ICLR 2025 Papers

Test-time adaptation (TTA) updates the model weights during the inference stage using testing data to enhance generalization. However, this practice exposes TTA to adversarial risks. Existing studies have shown that when TTA is updated with crafted adversarial test samples, also known as test-time poisoned data, the performance on benign samples can deteriorate. Nonetheless, the perceived adversarial risk may be overstated if the poisoned data is generated under overly strong assumptions. In this work, we first review realistic assumptions for test-time data poisoning, including white-box versus grey-box attacks, access to benign data, attack order, and more. We then propose an effective and realistic attack method that better produces poisoned samples without access to benign samples, and derive an effective in-distribution attack objective. We also design two TTA-aware attack objectives. Our benchmarks of existing attack methods reveal that the TTA methods are more robust than previously believed. In addition, we analyze effective defense strategies to help develop adversarially robust TTA methods. The source code is available at https://github.com/Gorilla-Lab-SCUT/RTTDP.

Episodic memory -- the ability to recall specific events grounded in time and space -- is a cornerstone of human cognition, enabling not only coherent storytelling, but also planning and decision-making. Despite their remarkable capabilities, Large Language Models (LLMs) lack a robust mechanism for episodic memory: we argue that integrating episodic memory capabilities into LLM is essential for advancing AI towards human-like cognition, increasing their potential to reason consistently and ground their output in real-world episodic events, hence avoiding confabulations. To address this challenge, we introduce a comprehensive framework to model and evaluate LLM episodic memory capabilities. Drawing inspiration from cognitive science, we develop a structured approach to represent episodic events, encapsulating temporal and spatial contexts, involved entities, and detailed descriptions. We synthesize a unique episodic memory benchmark, free from contamination, and release open source code and datasets to assess LLM performance across various recall and episodic reasoning tasks. Our evaluation of state-of-the-art models, including GPT-4 and Claude variants, Llama 3.1, and o1-mini, reveals that even the most advanced LLMs struggle with episodic memory tasks, particularly when dealing with multiple related events or complex spatio-temporal relationships -- even in contexts as short as 10k-100k tokens.

Computationally intensive decoding procedures---including search, reranking, and self-critique---can improve the quality of language model (LM) outputs in problems spanning code generation, numerical reasoning, and dialog. Existing work typically applies the same decoding procedure for every input to an LM. But not all inputs require the same amount of computation to process. Can we allocate decoding computation adaptively, using more resources to answer questions whose answers will be harder to compute? We present an approach that predicts the distribution of rewards given an input and computation budget, then allocates additional computation to inputs for which it is predicted to be most useful. We apply this approach in two decoding procedures: first, an adaptive best-of-$k$ procedure that dynamically selects the number of samples to generate as input to a reranker; second, a routing procedure that dynamically responds to a query using a decoding procedure that is expensive but accurate, or one that is cheaper but less capable. Across a suite of programming, mathematics, and dialog tasks, we show that accurate computation-allocation procedures can be learned, and reduce computation by up to 50% at no cost to quality.

This paper introduces rStar, a self-play mutual reasoning approach that significantly improves reasoning capabilities of small language models (SLMs) without fine-tuning or superior models. rStar decouples reasoning into a self-play mutual generation-discrimination process. First, a target SLM augments the Monte Carlo Tree Search (MCTS) with a rich set of human-like reasoning actions to construct higher quality reasoning trajectories. Next, another SLM, with capabilities similar to the target SLM, acts as a discriminator to verify each trajectory generated by the target SLM. The mutually agreed reasoning trajectories are considered mutual consistent, thus are more likely to be correct. Extensive experiments across five SLMs demonstrate rStar can effectively solve diverse reasoning problems, including GSM8K, GSM-Hard, MATH, SVAMP, and StrategyQA. Remarkably, rStar boosts GSM8K accuracy from 12.51\% to 63.91\% for LLaMA2-7B, from 36.46\% to 81.88\% for Mistral-7B, from 74.53\% to 91.13\% for LLaMA3-8B-Instruct. Code is available at https://github.com/zhentingqi/rStar.

