COLM 2025 Papers

Recognizing the information flows and operations comprising data science and machine learning Python notebooks is critical for evaluating, reusing, and adapting notebooks for new tasks. Investigating a notebook via re-execution often is impractical due to the challenges of resolving data and software dependencies. While Large Language Models (LLMs) pre-trained on large codebases have demonstrated effectiveness in understanding code without running it, we observe that they fail to understand some realistic notebooks due to hallucinations and long-context challenges. To address these issues, we propose a notebook understanding task yielding an information flow graph and corresponding cell execution dependency graph for a notebook, and demonstrate the effectiveness of a pincer strategy that uses limited syntactic analysis to assist full comprehension of the notebook using an LLM. Our Capture and Resolve Assisted Bounding Strategy (CRABS) employs shallow syntactic parsing and analysis of the abstract syntax tree (AST) to capture the correct interpretation of a notebook between lower and upper estimates of the inter-cell I/O set$\textemdash$the flows of information into or out of cells via variables$\textemdash$then uses an LLM to resolve remaining ambiguities via cell-by-cell zero-shot learning, thereby identifying the true data inputs and outputs of each cell. We evaluate and demonstrate the effectiveness of our approach using an annotated dataset of 50 representative, highly up-voted Kaggle notebooks that together represent 3454 actual cell inputs and outputs. The LLM correctly resolves 1397 of 1425 (98%) ambiguities left by analyzing the syntactic structure of these notebooks. Across 50 notebooks, CRABS achieves average $F_1$ scores of 98% identifying cell-to-cell information flows and 99% identifying transitive cell execution dependencies. Moreover, 37 out of the 50 (74%) individual information flow graphs and 41 out of 50 (82%) cell execution dependency graphs match the ground truth exactly.

The inference-time resource costs of large language and vision models present a growing challenge in production deployments. We propose the use of ***foundation model programs***, i.e., programs that can invoke foundation models with varying resource costs and performance, as an approach to this problem. Specifically, we present a method that translates a task into a program, then learns a policy for resource allocation that, on each input, selects foundation model "backends" for each program module. The policy uses smaller, cheaper backends to handle simpler subtasks, while allowing more complex subtasks to leverage larger, more capable models. We evaluate the method on two new "streaming" visual question-answering tasks in which a system answers a question on a sequence of inputs, receiving ground-truth feedback after each answer. Compared to monolithic multi-modal models, our implementation achieves up to 98\% resource savings with minimal accuracy loss, demonstrating its potential for scalable and resource-efficient multi-modal inference. The source code and the benchmarks are available at [GitHub](https://github.com/Flitternie/FMProgramming).

Large Language Models trained on large-scale uncontrolled corpora often encode stereotypes and biases, which can be displayed through harmful text generation or biased associations. However, do they also pick up subtler linguistic patterns that can potentially reinforce and communicate biases and stereotypes, as humans do? We aim to bridge theoretical insights from social science with bias research in NLP by designing controlled, theoretically motivated LLM experiments to elicit this type of bias. Our case study is negation bias, the bias that humans have towards using negation to describe situations that challenge common stereotypes. We construct an evaluation dataset containing negated and affirmed stereotypical and anti-stereotypical sentences and evaluate the performance of eight language models using perplexity as a metric for measuring model surprisal. We find that the autoregressive decoder models in our experiment exhibit this bias, while we do not find evidence for it among the stacked encoder models.

Large language models (LLMs) trained with Reinforcement Learning from Human Feedback (RLHF) have demonstrated remarkable capabilities, but their underlying reward functions and decision-making processes remain opaque. This paper introduces a novel approach to interpreting LLMs by applying inverse reinforcement learning (IRL) to recover their implicit reward functions. We conduct experiments on toxicity-aligned LLMs of varying sizes, extracting reward models that achieve up to 85\% accuracy in predicting human preferences. Our analysis reveals key insights into the non-identifiability of reward functions, the relationship between model size and interpretability, and potential pitfalls in the RLHF process. We demonstrate that IRL-derived reward models can be used to fine-tune new LLMs, resulting in comparable or improved performance on toxicity benchmarks. This work provides a new lens for understanding and improving LLM alignment, with implications for the responsible development and deployment of these powerful systems.

