COLM 2025 论文列表

Training large, general-purpose language models poses significant challenges. The growing availability of specialized *expert* models, fine-tuned from pretrained models for specific tasks or domains, offers a promising alternative. Leveraging the potential of these existing expert models in real-world applications requires effective methods to select or merge the models best suited for a given task. This paper introduces SpectR, an approach for dynamically composing expert models at each time step during inference. Notably, our method requires no additional training and enables flexible, token- and layer-wise model combinations. Our experimental results demonstrate that SpectR improves routing accuracy over alternative training-free methods, increasing task performance across expert domains.

We introduce D3 ("Diverse Data for Diff-by-Diff Coding"), a large dataset for training LMs to iteratively synthesize general-purpose Python source code by generating file diffs. D3 frames code synthesis as a goal-conditioned sequential decision-making problem, where goals, states, and actions are represented by token sequences corresponding to the description of a functionality to add, the current contents of a file, and a file diff, respectively. The dataset contains 8 billion tokens of instruction + file-state + file-diff-sequence examples sampled from 850,000 human-written Python source files. To construct D3, we filter, augment, and annotate source code from The Stack by sampling synthetic file-diff sequences with a code analysis tool and labeling each sample with an LLM-generated rationale. In our experiments, we show that mid-training LMs like Llama 3.2 1b and 3b on D3 prior to supervised fine-tuning (SFT) on task-curated data improves performance on synthesis & editing tasks. On benchmarks like HumanEvalSynth and HumanEvalFix, we observe improvements in pass@1 of 3 to 6 points compared to direct SFT. D3-trained models are particularly strong at completing partial human-written solutions to programming problems.

Understanding and mitigating hallucinations in Large Language Models (LLMs) is crucial for ensuring reliable content generation. While previous research has primarily focused on "when" LLMs hallucinate, our work explains "why" and directly links model behaviour to the pre-training data that forms their prior knowledge. Specifically, we demonstrate that an asymmetry exists in the recognition of logically equivalent facts, which can be attributed to frequency discrepancies of entities appearing as subjects versus objects. Given that most pre-training datasets are inaccessible, we leverage the fully open-source $\texttt{OLMo}$ series by indexing its $\texttt{Dolma}$ dataset to estimate entity frequencies. Using relational facts (represented as triples) from $\texttt{Wikidata5M}$, we construct probing datasets to isolate this effect. Our experiments reveal that facts with a high-frequency subject and a low-frequency object are better recognised than their inverse, despite their logical equivalence. The pattern reverses in low-to-high frequency settings, and no statistically significant asymmetry emerges when both entities are high-frequency. These findings highlight the influential role of pre-training data in shaping model predictions and provide insights for inferring the characteristics of pre-training data in closed or partially closed LLMs.

Fine-tuning large pre-trained LLMs generally demands extensive GPU memory. Traditional first-order optimizers like SGD encounter substantial difficulties due to increased memory requirements from storing activations and gradients during both the forward and backward phases as the model size expands. Alternatively, zeroth-order (ZO) techniques can compute gradients using just forward operations, eliminating the need to store activations. Furthermore, by leveraging CPU capabilities, it's feasible to enhance both the memory and processing power available to a single GPU. We propose a novel framework, ZO2 (Zeroth-Order Offloading), for efficient zeroth-order fine-tuning of LLMs with only limited GPU memory. Our framework dynamically shifts model parameters between the CPU and GPU as required, optimizing computation flow and maximizing GPU usage by minimizing downtime. This integration of parameter adjustments with ZO's double forward operations reduces unnecessary data movement, enhancing the fine-tuning efficacy. Additionally, our framework supports an innovative low-bit precision approach in AMP (Automatic Mixed Precision) mode to streamline data exchanges between the CPU and GPU. Employing this approach allows us to fine-tune extraordinarily large models, such as the OPT-175B with 175 billion parameters, on a mere 18GB GPU. Moreover, our framework achieves these results with almost no additional time overhead and absolutely no accuracy loss compared to standard zeroth-order methods. ZO2's code has been open-sourced in https://github.com/liangyuwang/zo2.

We introduce MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. This is the first Gym environment for machine learning (ML) tasks, enabling research on reinforcement learning (RL) algorithms for training such agents. MLGym-bench consists of 13 diverse and open-ended AI research tasks from diverse domains such as computer vision, natural language processing, reinforcement learning, and game theory. Solving these tasks requires real-world AI research skills such as generating new ideas and hypotheses, creating and processing data, implementing ML methods, training models, running experiments, analyzing the results, and iterating through this process to improve on a given task. We evaluate a number of frontier large language models (LLMs) on our benchmarks such as Claude-3.5-Sonnet, Llama-3.1 405B, GPT-4o, o1-preview, and Gemini-1.5 Pro. Our MLGym framework makes it easy to add new tasks, integrate and evaluate models or agents, generate synthetic data at scale, as well as develop new learning algorithms for training agents on AI research tasks. We find that current frontier models can improve on the given baselines, usually by finding better hyperparameters, but do not generate novel hypotheses, algorithms, architectures, or substantial improvements. We open-source our framework and benchmark to facilitate future research in advancing the AI research capabilities of LLM agents.

