ICLR 2025 论文列表

Recent advances in Code Large Language Models (CodeLLMs) have primarily focused on open-ended code generation, often overlooking the crucial aspect of code understanding & reasoning. To bridge this gap, we introduce CodeMMLU, a comprehensive multiple-choice benchmark designed to evaluate the depth of software and code comprehension in LLMs. CodeMMLU includes nearly 20,000 questions spanning diverse domains, including code analysis, defect detection, and software engineering principles across multiple programming languages. Unlike traditional benchmarks that emphasize code generation, CodeMMLU assesses a model’s ability to reason about programs across a wide-range of tasks such as code repair, execution reasoning, and fill-in-the-blank challenges. Our extensive evaluation reveals that even state-of-the-art models struggle with CodeMMLU, highlighting significant gaps in comprehension beyond generation. By emphasizing the essential connection between code understanding and effective AI-assisted development, CodeMMLU provides a critical resource for advancing more reliable and capable coding assistants.

Math Word Problems (MWPs) play a vital role in assessing the capabilities of Large Language Models (LLMs), yet current research primarily focuses on questions with concise contexts. The impact of longer contexts on mathematical reasoning remains under-explored. This study pioneers the investigation of Context Length Generalizability (CoLeG), which refers to the ability of LLMs to solve MWPs with extended narratives. We introduce Extended Grade-School Math (E-GSM), a collection of MWPs featuring lengthy narratives, and propose two novel metrics to evaluate the efficacy and resilience of LLMs in tackling these problems. Our analysis of existing zero-shot prompting techniques with proprietary LLMs along with open-source LLMs reveals a general deficiency in CoLeG. To alleviate these issues, we propose tailored approaches for different categories of LLMs. For proprietary LLMs, we introduce a new instructional prompt designed to mitigate the impact of long contexts. For open-source LLMs, we develop a novel auxiliary task for fine-tuning to enhance CoLeG. Our comprehensive results demonstrate the effectiveness of our proposed methods, showing improved performance on E-GSM. Additionally, we conduct an in-depth analysis to differentiate the effects of semantic understanding and reasoning efficacy, showing that our methods improves the latter. We also establish the generalizability of our methods across several other MWP benchmarks. Our findings highlight the limitations of current LLMs and offer practical solutions correspondingly, paving the way for further exploration of model generalizability and training methodologies.

First-order logic (FOL) reasoning, which involves sequential deduction, is pivotal for intelligent systems and serves as a valuable task for evaluating reasoning capabilities, particularly in chain-of-thought (CoT) contexts. Existing benchmarks often rely on extensive human annotation or handcrafted templates, making it difficult to achieve the necessary complexity, scalability, and diversity for robust evaluation. To address these limitations, we propose a novel framework called ProverGen that synergizes the generative strengths of Large Language Models (LLMs) with the rigor and precision of symbolic provers, enabling the creation of a scalable, diverse, and high-quality FOL reasoning dataset, ProverQA. ProverQA is also distinguished by its inclusion of accessible and logically coherent intermediate reasoning steps for each problem. Our evaluation shows that state-of-the-art LLMs struggle to solve ProverQA problems, even with CoT prompting, highlighting the dataset's challenging nature. We also finetune Llama3.1-8B-Instruct on a separate training set generated by our framework. The finetuned model demonstrates consistent improvements on both in-distribution and out-of-distribution test sets, suggesting the value of our proposed data generation framework. Code available at: \url{https://github.com/opendatalab/ProverGen}

The success of most existing cross-modal retrieval methods heavily relies on the assumption that the given queries follow the same distribution of the source domain. However, such an assumption is easily violated in real-world scenarios due to the complexity and diversity of queries, thus leading to the query shift problem. Specifically, query shift refers to the online query stream originating from the domain that follows a different distribution with the source one. In this paper, we observe that query shift would not only diminish the uniformity (namely, within-modality scatter) of the query modality but also amplify the gap between query and gallery modalities. Based on the observations, we propose a novel method dubbed Test-time adaptation for Cross-modal Retrieval (TCR). In brief, TCR employs a novel module to refine the query predictions (namely, retrieval results of the query) and a joint objective to prevent query shift from disturbing the common space, thus achieving online adaptation for the cross-modal retrieval models with query shift. Expensive experiments demonstrate the effectiveness of the proposed TCR against query shift. Code is available at https://github.com/XLearning-SCU/2025-ICLR-TCR.

