COLM 2025 Papers

Large Language Models (LLMs) have been extensively tuned to mitigate explicit biases, yet they often exhibit subtle implicit biases rooted in their pre-training data. Rather than directly probing LLMs with human-crafted questions that may trigger guardrails, we propose studying how models behave when they proactively ask questions themselves. The 20 Questions game, a multi-turn deduction task, serves as an ideal testbed for this purpose. We systematically evaluate geographic performance disparities in entity deduction using a new dataset, Geo20Q+, consisting of both notable people and culturally significant objects (e.g., foods, landmarks, animals) from diverse regions. We test popular LLMs across two gameplay configurations (canonical 20-question and unlimited turns) and in seven languages (English, Hindi, Mandarin, Japanese, French, Spanish, and Turkish). Our results reveal geographic disparities: LLMs are substantially more successful at deducing entities from the _Global North_ than the _Global South_, and the _Global West_ than the _Global East_. While Wikipedia pageviews and pre-training corpus frequency correlate mildly with performance, they fail to fully explain these disparities. Notably, the language in which the game is played has minimal impact on performance gaps. These findings demonstrate the value of creative, _free-form_ evaluation frameworks for uncovering subtle biases in LLMs that remain hidden in standard prompting setups. By analyzing how models initiate and pursue reasoning goals over multiple turns, we find geographic and cultural disparities embedded in their reasoning processes. We release the dataset (Geo20Q+) and code at https://sites.google.com/view/llmbias20q/home.

The emergence of generative pre-trained models has facilitated the synthesis of high-quality text but has also posed challenges in identifying factual errors in the generated text. In particular: (1) A wider range of tasks now face an increasing risk of containing factual errors when handled by generative models. (2) The content generated by these models tends to be lengthy and lacks clearly defined granularity for individual facts. (3) There is a scarcity of explicit evidence available during the process of fact checking. With the above challenges in mind, in this paper, we propose FacTool, a tool augmented multi-task and multi-domain framework for detecting factual errors of texts generated by large language models (e.g., ChatGPT). Experiments on four different tasks (knowledge-based QA, code generation, mathematical reasoning, and scientific literature review) show the efficacy of the proposed method. We release the code of FacTool with ChatGPT plugin in https://github.com/GAIR-NLP/factool.

A recent trend in LLMs is developing recurrent sub-quadratic models that improve long-context processing efficiency. We investigate leading large long-context models, focusing on how their fixed-size recurrent memory affects their performance. Our experiments reveal that, even when these models are trained for extended contexts, their use of long contexts remains underutilized. Specifically, we demonstrate that a chunk-based inference procedure, which identifies and processes only the most relevant portion of the input can mitigate recurrent memory failures and be effective for many long-context tasks: On LongBench, our method improves the overall performance of Falcon3-Mamba-Inst-7B by 14%, Falcon-Mamba-Inst-7B by 28%, RecurrentGemma-IT-9B by 50%, and RWKV6-Finch-7B by 51%. Surprisingly, this simple approach also leads to state-of-the-art results in the challenging LongBench v2 benchmark, showing competitive performance with equivalent size Transformers. Furthermore, our findings raise questions about whether recurrent models genuinely exploit long-range dependencies across multiple chunks, since our single-chunk strategy delivers stronger performance - even in tasks that presumably require cross-segment relations. We will release our code.

Language models (LMs) tend to show human-like preferences on a number of syntactic phenomena, but the extent to which these are attributable to direct exposure to the phenomena or more general properties of language is unclear. We explore this with the English dative alternation (DO: "gave Y the X" vs. PO: "gave the X to Y"), using a controlled rearing paradigm wherein we iteratively train small LMs on systematically manipulated input. We focus on two properties that affect the choice of alternant: length and animacy. Both properties are directly present in datives but also reflect more global tendencies for shorter elements to precede longer ones and animates to precede inanimates. First, by manipulating and ablating datives for these biases in the input, we show that direct evidence of length and animacy matters, but easy-first preferences persist even without such evidence. Then, using LMs trained on systematically perturbed datasets to manipulate global length effects (re-linearizing sentences globally while preserving dependency structure), we find that dative preferences can emerge from indirect evidence. We conclude that LMs' emergent syntactic preferences come from a mix of direct and indirect sources.