Despite their proficiency in general tasks, Multi-modal Large Language Models (MLLMs) struggle with automatic Geometry Problem Solving (GPS), which demands understanding diagrams, interpreting symbols, and performing complex reasoning. This limitation arises from their pre-training on natural images and texts, along with the lack of automated verification in the problem-solving process. Besides, current geometric specialists are limited by their task-specific designs, making them less effective for broader geometric problems. To this end, we present GeoX, a multi-modal large model focusing on geometric understanding and reasoning tasks. Given the significant differences between geometric diagram-symbol and natural image-text, we introduce unimodal pre-training to develop a diagram encoder and symbol decoder, enhancing the understanding of geometric images and corpora. Furthermore, we introduce geometry-language alignment, an effective pre-training paradigm that bridges the modality gap between unimodal geometric experts. We propose a Generator-And-Sampler Transformer (GS-Former) to generate discriminative queries and eliminate uninformative representations from unevenly distributed geometric signals. Finally, GeoX benefits from visual instruction tuning, empowering it to take geometric images and questions as input and generate verifiable solutions. Experiments show that GeoX outperforms both generalists and geometric specialists on publicly recognized benchmarks, such as GeoQA, UniGeo, Geometry3K, and PGPS9k. Our data and code will be released soon to accelerate future research on automatic GPS.

Accurate nutrition estimation helps people make informed dietary choices and is essential in the prevention of serious health complications. We present NutriBench, the first publicly available natural language meal description nutrition benchmark. NutriBench consists of 11,857 meal descriptions generated from real-world global dietary intake data. The data is human-verified and annotated with macro-nutrient labels, including carbohydrates, proteins, fats, and calories. We conduct an extensive evaluation of Nutribench on the task of carbohydrate estimation, testing twelve leading Large Language Models (LLMs), including GPT-4o, Llama3.1, Qwen2, Gemma2, and OpenBioLLM models, using standard, Chain-of-Thought and Retrieval-Augmented Generation strategies. Additionally, we present a study involving professional nutritionists, finding that LLMs can provide comparable but significantly faster estimates. Finally, we perform a real-world risk assessment by simulating the effect of carbohydrate predictions on the blood glucose levels of individuals with type 1 diabetes. Our work highlights the opportunities and challenges of using LLMs for nutrition estimation, demonstrating their potential to aid professionals and laypersons and improve health outcomes. Our benchmark is publicly available at: https://mehak126.github.io/nutribench.html

With the rapid advancement of test-time compute search strategies to improve the mathematical problem-solving capabilities of large language models (LLMs), the need for building robust verifiers has become increasingly important. However, all these inference strategies rely on existing verifiers originally designed for Best-of-N search, which makes them sub-optimal for tree search techniques at test time. During tree search, existing verifiers can only offer indirect and implicit assessments of partial solutions or under-value prospective intermediate steps, thus resulting in the premature pruning of promising intermediate steps. To overcome these limitations, we propose token-supervised value models (TVMs) -- a new class of verifiers that assign each token a probability that reflects the likelihood of reaching the correct final answer. This new token-level supervision enables TVMs to directly and explicitly evaluate partial solutions, effectively distinguishing between promising and incorrect intermediate steps during tree search at test time. Experimental results demonstrate that combining tree-search-based inference strategies with TVMs significantly improves the accuracy of LLMs in mathematical problem-solving tasks, surpassing the performance of existing verifiers.

Humans can leverage both symbolic reasoning and intuitive responses. In contrast, reinforcement learning policies are typically encoded in either opaque systems like neural networks or symbolic systems that rely on predefined symbols and rules. This disjointed approach severely limits the agents’ capabilities, as they often lack either the flexible low-level reaction characteristic of neural agents or the interpretable reasoning of symbolic agents. To overcome this challenge, we introduce *BlendRL*, a neuro-symbolic RL framework that harmoniously integrates both paradigms. We empirically demonstrate that BlendRL agents outperform both neural and symbolic baselines in standard Atari environments, and showcase their robustness to environmental changes. Additionally, we analyze the interaction between neural and symbolic policies, illustrating how their hybrid use helps agents overcome each other's limitations.