Large language models (LLMs) can solve an increasing number of complex reasoning tasks while making surprising mistakes in basic numerical understanding and processing (such as $9.11 > 9.9$). The latter ability is essential for tackling complex arithmetic and mathematical problems and serves as a foundation for most reasoning tasks, but previous work paid little attention to it or only discussed several restricted tasks (like integer addition). In this paper, we comprehensively investigate the numerical understanding and processing ability (NUPA) of LLMs. Firstly, we introduce a benchmark covering four common numerical representations and 17 distinct numerical tasks in four major categories, resulting in 41 meaningful combinations in total. These tasks are derived from primary and secondary education curricula, encompassing nearly all everyday numerical understanding and processing scenarios, and the rules of these tasks are very simple and clear. Through the benchmark, we find that current LLMs fail frequently in many of the tasks. To study the problem, we train small models with existing and potential techniques for enhancing NUPA (such as tokenizers, PEs, and number formats), comprehensively evaluating their effectiveness using our testbed. We also finetune practical-scale LLMs on our proposed NUPA tasks and find that 1) naive finetuning can improve NUPA a lot on many but not all tasks, and 2) surprisingly, techniques designed to enhance NUPA prove ineffective for finetuning pretrained models. We further explore the impact of chain-of-thought techniques on NUPA. Our work provides a more detailed and comprehensive understanding of NUPA in LLMs.

Software engineers mainly write code by editing existing programs. In contrast, language models (LMs) autoregressively synthesize programs in a single pass. One explanation for this is the scarcity of sequential edit data. While high-quality instruction data for code synthesis is scarce, edit data for synthesis is even scarcer. To fill this gap, we develop a synthetic data generation algorithm called LintSeq. This algorithm refactors programs into sequences of synthetic edits by using a linter to procedurally sample across interdependent lines of source code. Synthetic edits sampled with LintSeq reflect the syntax and semantics of their programming language. To test the algorithm, we use it to refactor a dataset of instruction + program pairs into instruction + program-diff-sequence tuples. Then, we fine-tune a series of smaller LMs ranging from 2.6B to 14B parameters on both the re-factored and original versions of this dataset. We perform comprehensive evaluations comparing edit sequence code LMs against baselines on HumanEval, MBPP(+), CodeContests, DS-1000, and BigCodeBench. We show that models fine-tuned to iteratively synthesize code match or outperform baselines on pass@1, and exhibit better scaling across higher pass@k as a function of total test-time FLOPs. Finally, we also pretrain our own tiny LMs for code understanding. We show that fine-tuning these models to synthesize code edit-by-edit results in strong performance on HumanEval and MBPP(+) compared to existing code language models of similar scale such as CodeT5+, AlphaCode, and Codex.

Chain-of-thought (CoT) significantly enhances the reasoning performance of large language models (LLM). While current theoretical studies often attribute this improvement to increased expressiveness and computational capacity, we argue that expressiveness is not the primary limitation in the LLM regime, as current large models will fail on simple tasks. Using a parity-learning setup, we demonstrate that CoT can substantially improve sample efficiency even when the representation power is sufficient. Specifically, with CoT, a transformer can learn the function within polynomial samples, whereas without CoT, the required sample size is exponential. Additionally, we show that CoT simplifies the learning process by introducing sparse sequential dependencies among input tokens, and leads to a sparse and interpretable attention. We validate our theoretical analysis with both synthetic and real-world experiments, confirming that sparsity in attention layers is a key factor of the improvement induced by CoT.