Dynamically integrating new or rapidly evolving information after (Large) Language Model pre-training remains challenging, particularly in low-data scenarios or when dealing with private and specialized documents. In-context learning and retrieval-augmented generation (RAG) face limitations, including their high inference costs and their inability to capture global document information. In this paper, we propose a way of modularizing knowledge by training document-level Knowledge Modules (KMs). KMs are lightweight components implemented as parameter-efficient LoRA modules, which are trained to store information about new documents and can be easily plugged into models on demand. We show that next-token prediction performs poorly as the training objective for KMs. We instead propose Deep Context Distillation: we learn KMs parameters such as to simulate hidden states and logits of a teacher that takes the document in context. Our method outperforms standard next-token prediction and pre-instruction training techniques, across two datasets. Finally, we highlight synergies between KMs and retrieval-augmented generation.

Standard decoding in a Transformer based language model is inherently sequential as we wait for a token’s embedding to pass through all the layers in the network before starting the generation of the next token. In this work, we propose anew architecture StagFormer (Staggered Transformer), which staggered execution along the time axis and thereby enables parallelizing the decoding process along the depth of the model. We achieve this by breaking the dependency of the token representation at time step $i$ in layer $l$ upon the representations of tokens until time step $i$ from layer $l−1$. Instead, we stagger the execution and only allow a dependency on token representations until time step $i−1$. The later sections of the Transformer still get access to the ”rich” representations from the prior section but only from those token positions which are one time step behind. StagFormer allows for different sections of the model to be executed in parallel yielding up to 33% speedup in decoding while being quality neutral. We also explore many natural variants of this idea. We present how weight-sharing across the different sections being staggered can be more practical in settings with limited memory. We show how one can approximate a recurrent model during inference using such weight-sharing. We explore the efficacy of using a bounded window attention to pass information from one section to another which helps drive further latency gains for some applications. We also explore demonstrate the scalability of the staggering idea over more than 2 sections of the Transformer.

Multi-turn interactions with language models (LMs) pose critical safety risks, as harmful intent can be strategically spread across exchanges. Yet, the vast majority of prior work has focused on single-turn safety, while adaptability and diversity remain among the key challenges of multi-turn red-teaming. To address these challenges, we present X-Teaming, a scalable framework that systematically explores how seemingly harmless interactions escalate into harmful outcomes and generates corresponding attack scenarios. X-Teaming employs collaborative agents for planning, attack optimization, and verification, achieving state-of-the-art multi-turn jailbreak effectiveness and diversity with success rates up to 98.1% across representative leading open-weight and closed-source models. In particular, X-Teaming achieves a 96.2% attack success rate against the latest Claude 3.7 Sonnet model, which has been considered nearly immune to single-turn attacks. Building on X-Teaming, we introduce X-Guard-Train, an open-source multi-turn safety training dataset that's $~20\times$ larger than the previous best resource, comprising 30K interactive jailbreaks, designed to enable robust multi-turn safety alignment for LMs. Our work offers essential tools and insights for mitigating sophisticated conversational attacks, advancing the multi-turn safety of LMs.

The key-value (KV) cache accelerates LLMs decoding by storing KV tensors from previously generated tokens. It reduces redundant computation at the cost of increased memory usage. To mitigate this overhead, existing approaches compress KV tensors into lower-bit representations; however, quantization errors can accumulate as more tokens are generated, potentially resulting in undesired outputs. In this paper, we introduce SQuat (Subspace-orthogonal KV cache quantization). It first constructs a subspace spanned by query tensors to capture the most critical task-related information. During key tensor quantization, it enforces that the difference between the (de)quantized and original keys remains orthogonal to this subspace, minimizing the impact of quantization errors on the attention mechanism’s outputs. SQuat requires no model fine-tuning, no additional calibration dataset for offline learning, and is grounded in a theoretical framework we develop. Through numerical experiments, we show that our method reduces peak memory by $2.17\times \sim 2.82\times$, improves throughput by $2.45\times \sim 3.60 \times$, and achieves more favorable benchmark scores than existing KV cache quantization algorithms.