Recent advancements in reasoning language models have demonstrated remarkable performance in complex tasks, but their extended chain-of-thought reasoning process increases inference overhead. While quantization has been widely adopted to reduce the inference cost of large language models, its impact on reasoning models remains understudied. In this paper, we conduct the first systematic study on quantized reasoning models, evaluating the open-sourced DeepSeek-R1-Distilled Qwen and LLaMA families ranging from 1.5B to 70B parameters, QwQ-32B, and Qwen3-8B. Our investigation covers weight, KV cache, and activation quantization using state-of-the-art algorithms at varying bit-widths, with extensive evaluation across mathematical (AIME, MATH-500), scientific (GPQA), and programming (LiveCodeBench) reasoning benchmarks. Our findings reveal that while lossless quantization can be achieved with W8A8 or W4A16 quantization, lower bit-widths introduce significant accuracy risks. We further identify model size, model origin, and task difficulty as critical determinants of performance. Contrary to expectations, quantized models do not exhibit increased output lengths. In addition, strategically scaling the model sizes or reasoning steps can effectively enhance the performance. All quantized models and codes are open-sourced in https://github.com/ruikangliu/Quantized-Reasoning-Models.

Achieving robust textual comprehension and in-context learning requires language models capable of interpreting entire document contexts. However, scaling these models directly to long contexts remains technically challenging, prompting a surge of “extension” strategies. To date, rigorous comparisons among these approaches have been complicated by inconsistent base models, training data, and evaluation metrics, limiting our understanding of how long-context performance may differ from standard benchmarks. In this work, we introduce a controlled extension protocol and a standardized evaluation pipeline, enabling an apples-to-apples comparison across diverse long-context methods. Through extensive experiments, we uncover three key insights: (1) perplexity emerges as a helpful (albeit imperfect) indicator for gauging model quality on lengthy-context tasks, (2) approximate attention mechanisms exhibit systematic performance deficits on long-context benchmarks, and (3) exact fine-tuning remains robust within its extension range, although extrapolation beyond that range continues to pose challenges. All codebases, trained models, and checkpoints will be released, fostering transparency and accelerating progress in this critical area of AI research. Our results not only help clarify the current landscape of long-context modeling but also offer guidance for building more capable, context-aware language models.

Large Language Models (LLMs) are increasingly used for social character simulations, enabling applications in role-playing agents and Computational Social Science (CSS). However, their inherent flaws—such as inconsistencies in simulated roles—raise concerns about their reliability and trustworthiness. In this paper, we systematically investigate these flaws and explore potential solutions. To assess the reliability of LLM-based simulations, we introduce TrustSim, a benchmark dataset covering 10 CSS-related topics. Through experiments on 14 LLMs, we uncover persistent inconsistencies in simulated roles and find that higher general model performance does not necessarily correlate with greater simulation reliability. To mitigate these flaws, we propose Adaptive Learning Rate Based ORPO (AdaORPO), a reinforcement learning-based algorithm that improves simulation consistency across seven LLMs. Our study not only exposes critical weaknesses in LLM-driven social character simulations but also offers a pathway toward more robust and trustworthy simulations, laying the foundation for future advancements in this field.

Large language models (LLMs) are typically multilingual due to pretraining on diverse multilingual corpora. But can these models relate corresponding concepts across languages, i.e., be crosslingual? This study evaluates state-of-the-art LLMs on inherently crosslingual tasks. We observe that while these models show promising surface-level crosslingual abilities on machine translation and embedding space analyses, they struggle with deeper crosslingual knowledge transfer, revealing a crosslingual knowledge barrier in both general (MMLU benchmark) and domain-specific (Harry Potter quiz and TOFU benchmark) contexts. Since simple inference-time mitigation methods offer only limited improvement, we propose fine-tuning of LLMs on mixed-language data, which effectively reduces these gaps, even when using out-of-domain datasets like WikiText. Our findings suggest the need for explicit optimization to unlock the full crosslingual potential of LLMs. Our code is available at https://github.com/google-research/crosslingual-knowledge-barriers.