Recent advancements in Large Language Models (LLMs) have sparked interest in their mathematical reasoning capabilities. While performance on the widely popular GSM8K benchmark has improved, questions remain about whether reported evaluation metrics are reliable, and reasoning abilities of LLMs have advanced. To overcome the limitations of existing evaluations, we introduce GSM-Symbolic, an improved benchmark created from symbolic templates that allow for the generation of a diverse set of questions. GSM-Symbolic enables more controllable evaluations, providing key insights and more reliable metrics for measuring the reasoning capabilities of models. Our findings reveal that LLMs exhibit noticeable variance when responding to different instantiations of the same question. Specifically, the performance of models declines when only the numerical values in the question are altered in the GSM-Symbolic benchmark. Furthermore, we investigate the fragility of mathematical reasoning in these models and demonstrate that their performance significantly deteriorates as the number of clauses in a question increases. We hypothesize that this decline is due to the fact that current LLMs are not capable of genuine logical reasoning; instead, they attempt to replicate the reasoning steps observed in their training data. When we add a single clause that appears relevant to the question, we observe significant performance drops (up to 65%) across all state-of-the-art models, even though the added clause does not contribute to the reasoning chain needed to reach the final answer. Overall, our work provides a more nuanced understanding of LLMs' capabilities and limitations in mathematical reasoning.

The potential of large language models (LLMs) as decision support tools is increasingly being explored in fields such as business, engineering, and medicine, which often face challenging tasks of *decision-making under uncertainty*. In this paper, we show that directly prompting LLMs on these types of decision-making problems can yield poor results, especially as the problem complexity increases. To aid in these tasks, we propose DeLLMa (Decision-making Large Language Model assistant), a framework designed to enhance decision-making accuracy in uncertain environments. DeLLMa involves a multi-step reasoning procedure that integrates recent best practices in scaling *inference-time reasoning*, drawing upon principles from decision theory and utility theory, to provide an accurate and human-auditable decision-making process. We validate our procedure on multiple realistic decision-making environments, demonstrating that DeLLMa can consistently enhance the decision-making performance of leading language models, and achieve up to a 40% increase in accuracy over competing methods. Additionally, we show how performance improves when scaling compute at test time, and carry out human evaluations to benchmark components of DeLLMa.

Multimodal Large Language Models (MLLMs) have emerged as a central focus in both industry and academia, but often suffer from biases introduced by visual and language priors, which can lead to multimodal hallucination. These biases arise from the visual encoder and the Large Language Model (LLM) backbone, affecting the attention mechanism responsible for aligning multimodal inputs. Existing decoding-based mitigation methods focus on statistical correlations and overlook the causal relationships between attention mechanisms and model output, limiting their effectiveness in addressing these biases. To tackle this issue, we propose a causal inference framework termed CausalMM that applies structural causal modeling to MLLMs, treating modality priors as a confounder between attention mechanisms and output. Specifically, by employing backdoor adjustment and counterfactual reasoning at both the visual and language attention levels, our method mitigates the negative effects of modality priors and enhances the alignment of MLLM's inputs and outputs, with a maximum score improvement of 65.3% on 6 VLind-Bench indicators and 164 points on MME Benchmark compared to conventional methods. Extensive experiments validate the effectiveness of our approach while being a plug-and-play solution. Our code is available at: https://github.com/The-Martyr/CausalMM.

Language models are increasingly being used in important decision pipelines, so ensuring the correctness of their outputs is crucial. Recent work has proposed evaluating the “factuality” of claims decomposed from a language model generation and applying conformal prediction techniques to filter out those claims that are not factual. This can be effective for tasks such as information retrieval, where constituent claims may be evaluated in isolation for factuality, but is not appropriate for reasoning tasks, as steps of a logical argument can be evaluated for correctness only within the context of the claims that precede them. To capture this, we define “coherent factuality” and develop a conformal-prediction-based method to guarantee coherent factuality for language model outputs. Our approach applies split conformal prediction to subgraphs within a "deducibility" graph that represents the steps of a reasoning problem. We evaluate our method on mathematical reasoning problems from the MATH and FELM datasets and find that our algorithm consistently produces correct and substantiated orderings of claims, achieving coherent factuality across target coverage levels. Moreover, we achieve 90\% factuality on our stricter definition while retaining 80\% or more of the original claims, highlighting the utility of our deducibility-graph-guided approach.