Key-Value (KV) cache quantization has become a widely adopted optimization technique for efficient large language models (LLMs) inference by reducing KV cache memory usage and mitigating memory-bound constraints. Recent studies have emphasized the importance of preserving the original precision of KVs for the first few tokens to ensure the protection of attention sinks. While this approach has proven effective in mitigating performance degradation, its underlying principles remain insufficiently understood. Moreover, it fails to address the recent discovery that attention sinks can emerge beyond the initial token positions. In this work, we elucidate the underlying mechanisms of attention sinks during inference by examining their role in the cross-layer evolution of extreme activation outliers. Additionally, we provide a comprehensive analysis of the interplay between attention sinks and KV cache quantization. Based on our enhanced understanding, we introduce KVSink, a plug-and-play method that effectively predicts sink tokens with negligible overhead, enabling more thorough preservation. Extensive experiments demonstrate that KVSink outperforms the existing Preserve-First-N (PFN) strategy, offering more effective preservation of attention sinks during KV cache quantization. Moreover, when applied to the well-established KVQuant method, KVSink further improves perplexity (PPL) and reduces reliance on 16-bit numerical outliers.

As Large Language Models (LLMs) gain widespread practical applica- tion, offering model families with varying parameter sizes has become standard practice to accommodate diverse computational requirements. Traditionally, each model in the family is trained independently, incurring computational costs that scale additively with the number of models. In this work, we propose an efficient method for constructing model families via progressive training, where smaller models are incrementally expanded to larger sizes to create a complete model family. Through extensive ex- periments on a model family ranging from 1B to 8B parameters, we show that our approach reduces total computational cost by approximately 25% while maintaining comparable performance to independently trained mod- els. Moreover, by strategically adjusting the maximum learning rate based on model size, our method outperforms the independent training across various metrics. Beyond these improvements, our approach also fosters greater consistency in behavior across model sizes.

Large language models (LLMs) have demonstrated potential in reasoning tasks, but their performance on linguistics puzzles remains consistently poor. These puzzles, often derived from Linguistics Olympiad (LO) contests, provide a minimal contamination environment to assess LLMs' linguistic reasoning abilities across low-resource languages. This work analyses LLMs' performance on 629 problems across 41 low-resource languages by labelling each with linguistically informed features to unveil weaknesses. Our analyses show that LLMs struggle with puzzles involving higher morphological complexity and perform better on puzzles involving linguistic features that are also found in English. We also show that splitting words into morphemes as a pre-processing step improves solvability, indicating a need for more informed and language-specific tokenisers. These findings thus offer insights into some challenges in linguistic reasoning and modelling of low-resource languages.

Web agents enable users to perform tasks on web browsers through natural language interaction. Evaluating web agents trajectories is an important problem, since it helps us determine whether the agent successfully completed the tasks. Rule-based methods are widely used for this purpose, but they are challenging to extend to new tasks and may not always recognize successful trajectories. We may achieve higher accuracy through human evaluation, but the process would be substantially slower and more expensive. Automatic evaluations with LLMs may avoid the challenges of designing new rules and manually annotating trajectories, enabling faster and cost-effective evaluation. However, it is unclear how effective they are at evaluating web agents. To this end, we propose AgentRewardBench, the first benchmark to assess the effectiveness of LLM judges for evaluating web agents. AgentRewardBench contains 1302 trajectories across 5 benchmarks and 4 LLMs. Each trajectory in AgentRewardBench is reviewed by an expert, who answers questions pertaining to the success, side effects, and repetitiveness of the agent. Using our benchmark, we evaluate 12 LLM judges and find that no single LLM excels across all benchmarks. We also find that the rule-based evaluation used by common benchmarks tends to underreport the success rate of web agents, highlighting a key weakness of rule-based evaluation and the need to develop more flexible automatic evaluations. We release the benchmark at: https://agent-reward-bench.github.io