Employing language models as evaluators of long-form output (LLM-as-a-Judge) has become the \textit{de facto} standard for automatic evaluation. However, most LLM judges have been optimized exclusively for English outputs, with strategies for enhancing judges' multilingual evaluation capabilities remaining largely unexplored in the current literature. This has created a disparity in the quality of automatic evaluation methods for other languages, ultimately hindering the development of models with better multilingual capabilities. To bridge this gap, we introduce M-Prometheus, a suite of open-weight LLM judges ranging from 3B to 14B parameters that can provide both direct assessment and pairwise comparison feedback on multilingual outputs. M-Prometheus models outperform state-of-the-art open LLM judges on multilingual reward benchmarks spanning more than 20 languages, as well as on literary machine translation evaluation covering 4 language pairs. Furthermore, we find M-Prometheus models can be used with quality-aware decoding methods to significantly improve generated outputs, showcasing their utility for the development of better multilingual models. Crucially, through extensive ablations, we identify key strategies for training an effective multilingual judge. Our findings highlight the significance of model size and base model selection, and the advantages of using natively multilingual data rather than translated data. We release our models, training dataset, and code to reproduce our experiments.

There has been recent interest in whether large language models (LLMs) can introspect about their own internal states. Such abilities would make LLMs more interpretable, and also validate the use of standard introspective methods in linguistics to evaluate grammatical knowledge in models (e.g., asking "Is this sentence grammatical?"). We systematically investigate emergent introspection across 21 open-source LLMs, in two domains where introspection is of theoretical interest: grammatical knowledge and word prediction. Crucially, in both domains, a model’s internal linguistic knowledge can be theoretically grounded in direct measurements of string probability. We then evaluate whether models' responses to metalinguistic prompts faithfully reflect their internal knowledge. We propose a new measure of introspection: the degree to which a model’s prompted responses predict its own string probabilities, beyond what would be predicted by another model with nearly identical internal knowledge. While both metalinguistic prompting and probability comparisons lead to high task accuracy, we do not find evidence that LLMs have privileged "self-access". By using general tasks, controlling for model similarity, and evaluating a wide range of open-source models, we show that LLMs cannot introspect, and add new evidence to the argument that prompted responses should not be conflated with models' linguistic generalizations.

Large language models (LLMs) have demonstrated strong performance in sentence-level machine translation, but scaling to document-level translation remains challenging, particularly in modeling long-range dependencies and discourse phenomena across sentences and paragraphs. In this work, we propose a method to improve LLM-based long-document translation through targeted fine-tuning on high-quality document-level data, which we curate and introduce as DocBlocks. Our approach supports multiple translation paradigms, including direct document-to-document and chunk-level translation, by integrating instructions both with and without surrounding context. This enables models to better capture cross-sentence dependencies while maintaining strong sentence-level translation performance. Experimental results show that incorporating multiple translation paradigms improves document-level translation quality and inference speed compared to prompting and agent-based methods.

Large Language Models (LLMs) have become indispensable in real-world applications. However, their widespread adoption raises significant safety concerns, particularly in responding to socially harmful questions. Despite substantial efforts to improve model safety through alignment, aligned models can still have their safety protections undermined by subsequent fine-tuning—even when the additional training data appears benign. In this paper, we empirically demonstrate that this vulnerability stems from the sensitivity of safety-critical low-rank subspaces in LLM parameters to fine-tuning. Building on this insight, we propose a novel training-free method, termed Low-Rank Extrapolation (LoX), to enhance safety robustness by extrapolating the safety subspace of an aligned LLM. Our experimental results confirm the effectiveness of LoX, demonstrating significant improvements in robustness against both benign and malicious fine-tuning attacks while preserving the model’s adaptability to new tasks. For instance, LoX leads to 11% to 54% absolute reductions in attack success rates (ASR) facing benign or malicious fine-tuning attacks. By investigating the ASR landscape of parameters, we attribute the success of LoX to that the extrapolation moves LLM parameters to a flatter zone, thereby less sensitive to perturbations. The code will be released upon acceptance.