Problems in fields such as healthcare, robotics, and finance requires reasoning about the value both of what decision or action to take and when to take it. The prevailing hope is that artificial intelligence will support such decisions by estimating the causal effect of policies such as how to treat patients or how to allocate resources over time. However, existing methods for estimating the effect of a policy struggle with \emph{irregular time}. They either discretize time, or disregard the effect of timing policies. We present a new deep-Q algorithm that estimates the effect of both when and what to do called Earliest Disagreement Q-Evaluation (EDQ). EDQ makes use of recursion for the Q-function that is compatible with flexible sequence models, such as transformers. EDQ provides accurate estimates under standard assumptions. We validate the approach through experiments on survival time and tumor growth tasks.

Deep learning models often struggle with generalization when deploying on real-world data, due to the common distributional shift to the training data. Test-time adaptation (TTA) is an emerging scheme used at inference time to address this issue. In TTA, models are adapted online at the same time when making predictions to test data. Neighbor-based approaches have gained attention recently, where prototype embeddings provide location information to alleviate the feature shift between training and testing data. However, due to their inherit limitation of simplicity, they often struggle to learn useful patterns and encounter performance degradation. To confront this challenge, we study the TTA problem from a geometric point of view. We first reveal that the underlying structure of neighbor-based methods aligns with the Voronoi Diagram, a classical computational geometry model for space partitioning. Building on this observation, we propose the Test-Time adjustment by Voronoi Diagram guidance (TTVD), a novel framework that leverages the benefits of this geometric property. Specifically, we explore two key structures: 1) Cluster-induced Voronoi Diagram (CIVD): This integrates the joint contribution of self-supervision and entropy-based methods to provide richer information. 2) Power Diagram (PD): A generalized version of the Voronoi Diagram that refines partitions by assigning weights to each Voronoi cell. Our experiments under rigid, peer-reviewed settings on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and ImageNet-R shows that TTVD achieves remarkable improvements compared to state-of-the-art methods. Moreover, extensive experimental results also explore the effects of batch size and class imbalance, which are two scenarios commonly encountered in real-world applications. These analyses further validate the robustness and adaptability of our proposed framework.

Large language models (LLMs) can solve arithmetic word problems with high accuracy, but little is known about how well they generalize to more complex problems. This is difficult to study, as (i) much of the available evaluation data has already been seen by the most capable models during training, and (ii) existing benchmarks do not capture how problem proofs may be arbitrarily complex in various ways. In this paper, we present a data-generation framework for evaluating LLMs on problems with arbitrarily complex arithmetic proofs, called MathGAP. MathGAP generates problem statements and chain-of-thought reasoning traces according to specifications about their arithmetic proof structure, enabling systematic studies on easy-to-hard generalization with respect to complexity of proof trees. Using MathGAP, we find that LLMs show a significant decrease in performance as proofs get deeper and wider. This effect is more pronounced in complex, nonlinear proof structures, which are challenging even for the most capable models. The models are also sensitive to simple changes in sentence ordering. However, they remain capable of solving some complex problems, suggesting that reasoning generalization is noisy.

While a lot of recent research focuses on enhancing the textual reasoning capabilities of Large Language Models (LLMs) by optimizing the multi-agent framework or reasoning chains, several benchmark tasks can be solved with 100\% success through direct coding, which is more scalable and avoids the computational overhead associated with textual iterating and searching. Textual reasoning has inherent limitations in solving tasks with challenges in math, logics, optimization, and searching, which is unlikely to be solved by simply scaling up the model and data size. The recently released OpenAI GPT Code Interpreter and multi-agent frameworks such as AutoGen have demonstrated remarkable proficiency of integrating code generation and execution to solve complex tasks using LLMs. However, based on our experiments on 7 existing popular methods for steering code/text generation in both single- and multi-turn settings with 14 tasks and 6 types of LLMs (including the new O1-preview), currently there is no optimal method to correctly steer LLMs to write code when needed. We discover some interesting patterns on when models use code vs. textual reasoning with the evolution to task complexity and model sizes, which even result in an astonishingly inverse scaling behavior. We also discover that results from LLM written code are not always better than using textual reasoning, even if the task could be solved through code. To mitigate the above issues, we propose three methods to better steer LLM code/text generation and achieve a notable improvement. The costs of token lengths and runtime are thoroughly discussed for all the methods. We believe the problem of steering LLM code/text generation is critical for future research and has much space for further improvement. Project Page, Datasets, and Codes are available at https://yongchao98.github.io/CodeSteer/.