This work presents a framework for assessing whether large language models (LLMs) encode more factual knowledge in their parameters than what they express in their outputs. While a few studies hint at this possibility, none has clearly defined or demonstrated this phenomenon. We first propose a formal definition of knowledge, quantifying it for a given question as the fraction of correct-incorrect answer pairs where the correct one is ranked higher. This gives rise to external and internal knowledge, depending on the information used to score individual answer candidates: either the model’s observable token-level probabilities or its intermediate computations. Hidden knowledge arises when internal knowledge exceeds external knowledge. We then present a case study, applying this framework to three popular open-weights LLMs in a closed-book QA setup. Our results indicate that: (1) LLMs consistently encode more factual knowledge internally than what they express externally, with an average gap of 40%. (2) Surprisingly, some knowledge is so deeply hidden that a model can internally know an answer perfectly, yet fail to generate it even once, despite large-scale repeated sampling of 1,000 answers. This reveals fundamental limitations in the generation capabilities of LLMs, which (3) puts a practical constraint on scaling test-time compute via repeated answer sampling in closed-book QA: significant performance improvements remain inaccessible because some answers are practically never sampled, yet if they were, we would be guaranteed to rank them first.

Unlearning in Large Language Models (LLMs) aims to enhance safety, mitigate biases, and comply with legal mandates, such as the right to be forgotten. However, existing unlearning methods are brittle: minor query modifications, such as multi-hop reasoning and entity aliasing, can recover supposedly forgotten information. As a result, current evaluation metrics often create an illusion of effectiveness, failing to detect these vulnerabilities due to reliance on static, unstructured benchmarks. We propose a dynamic framework that stress tests unlearning robustness using complex structured queries. Our approach first elicits knowledge from the target model (pre-unlearning) and constructs targeted probes, ranging from simple queries to multi-hop chains, allowing precise control over query difficulty. Our experiments show that the framework: (1) shows comparable coverage to existing benchmarks by automatically generating semantically equivalent Q&A probes, (2) aligns with prior evaluations, and (3) uncovers new unlearning failures missed by other benchmarks, particularly in multi-hop settings. Furthermore, activation analyses show that single-hop queries typically follow dominant computation pathways, which are more likely to be disrupted by unlearning methods. In contrast, multi-hop queries tend to use alternative pathways that often remain intact, explaining the brittleness of unlearning techniques in multi-hop settings. Our framework enables practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications. We release the pip package and the code at https://sites.google.com/view/unlearningmirage/home.

Evaluation of language model outputs on structured writing tasks is typically conducted with a number of desirable criteria presented to human evaluators or large language models (LLMs). For instance, on a prompt like "Help me draft an academic talk on coffee intake vs research productivity", a model response may be evaluated for criteria like accuracy and coherence. However, high-quality responses should do more than just satisfy basic task requirements. An effective response to this query should include quintessential features of an academic talk, such as a compelling opening, clear research questions, and a takeaway. To help identify these implicit criteria, we introduce EvalAgent, a novel framework designed to automatically uncover nuanced and task-specific criteria. EvalAgent first mines expert-authored online guidance. It then uses this evidence to propose diverse, long-tail evaluation criteria that are grounded in reliable external sources. Our experiments demonstrate that the grounded criteria produced by EvalAgent are often implicit (not directly stated in the user's prompt), yet specific (high degree of lexical precision). Further, EvalAgent criteria are often not satisfied by initial responses but they are actionable, such that responses can be refined to satisfy them. Finally, we show that combining LLM-generated and EvalAgent criteria uncovers more human-valued criteria than using LLMs alone.

Large language models (LLMs) possess impressive linguistic capabilities but often fail to faithfully retain factual knowledge, leading to hallucinations and unreliable outputs. Understanding LLMs' knowledge deficiencies by exhaustively evaluating against full-scale knowledge bases is computationally prohibitive, especially for closed-weight models. We propose stochastic error ascent (SEA), a scalable and efficient framework for discovering knowledge deficiencies (errors) in closed-weight LLMs under a strict query budget. Rather than naively probing all knowledge candidates, SEA formulates error discovery as a stochastic optimization process: it iteratively retrieves new high-error candidates by leveraging the semantic similarity to previously observed failures. To further enhance search efficiency and coverage, SEA employs hierarchical retrieval across document and paragraph levels, and constructs a relation directed acyclic graph to model error propagation and identify systematic failure modes. Empirically, SEA uncovers 40.7× more knowledge errors than Automated Capability Discovery and 26.7% more than AutoBencher, while reducing the cost-per-error by 599× and 9×, respectively. Human evaluation confirms the high quality of generated questions, while ablation and convergence analyses validate the contribution of each component in SEA. Further analysis on the discovered errors reveals correlated failure patterns across LLM families and recurring deficits, highlighting the need for better data coverage and targeted fine-tuning in future LLM development.