A promising approach for improving reasoning in large language models is to use process reward models (PRMs). PRMs provide feedback at each step of a multi-step reasoning trace, improving credit assignment over outcome reward models (ORMs) that only provide feedback at the final step. However, collecting dense, per-step human labels is not scalable, and training PRMs from automatically-labeled data has thus far led to limited gains. With the goal of using PRMs to improve a *base* policy via test-time search and reinforcement learning (RL), we ask: ``How should we design process rewards?'' Our key insight is that, to be effective, the process reward for a step should measure *progress*: a change in the likelihood of producing a correct response in the future, before and after taking the step, as measured under a *prover* policy distinct from the base policy. Such progress values can {distinguish} good and bad steps generated by the base policy, even though the base policy itself cannot. Theoretically, we show that even weaker provers can improve the base policy, as long as they distinguish steps without being too misaligned with the base policy. Our results show that process rewards defined as progress under such provers improve the efficiency of exploration during test-time search and online RL. We empirically validate our claims by training **process advantage verifiers (PAVs)** to measure progress under such provers and show that compared to ORM, they are >8% more accurate, and 1.5-5x more compute-efficient. Equipped with these insights, our PAVs enable **one of the first results** showing a 6x gain in sample efficiency for a policy trained using online RL with PRMs vs. ORMs.

Search is an ability foundational in many important tasks, and recent studies have shown that large language models (LLMs) struggle to perform search robustly. It is unknown whether this inability is due to a lack of data, insufficient model parameters, or fundamental limitations of the transformer architecture. In this work, we use the foundational graph connectivity problem as a testbed to generate effectively limitless high-coverage data to train small transformers and test whether they can learn to perform search. We find that, when given the right training distribution, the transformer is able to learn to search. We analyze the algorithm that the transformer has learned through a novel mechanistic interpretability technique that enables us to extract the computation graph from the trained model. We find that for each vertex in the input graph, transformers compute the set of vertices reachable from that vertex. Each layer then progressively expands these sets, allowing the model to search over a number of vertices exponential in the number of layers. However, we find that as the input graph size increases, the transformer has greater difficulty in learning the task. This difficulty is not resolved even as the number of parameters is increased, suggesting that increasing model scale will not lead to robust search abilities. We also find that performing search in-context (i.e., chain-of-thought) does not resolve this inability to learn to search on larger graphs.

The Mixture-of-Experts (MoE) architecture enables a significant increase in the total number of model parameters with minimal computational overhead. However, it is not clear what performance tradeoffs, if any, exist between MoEs and standard dense transformers. In this paper, we show that as we increase the number of experts (while fixing the number of active parameters), the memorization performance consistently increases while the reasoning capabilities saturate. We begin by analyzing the theoretical limitations of MoEs at reasoning. We prove that there exist graph problems that cannot be solved by any number of experts of a certain width; however, the same task can be easily solved by a dense model with a slightly larger width. On the other hand, we find that on memory-intensive tasks, MoEs can effectively leverage a small number of active parameters with a large number of experts to memorize the data. We empirically validate these findings on synthetic graph problems and memory-intensive closed book retrieval tasks. Lastly, we pre-train a series of MoEs and dense transformers and evaluate them on commonly used benchmarks in math and natural language. We find that increasing the number of experts helps solve knowledge-intensive tasks, but fails to yield the same benefits for reasoning tasks.

Real-time monitoring in modern medical research introduces functional longitudinal data, characterized by continuous-time measurements of outcomes, treatments, and confounders. This complexity leads to uncountably infinite treatment-confounder feedbacks, which traditional causal inference methodologies cannot handle. Inspired by the coarsened data framework, we adopt stochastic process theory, measure theory, and net convergence to propose a nonparametric causal identification framework. This framework generalizes classical g-computation, inverse probability weighting, and doubly robust formulas, accommodating time-varying outcomes subject to mortality and censoring for functional longitudinal data. We examine our framework through Monte Carlo simulations. Our approach addresses significant gaps in current methodologies, providing a solution for functional longitudinal data and paving the way for future estimation work in this domain.