Domain reweighting is an emerging research area aimed at adjusting the relative weights of different data sources to improve the effectiveness and efficiency of LLM pre-training. We show that data mixtures that perform well at smaller scales may not retain their advantage at larger scales, challenging the existing practice of determining competitive mixtures in small-scale experiments and *directly* applying them at much larger scales. To address this, we propose AutoScale, a two-stage, scale-aware data composition framework. First, AutoScale fits a parametric model that predicts the model’s loss under different data compositions, then uses it to find an approximate best allocation at smaller, more manageable budgets. Next, leveraging a novel theoretical analysis of how optimal compositions evolve with scale, AutoScale extrapolates that composition to larger budgets without further retraining. Empirically, AutoScale accelerates convergence and improves downstream performance. For instance, when pre-training GPT-2 Large, it achieves a 28\% faster perplexity reduction than baselines and up to a 38\% speed-up over unweighted training, while yielding best-average results on various downstream tasks. Overall, our findings illustrate how domain importance shifts with training scale, underscoring the need for scale-dependent data curation in LLM training. Our code is open-sourced.

Existing benchmarks for evaluating long-context language models (LCLMs) primarily focus on long-context recall, requiring models to produce short responses based on a few critical snippets while processing thousands of irrelevant tokens. We introduce LongProc (Long Procedural Generation), a new benchmark that requires both the integration of highly dispersed information and long-form generation. LongProc consists of six diverse procedural generation tasks, such as extracting structured information from HTML pages into a TSV format and executing complex search procedures to create travel plans. These tasks challenge LCLMs by testing their ability to follow detailed procedural instructions, synthesize and reason over dispersed information, and generate structured, long-form outputs (up to 8K tokens). Furthermore, as these tasks adhere to deterministic procedures and yield structured outputs, they enable reliable rule-based evaluation. We evaluated 23 LCLMs, including instruction-tuned models and recent reasoning models, on LongProc at three difficulty levels, with the maximum number of output tokens set at 500, 2K, and 8K. Notably, while all tested models claim a context window size above 32K tokens, open-weight models typically falter on 2K-token tasks, and closed-source models like GPT-4o show significant degradation on 8K-token tasks. Reasoning models achieve stronger overall performance in long-form generation, benefiting from long CoT training. Further analysis reveals that LCLMs struggle to maintain long-range coherence in long-form generations. These findings highlight critical limitations in current LCLMs and suggest substantial room for improvement.

We present significant extensions to diffusion-based sequence generation models, blurring the line with autoregressive language models. We introduce *hyperschedules*, which assign distinct noise schedules to individual token positions, generalizing both autoregressive models (*e.g.*, GPT) and conventional diffusion models (*e.g.*, SEDD, MDLM) as special cases. Second, we propose two \emph{hybrid token-wise noising processes} that interpolate between absorbing and uniform processes, enabling the model to fix past mistakes, and we introduce a *novel inference algorithm* that leverages this new feature in a simplified context inspired from MDLM. To support efficient training and inference, we design attention masks compatible with KV-caching. Our methods achieve state-of-the-art perplexity and generate diverse, high-quality sequences across standard benchmarks, suggesting a promising path for autoregressive diffusion-based sequence generation. See code and resources at https://hdlm-colm.github.io/ .

Although large language models (LLMs) have become more capable and accurate across many tasks, some fundamental sources of unreliability remain in their behavior. One key limitation is their inconsistency at reporting the same information when prompts are changed. In this paper, we consider the discrepancy between a model’s generated answer and their own verification of that answer, the generator-validator gap. We define this gap in a more stringent way than prior work: we expect correlation of scores from a generator and a validator over the entire set of candidate answers, i.e., candidate completions that could possibly arise during ordinary language use without breaking Gricean norms. We show that according to this measure, a large gap exists in various settings, including question answering, lexical semantics tasks, and next-word prediction. We then propose RankAlign, a ranking-based training method, and show that it significantly closes the gap, surpassing all baseline methods. Moreover, this approach generalizes well to out-of-domain tasks and lexical items.