The rapid increase in textual information means we need more efficient methods to sift through, organize, and understand it all. While retrieval-augmented generation (RAG) models excel in accessing information from large document collections, they struggle with complex tasks that require aggregation and reasoning over information spanning across multiple documents--what we call \textit{holistic reasoning}. Long-context language models (LCLMs) have great potential for managing large-scale documents, but their holistic reasoning capabilities remain unclear. In this work, we introduce HoloBench, a novel framework that brings database reasoning operations into text-based contexts, making it easier to systematically evaluate how LCLMs handle holistic reasoning across large documents. Our approach adjusts key factors such as context length, information density, distribution of information, and query complexity to evaluate LCLMs comprehensively. Our experiments show that the amount of information in the context has a bigger influence on LCLM performance than the actual context length. Furthermore, the complexity of queries affects performance more than the amount of information, particularly for different types of queries. Interestingly, queries that involve finding maximum or minimum values are easier for LCLMs and are less affected by context length, even though they pose challenges for RAG systems. However, tasks requiring the aggregation of multiple pieces of information show a noticeable drop in accuracy as context length increases. Additionally, we find that while grouping relevant information generally improves performance, the optimal positioning varies across models. Our findings surface both the advancements and the ongoing challenges in achieving a holistic understanding of long contexts. These can guide future developments in LCLMs and set the stage for creating more robust language models for real-world applications.

Machine learning models must continuously self-adjust themselves for novel data distribution in the open world. As the predominant principle, entropy minimization (EM) has been proven to be a simple yet effective cornerstone in existing test-time adaption (TTA) methods. While unfortunately its fatal limitation (i.e., overconfidence) tends to result in model collapse. For this issue, we propose to \textbf{\texttt{Co}}nservatively \textbf{\texttt{M}}inimize the \textbf{\texttt{E}}ntropy (\texttt{COME}), which is a simple drop-in replacement of traditional EM to elegantly address the limitation. In essence, \texttt{COME} explicitly models the uncertainty by characterizing a Dirichlet prior distribution over model predictions during TTA. By doing so, \texttt{COME} naturally regularizes the model to favor conservative confidence on unreliable samples. Theoretically, we provide a preliminary analysis to reveal the ability of \texttt{COME} in enhancing the optimization stability by introducing a data-adaptive lower bound on the entropy. Empirically, our method achieves state-of-the-art performance on commonly used benchmarks, showing significant improvements in terms of classification accuracy and uncertainty estimation under various settings including standard, life-long and open-world TTA, i.e., up to $34.5\%$ improvement on accuracy and $15.1\%$ on false positive rate. Our code is available at: \href{https://github.com/BlueWhaleLab/COME}{https://github.com/BlueWhaleLab/COME}.

Test-time adaptation updates pretrained models on the fly to handle distribution shifts in test data. While existing research has focused on stable optimization during adaptation, less attention has been given to enhancing model representations for adaptation capability. To address this gap, we propose Test-Time Ensemble (TTE) grounded in the intriguing property of linear mode connectivity. TTE leverages ensemble strategies during adaptation: 1) adaptively averaging the parameter weights of assorted test-time adapted models and 2) incorporating dropout to further promote representation diversity. These strategies encapsulate model diversity into a single model, avoiding computational burden associated with managing multiple models. Besides, we propose a robust knowledge distillation scheme to prevent model collapse, ensuring stable optimization and preserving the ensemble benefits during adaptation. Notably, TTE integrates seamlessly with existing TTA approaches, advancing their adaptation capabilities. In extensive experiments, integration with TTE consistently outperformed baseline models across various challenging scenarios, demonstrating its effectiveness and general applicability.