Recent advancements in instructional fine-tuning have injected noise into embeddings, with NEFTune (Jain et al., 2024) setting benchmarks using uniform noise. Despite NEFTune’s empirical findings that uniform noise outperforms Gaussian noise, the reasons for this remain unclear. This paper aims to clarify this by offering a thorough analysis, both theoretical and empirical, indicating comparable performance among these noise types. Additionally, we introduce a new fine-tuning method for language models, utilizing symmetric noise in embeddings. This method aims to enhance the model’s function by more stringently regulating its local curvature, demonstrating superior performance over the current method, NEFTune. When fine-tuning the LLaMA-2-7B model using Alpaca, standard techniques yield a 29.79% score on AlpacaEval. However, our approach, SymNoise, increases this score significantly to 69.04%, using symmetric noisy embeddings. This is a 6.7% improvement over the state-of-the-art method, NEFTune (64.69%). Furthermore, when tested on various models and stronger baseline instruction datasets, such as Evol-Instruct, ShareGPT, OpenPlatypus, SymNoise consistently outperforms NEFTune. The current literature, including NEFTune, has underscored the importance of more in-depth research into the application of noise-based strategies in the fine-tuning of language models. Our approach, SymNoise, is another significant step towards this direction, showing notable improvement over the existing state-of-the-art method.

Recent studies suggest that very small language models (SLMs) can generate surprisingly coherent text when trained on simplified, child-directed corpora such as TinyStories. These findings have been interpreted as evidence that readability—characterized by accessible vocabulary, familiar narrative structure, and simple syntax—plays a key role in enabling such capabilities to emerge. In this paper, we challenge that interpretation. We construct synthetic datasets with matched structure but varied readability, and find that readability alone does not predict coherence or learning efficiency in SLMs. Models trained on complex, adult-level text perform comparably to those trained on simplified language, and even exhibit faster development of coherence during training. Instead, we show that statistical simplicity, as measured by n-gram diversity, is a stronger predictor of learnability. Our findings caution against the growing trend of anthropomorphizing language model training—drawing parallels to human cognitive development without empirical basis—and argue for more precise reasoning about what properties actually support capability emergence in small models.

Improvements in language models are often driven by increasing the quality of the data we train them on, which can be limiting when strong supervision is not readily available. In this work, we show that paired preference data consisting of individually weak data points can enable gains beyond the strength of each individual sample. We formulate the **delta learning hypothesis** to explain this phenomenon, positing that the relative quality _delta_ between points suffices to drive learning via preference tuning—even when supervised finetuning on the weak data hurts. We validate our hypothesis in controlled experiments and at scale, where we post-train 8B models on preference data generated by pairing a small 3B model's responses with outputs from an even smaller 1.5B model to ensure a meaningful delta. Strikingly, on a standard 11-benchmark evaluation suite (MATH, MMLU, etc.), our simple recipe matches the performance of Tülu 3, a state-of-the-art open model that was tuned from the same base as our model while relying on vastly stronger supervisors (e.g., GPT-4o). Delta learning thus enables simpler and cheaper open recipes for state-of-the-art post-training, highlighting that models can learn a surprising amount from data that might typically be considered weak.