Sexism is generally defined as prejudice and discrimination based on sex or gender, affecting every sector of society, from social institutions to relationships and individual behavior. Social media platforms amplify the impact of sexism by conveying discriminatory content not only through text but also across multiple modalities, highlighting the critical need for a multimodal approach to the analysis of sexism online. With the rise of social media platforms where users share short videos, sexism is increasingly spreading through video content. Automatically detecting sexism in videos is a challenging task, as it requires analyzing the combination of verbal, audio, and visual elements to identify sexist content. In this study, (1) we introduce MuSeD, a new Multimodal Spanish dataset for Sexism Detection consisting of ≈ 11 hours of videos extracted from TikTok and BitChute; (2) we propose an innovative annotation framework for analyzing the contributions of textual, vocal, and visual modalities to the classification of content as either sexist or non-sexist; and (3) we evaluate a range of large language models (LLMs) and multimodal LLMs on the task of sexism detection. We find that visual information plays a key role in labeling sexist content for both humans and models. Models effectively detect explicit sexism; however, they struggle with implicit cases, such as stereotypes—instances where annotators also show low agreement. This highlights the inherent difficulty of the task, as identifying implicit sexism depends on the social and cultural context.

The deployment of large-scale models, such as large language models (LLMs), incurs substantial costs due to their computational demands. To mitigate these costs and address challenges related to scalability and data security, there is a growing shift towards decentralized systems for model deployment, where choosing efficient inference acceleration schemes become crucial to manage computational resources effectively and enhance system responsiveness. In this work, we address the challenge of selecting optimal acceleration methods in decentralized systems by introducing a meta-learning-based framework. This framework automates the selection process by learning from historical performance data of various acceleration techniques across different tasks. Unlike traditional methods that rely on random selection or expert intuition, our approach systematically identifies the best acceleration strategies based on the specific characteristics of each task. We demonstrate that our meta-learning framework not only streamlines the decision-making process but also consistently outperforms conventional methods in terms of efficiency and performance. Our results highlight the potential of inference acceleration in decentralized AI systems, offering a path towards more democratic and economically feasible artificial intelligence solutions. Our code and data will be released later.

Is automated hallucination detection fundamentally possible? In this paper, we introduce a theoretical framework to rigorously study the (im)possibility of automatically detecting hallucinations produced by large language models (LLMs). Our model builds on the classical Gold-Angluin framework of language identification and its recent adaptation by Kleinberg and Mullainathan to the language generation setting. Concretely, we investigate whether an algorithm—trained on examples from an unknown target language $K$, chosen from a countable collection of languages $\mathcal{L}$, and given access to an LLM—can reliably determine if the LLM’s outputs are correct or constitute hallucinations. First, we establish a strong equivalence between hallucination detection and the classical problem of language identification. Specifically, we prove that any algorithm capable of identifying languages (in the limit) can be efficiently transformed into one that reliably detects hallucinations, and conversely, successful hallucination detection strategy inherently implies language identification. Given the notorious difficulty of language identification, our first result implies that hallucination detection is *impossible* for most collections of languages. Second, we show that once we enrich the detector’s training data, i.e., providing it with both positive examples (correct statements) and negative examples (explicitly labeled incorrect statements)— the conclusion dramatically changes. Under this enriched training regime, we show that automated hallucination detection is *possible* for any countable collection $\mathcal{L}$. Our theoretical results, thus, underscore the fundamental importance of expert-labeled feedback in the practical deployment of hallucination detection methods, reinforcing why feedback-based approaches, such as reinforcement learning with human feedback (RLHF), have proven so crucial in improving the reliability and safety of real-world LLMs.