Large language models (LLMs) are capable of generating *plausible* explanations of how they arrived at an answer to a question. However, these explanations can misrepresent the model's "reasoning" process, i.e., they can be *unfaithful*. This, in turn, can lead to over-trust and misuse. We introduce a new approach for measuring the faithfulness of LLM explanations. First, we provide a rigorous definition of faithfulness. Since LLM explanations mimic human explanations, they often reference high-level *concepts* in the input question that purportedly influenced the model. We define faithfulness in terms of the difference between the set of concepts that the LLM's *explanations imply* are influential and the set that *truly* are. Second, we present a novel method for estimating faithfulness that is based on: (1) using an auxiliary LLM to modify the values of concepts within model inputs to create realistic counterfactuals, and (2) using a hierarchical Bayesian model to quantify the causal effects of concepts at both the example- and dataset-level. Our experiments show that our method can be used to quantify and discover interpretable patterns of unfaithfulness. On a social bias task, we uncover cases where LLM explanations hide the influence of social bias. On a medical question answering task, we uncover cases where LLM explanations provide misleading claims about which pieces of evidence influenced the model's decisions.

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n > 5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d = 28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.24% root mean square error (RMSE), (2) reconstruct the original total score with 0.58% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.94.

There has been considerable divergence of opinion on the reasoning abilities of Large Language Models (LLMs). While the initial optimism that reasoning might emerge automatically with scale has been tempered thanks to a slew of counterexamples--ranging from multiplication to simple planning--there persists a wide spread belief that LLMs can self-critique and improve their own solutions in an iterative fashion. This belief seemingly rests on the assumption that verification of correctness should be easier than generation--a rather classical argument from computational complexity--which should be irrelevant to LLMs to the extent that what they are doing is approximate retrieval. In this paper, we set out to systematically investigate the effectiveness of iterative prompting in the context of reasoning and planning. We present a principled empirical study of the performance of GPT-4 in three domains: Game of 24, Graph Coloring, and STRIPS planning. We experiment both with the model critiquing its own answers and with an external correct reasoner verifying proposed solutions. In each case, we analyze whether the content of criticisms actually affects bottom line performance, and whether we can ablate elements of the augmented system without losing performance. We observe significant performance collapse with self-critique and significant performance gains with sound external verification. We also note that merely re-prompting with a sound verifier maintains most of the benefits of more involved setups.

Enabling LLMs to improve their outputs by using more test-time compute is a critical step towards building self-improving agents that can operate on open-ended natural language. In this paper, we scale up inference-time computation in LLMs, with a focus on answering: if an LLM is allowed to use a fixed but non-trivial amount of inference-time compute, how much can it improve its performance on a challenging prompt? Answering this question has implications not only on performance, but also on the future of LLM pretraining and how to tradeoff inference-time and pre-training compute. Little research has attempted to understand the scaling behaviors of test-time inference methods, with current work largely providing negative results for a number of these strategies. In this work, we analyze two primary mechanisms to scale test-time computation: (1) searching against dense, process-based verifier reward models (PRMs); and (2) updating the model's distribution over a response adaptively, given the prompt at test time. We find that in both cases, the effectiveness of different approaches to scaling test-time compute critically varies depending on the difficulty of the prompt. This observation motivates applying a "compute-optimal" scaling strategy, which acts to, as effectively as possible, allocate test-time compute per prompt in an adaptive manner. Using this compute-optimal strategy, we can improve the efficiency of test-time compute scaling for math reasoning problems by more than 4x compared to a best-of-N baseline. Additionally, in a FLOPs-matched evaluation, we find that on problems where a smaller base model attains somewhat non-trivial success rates, test-time compute can be used to outperform a 14x larger model.