Most natural language reasoning tasks in the research community assume consistent input knowledge. Nevertheless, real-world scenarios often involve inconsistent information, which might lead to divergent conclusions and are typically associated with varying levels of uncertainty. This raises a key research question: can large language models (LLMs) effectively handle uncertainty in their reasoning process to maximize knowledge consistency? In this paper, we propose a framework for evaluating reasoning over inconsistent knowledge. Our approach models uncertainty via weights of logical rules, leveraging Markov logic networks (MLN), which integrate probabilistic reasoning with first-order logic. This enables us to quantify inconsistencies in knowledge bases, and hence rigorously evaluate LLM reasoning. We introduce two tasks using this framework: 1) QA, which involves answering questions by integrating inconsistent knowledge; and 2) knowledge rectification, where we aim to rectify language models' acquired knowledge to improve consistency. We curate a dataset of 3,000 MLN-formatted knowledge bases to implement these tasks. We evaluate state-of-the-art LLMs on these tasks and highlight their limitations in uncertainty-aware reasoning over inconsistent logical knowledge.

Neural sentence embedding models for dense retrieval typically rely on binary relevance labels, treating query-document pairs as either relevant or irrelevant. However, real-world relevance often exists on a continuum, and recent advances in large language models (LLMs) have made it feasible to scale the generation of fine-grained graded relevance labels. In this work, we propose \textbf{BiXSE}, a simple and effective pointwise training method that optimizes binary cross-entropy (BCE) over LLM-generated graded relevance scores. BiXSE interprets these scores as probabilistic targets, enabling granular supervision from a single labeled query-document pair per query. Unlike pairwise or listwise losses that require multiple annotated comparisons per query, BiXSE achieves strong performance with reduced annotation and compute costs by leveraging in-batch negatives. Extensive experiments across sentence embedding (MMTEB) and retrieval benchmarks (BEIR, TREC-DL) show that BiXSE consistently outperforms softmax-based contrastive learning (InfoNCE), and matches or exceeds strong pairwise ranking baselines when trained on LLM-supervised data. BiXSE offers a robust, scalable alternative for training dense retrieval models as graded relevance supervision becomes increasingly accessible.

Multimodal in-context learning (ICL) equips Large Vision-language Models (LVLMs) with the ability to adapt to new tasks via multiple user-provided demonstrations, without requiring any model parameter updates. However, its effectiveness is constrained by the token-intensive nature of multimodal inputs and the complexity of cross-modal few-shot reasoning, which together hinder LVLMs from extracting useful patterns from demonstrations. To address these challenges, we propose \textbf{M²IV}, a novel representation engineering approach that replaces explicit token-level demonstrations with a set of learnable Multimodal In-context Vectors directly injected into the residual streams of LVLMs. By analyzing the distinct roles of multi-head attention (MHA) and multi-layer perceptrons (MLP) in the ICL process, we design a training strategy that enables M²IV to perform fine-grained semantic distillation and robust cross-modal representation learning. M²IV not only improves performance across diverse tasks and LVLMs but also significantly reduces token overhead, enabling graceful scaling to many-shot scenarios. To further enhance usability, we introduce \textbf{VLibrary}, a repository that stores trained M²IVs for flexible retrieval and injection. With VLibrary, users can steer pre-trained LVLMs in a customized manner that meets diverse requirements. Extensive experiments demonstrate that M²IV consistently outperforms vanilla ICL and prior representation engineering baselines, achieving an average accuracy gain of 3.74\% with substantial improvements in overall efficiency.

While in-context learning is well-studied with decoder-only language models (LLMs), its utility for encoder-only models remains underexplored. We study in-context learning for encoder-only models for text retrieval tasks. Can incorporating in-context examples (query-document pairs) to the target query enhance retriever performance? Our approach, \texttt{RARe}, finetunes a pre-trained model with in-context examples whose query is semantically similar to the target query. This approach achieves performance gains of up to +2.72\% nDCG across open-domain retrieval datasets (BeIR, RAR-b) compared to using the target query only as an input. In particular, we find \texttt{RARe} exhibits stronger out-of-domain generalization compared to models using queries without in-context examples, similar to what is seen for in-context learning in LLMs. We further provide analysis on the design choices of in-context example augmentation for retrievers and lay the foundation for future work.