Spatial expressions in situated communication can be ambiguous, as their meanings vary depending on the frames of reference (FoR) adopted by speakers and listeners. While spatial language understanding and reasoning by vision-language models (VLMs) have gained increasing attention, potential ambiguities in these models are still under-explored. To address this issue, we present the COnsistent Multilingual Frame Of Reference Test (COMFORT), an evaluation protocol to systematically assess the spatial reasoning capabilities of VLMs. We evaluate nine state-of-the-art VLMs using COMFORT. Despite showing some alignment with English conventions in resolving ambiguities, our experiments reveal significant shortcomings of VLMs: notably, the models (1) exhibit poor robustness and consistency, (2) lack the flexibility to accommodate multiple FoRs, and (3) fail to adhere to language-specific or culture-specific conventions in cross-lingual tests, as English tends to dominate other languages. With a growing effort to align vision-language models with human cognitive intuitions, we call for more attention to the ambiguous nature and cross-cultural diversity of spatial reasoning.

General-purpose LLM judges capable of human-level evaluation provide not only a scalable and accurate way of evaluating instruction-following LLMs but also new avenues for supervising and improving their performance. One promising way of leveraging LLM judges for supervision is through Minimum Bayes Risk (MBR) decoding, which uses a reference-based evaluator to select a high-quality output from amongst a set of candidate outputs. In the first part of this work, we explore using MBR decoding as a method for improving the test-time performance of instruction-following LLMs. We find that MBR decoding with reference-based LLM judges substantially improves over greedy decoding, best-of-N decoding with reference-free judges and MBR decoding with lexical and embedding-based metrics on AlpacaEval and MT-Bench. These gains are consistent across LLMs with up to 70B parameters, demonstrating that smaller LLM judges can be used to supervise much larger LLMs. Then, seeking to retain the improvements from MBR decoding while mitigating additional test-time costs, we explore iterative self-training on MBR-decoded outputs. We find that self-training using Direct Preference Optimisation leads to significant performance gains, such that the self-trained models with greedy decoding generally match and sometimes exceed the performance of their base models with MBR decoding.

Scaling language models to larger and deeper sizes has led to significant boosts in performance. Even though the size of these models limits their application in compute-constrained environments, the race to continually develop ever larger and deeper foundational models is underway. At the same time---regardless of the model size---task-specific techniques continue to play a pivotal role in achieving optimal downstream performance. One of these techniques, called Chain-of-Thought (CoT), is particularly interesting since, as we point out in this work, it resembles employing a deeper transformer through re-applying the model multiple times. However, a key subtlety in computing the attention of past tokens differentiates CoT from simply applying the model several times. Based on this insight, we propose CoTFormer, a novel architecture which closely mimics CoT at the token level, allowing us to obtain significantly improved accuracies close to much larger models. While applying CoT introduces additional computation costs, we compensate for it by leveraging CoTFormer's special compatibility with token-wise variable depth. Through a compute adaptive model---which automatically allocates the compute to tokens that need it most---we show that it is possible to reduce the computation cost significantly without any reduction in accuracy, and with further compute cost reductions possible while maintaining a competitive accuracy.

Current large language models (LLMs) are far from reliable: they are prone to generate non-factual information and, more crucially, to contradict themselves when prompted to reason about relations between real entities of the world. These problems are currently addressed with large scale fine-tuning or by delegating consistent reasoning to external tools. In this work, we strive for a middle ground and leverage a training objective based on a principled neuro-symbolic loss that teaches a LLM to be consistent with external knowledge in the form of a set of facts and rules. Fine-tuning with such a loss on a limited set of facts enables our LLMs to be more logically consistent than previous baselines for a given constraint. Our approach also allows to easily combine multiple logical constraints at once in a principled way, delivering LLMs that are more consistent w.r.t. all the selected rules. Moreover, our method allows LLMs to extrapolate to unseen but semantically similar factual knowledge, represented in unseen datasets, more systematically.