In-Context Learning (ICL) is an intriguing ability of large language models (LLMs). Despite a substantial amount of work on its behavioral aspects and how it emerges in miniature setups, it remains unclear which mechanism assembles task information from the individual examples in a fewshot prompt. We use causal interventions to identify information flow in Gemma-2 2B for five naturalistic ICL tasks. We find that the model infers task information using a two-step strategy we call contextualize-then-aggregate: In the lower layers, the model builds up representations of individual fewshot examples, which are contextualized by preceding examples through connections between fewshot input and output tokens across the sequence. In the higher layers, these representations are aggregated to identify the task and prepare prediction of the next output. The importance of the contextualization step differs between tasks, and it may become more important in the presence of ambiguous examples. Overall, by providing rigorous causal analysis, our results shed light on the mechanisms through which ICL happens in language models.

Foundation models, particularly Large Language Models (LLMs), have garnered significant interest due to their wide range of applications. Yet these models demonstrate notable weaknesses when confronted with tasks involving iterative sub-problems or deliberately misleading content—exemplified by complex arithmetic operations and comprehensive fake news evaluation. Conventional instructional prompting frequently produces flawed outputs in these scenarios. While research has established that advanced techniques such as Chain-of-Thoughts and Least-to-Most methodologies can dramatically enhance LLM performance, emerging investigation indicates that a more streamlined divide-and-conquer (DaC) approach—which systematically partitions input sequences into discrete components—can yield remarkable improvements for particular problem classes like misinformation assessment. Our investigation rigorously examines the efficacy of DaC prompting strategies and precisely delineates the task characteristics that benefit most from this methodology. Through comprehensive theoretical analysis, we establish formal guarantees for performance enhancement in specifically identified task categories. We validate our theoretical framework through focused empirical studies on large integer multiplication and factual verification tasks, where experimental outcomes robustly confirm our analytical predictions, demonstrating DaC's practical superiority in these challenging domains.

Academic writing requires both coherent text generation and precise citation of relevant literature. Although recent Retrieval-Augmented Generation (RAG) systems have significantly improved factual accuracy in general-purpose text generation, their ability to support professional academic writing remains limited. In this work, we introduce ScholarCopilot, a unified framework designed to enhance existing large language models for generating professional academic articles with accurate and contextually relevant citations. ScholarCopilot dynamically determines when to retrieve scholarly references by generating a retrieval token [RET], which is then used to query a citation database. The retrieved references are fed into the model to augment the generation process. We jointly optimize both the generation and citation tasks within a single framework to improve efficiency. Our model is built upon Qwen-2.5-7B and trained on 500K papers from arXiv. It achieves a top-1 retrieval accuracy of 40.1% on our evaluation dataset, outperforming baselines such as E5-Mistral-7B-Instruct (15.0%) and BM25 (9.8%). On a dataset of 1,000 academic writing samples, ScholarCopilot scores 16.2/25 in generation quality--measured across relevance, coherence, academic rigor, completeness, and innovation--significantly surpassing all existing models, including much larger ones like the Retrieval-Augmented Qwen2.5-72B-Instruct. Human studies further demonstrate that ScholarCopilot, despite being a 7B model, significantly outperforms ChatGPT, achieving 100% preference in citation quality and over 70% in overall usefulness.

Transformers, while powerful, suffer from quadratic computational complexity and the ever-growing Key-Value (KV) cache of the attention mechanism. This paper introduces Trellis, a novel Transformer architecture with bounded memory that learns how to compress its key-value memory dynamically at test time. Trellis replaces the standard KV cache with a fixed-size memory and train a two-pass recurrent compression mechanism to store new keys and values into memory. To achieve this, it leverages an online gradient descent procedure with a forget gate, enabling the compressed memory to be updated recursively while learning to retain important contextual information from incoming tokens at test time. Extensive experiments on language modeling, common-sense reasoning, recall-intensive tasks, and time series show that the proposed architecture outperforms strong baselines. Notably, its performance gains increase as the sequence length increases, highlighting its potential for long-context applications.

State-of-the-art large language models (LLMs) are now claiming remarkable supported context lengths of 256k or even more. In contrast, the average context lengths of mainstream benchmarks are insufficient (5k-21k), and they suffer from potential knowledge leakage and inaccurate metrics, resulting in biased evaluation. This paper introduces LV-Eval, a challenging long-context benchmark with five length levels (16k, 32k, 64k, 128k, and 256k) reaching up to 256k words. LV-Eval features two main tasks, single-hop QA and multi-hop QA, comprising 11 bilingual datasets. The design of LV-Eval has incorporated three key techniques, namely confusing facts insertion, keyword and phrase replacement, and keyword-recall-based metric design. The advantages of LV-Eval include controllable evaluation across different context lengths, challenging test instances with confusing facts, mitigated knowledge leakage, and more objective evaluations. We evaluate 15 LLMs on LV-Eval and conduct ablation studies on the benchmarking techniques. The results reveal that: (i) Moonshot-v1 and recent large-scale open-source models, such as Qwen-2.5-72B and Llama-3.1-70B, achieve the highest performance on LV-Eval, particularly at lengths below $64k$. (ii) Models exhibit distinct score trends. For example, GLM-4-9B-128k, Yi-6B-200k, and Llama3-8B-1M exhibit a relatively gentle degradation of performance, but their absolute performances may not necessarily be higher than those of LLMs with shorter context lengths. (iii) LLMs' performances can significantly degrade in the presence of confusing information, especially in the pressure test of "needle in a haystack". (iv) Issues related to knowledge leakage and inaccurate metrics introduce bias in evaluation, and these concerns are alleviated in LV-Eval.

