to the diversity of model capabilities can be accelerated by leveraging existing models as stepping stones to increasingly powerful models.

Large language models show great potential in unstructured data understanding, but still face significant challenges with graphs due to their structural hallucination. Existing approaches mainly either verbalize graphs into natural language, which leads to excessive token consumption and scattered attention, or transform graphs into trainable continuous embeddings (i.e., soft prompt), but exhibit severe misalignment with original text tokens. To solve this problem, we propose to incorporate one special token <SO$G_k$> to fully represent the \textbf{\underline{S}}tructure \textbf{\underline{O}}f \textbf{\underline{G}}raph within a unified token space, facilitating explicit topology input and structural information sharing. Specifically, we propose a topology-aware structural tokenizer that maps each graph topology into a highly selective single token. Afterwards, we construct a set of hybrid structure Question-Answering corpora to align new structural tokens with existing text tokens. With this approach, <SO$G_k$> empowers LLMs to understand, generate, and reason in a concise and accurate manner. Extensive experiments on five graph-level benchmarks demonstrate the superiority of our method, achieving a performance improvement of 9.9–41.4\% compared to the baselines while exhibiting interpretability and consistency. Furthermore, our method provides a flexible extension to node-level tasks, enabling both global and local structural understanding. The codebase is publicly available\footnote{The code of our project is available at \href{https://anonymous.4open.science/r/SOG-8432}{https://anonymous.4open.science/r/SOG-8432}.}.

Minimizing the inference cost and latency of foundation models has become a crucial area of research. Optimization approaches include theoretically lossless methods and others without accuracy guarantees like quantization. In all of these cases it is crucial to ensure that the model quality has not degraded. However, even at temperature zero, model generations are not necessarily robust even to theoretically lossless model optimizations due to numerical errors. We thus require statistical tools to decide whether a finite-sample accuracy deviation is an evidence of a model's degradation or whether it can be attributed to (harmless) noise in the evaluation. We propose a statistically sound hypothesis testing framework based on McNemar's test allowing to efficiently detect model degradations, while guaranteeing a controlled rate of false positives. The crucial insight is that we have to confront the model scores on each sample, rather than aggregated on the task level. Furthermore, we propose three approaches to aggregate accuracy estimates across multiple benchmarks into a single decision. We provide an implementation on top of the largely adopted open source LM Evaluation Harness and provide a case study illustrating that the method correctly flags degraded models, while not flagging model optimizations that are provably lossless. We find that with our tests even empirical accuracy degradations of 0.3% can be confidently attributed to actual degradations rather than noise. Code: https://github.com/amazon-science/LLM-Accuracy-Stats

Large Language Models (LLMs) using Chain-of-Thought (CoT) prompting excel at complex reasoning but generate verbose thought processes with considerable redundancy, leading to increased inference costs and reduced efficiency. We introduce a novel CoT compression framework based on step entropy, a metric that quantifies the informational contribution of individual reasoning steps to identify redundancy. Through theoretical analysis and extensive empirical validation on mathematical reasoning benchmarks, we demonstrate that steps with low entropy are indeed highly redundant. Our experiments reveal that an astonishing 80% of low-entropy intermediate steps can be pruned without significant degradation in the final answer accuracy across DeepSeek-R1-7B, 14B and Qwen3-8B. This finding sharply contrasts with random or high-entropy pruning, which severely impairs reasoning performance. Building on this, we propose a novel two-stage training strategy combining Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) reinforcement learning. This approach enables LLMs to autonomously learn to generate compressed COTs during inference by strategically incorporating [SKIP] tokens. Our method significantly enhances LLM inference efficiency while rigorously preserving accuracy, offering profound implications for practical LLM deployment and a deeper understanding of reasoning structures.

Unlocking advanced reasoning in large language model agents is hindered by a scarcity of training data situated at the very frontier of their capabilities. We address this with a novel data synthesis approach inspired by the educational theory of the Zone of Proximal Development (ZPD), which conceptualizes this frontier as tasks an LLM cannot solve independently but can master with guidance. We operationalize this principle through the AgentFrontier Data Engine, an automated pipeline that synthesizes high-quality, multidisciplinary data situated precisely within an LLM's ZPD. The engine yields two synergistic outputs: knowledge-intensive data for continued pre-training and frontier-level reasoning trajectories for post-training. Concurrently, it produces the ZPD Exam, a self-evolving benchmark for evaluating agent capabilities by compelling them to reason beyond their parameterized knowledge. By training our AgentFrontier-30B-A3B model on the synthesized data, we achieve state-of-the-art results on demanding benchmarks like Humanity's Last Exam, outperforming several leading proprietary agents. This work establishes ZPD-guided data synthesis as a scalable and effective paradigm for cultivating increasingly capable LLM agents.