Modern Vision-Language Models (VLMs) exhibit remarkable visual and linguistic capabilities, achieving impressive performance in various tasks such as image recognition and object localization. However, their effectiveness in fine-grained tasks remains an open question. In everyday scenarios, individuals encountering design materials, such as magazines, typography tutorials, research papers, or branding content, may wish to identify aesthetically pleasing fonts used in the text. Given their multimodal capabilities and free accessibility, many VLMs are often considered potential tools for font recognition. This raises a fundamental question: Do VLMs truly possess the capability to recognize fonts? To investigate this, we introduce the Font Recognition Benchmark (FRB), a compact and well-structured dataset comprising 15 commonly used fonts. FRB includes two versions: (i) an easy version, where 10 sentences are rendered in different fonts, and (ii) a hard version, where each text sample consists of the names of the 15 fonts themselves, introducing a stroop effect that challenges model perception. Through extensive evaluation of various VLMs on font recognition tasks, we arrive at the following key findings: (i) Current VLMs exhibit limited font recognition capabilities, with many state-of-the-art models failing to achieve satisfactory performance and being easily affected by the stroop effect introduced by textual information. (ii) Few-shot learning and Chain-of-Thought (CoT) prompting provide minimal benefits in improving font recognition accuracy across different VLMs. (iii) Attention analysis sheds light on the inherent limitations of VLMs in capturing semantic features.

Humans build viewpoint-independent cognitive maps through navigation, enabling intuitive reasoning about object permanence and spatial relations. We argue that multimodal large language models (MLLMs), despite extensive video training, lack this fundamental spatial reasoning capability, a critical limitation for embodied applications. To demonstrate these limitations and drive research, we introduce REM: Reasoning over Embodied Multi-Frame Trajectories, a benchmark using controllable 3D environments for long-horizon embodied spatial reasoning. REM systematically evaluates key aspects like object permanence/distinction, spatial relationships, and numerical tracking across dynamic embodied viewpoints. Our evaluation shows that the best-performing current models exhibit promising overall performance, but become increasingly unreliable at even moderate complexity levels easily handled by humans. These findings highlight challenges MLLMs face in developing robust spatial representations from sequential visual input. Consequently, REM provides targeted metrics and diagnostics to foster improved spatial understanding in future models.

Large Language Models (LLMs) have shown remarkable performance across a wide range of natural language processing tasks. Quality Estimation (QE) for Machine Translation (MT), which assesses the quality of a source-target pair without relying on reference translations, remains a challenging cross-lingual task for LLMs. The challenges stem from the inherent limitations of existing LLM-based QE systems, which are pre-trained for causal language modelling rather than regression-specific tasks, further elevated by the presence of low-resource languages given pre-training data distribution. This paper introduces ALOPE, an adaptive layer-optimization framework designed to enhance LLM-based QE by restructuring Transformer representations through layer-wise adaptation for improved regression-based prediction. Our framework integrates low-rank adapters (LoRA) with regression task heads, leveraging selected pre-trained Transformer layers for improved cross-lingual alignment. In addition to the layer-specific adaptation, ALOPE introduces two strategies—dynamic weighting, which adaptively combines representations from multiple layers, and multi-head regression, which aggregates regression losses from multiple heads for QE. Our framework shows improvements over various existing LLM-based QE approaches. Empirical evidence suggests that intermediate Transformer layers in LLMs provide contextual representations that are more aligned with the cross-lingual nature of the QE task. We make resultant models and framework code publicly available for further research, also allowing existing LLM-based MT frameworks to be scaled with QE capabilities.

Language provides a natural interface to specify and evaluate performance on visual tasks. To realize this possibility, vision language models (VLMs) must successfully integrate visual and linguistic information. Our work compares VLMs to a direct readout of their visual encoders to understand their ability to integrate across these modalities. Across a series of vision-centric benchmarks (e.g., depth estimation, correspondence), we find that VLMs perform substantially worse than their visual encoders, dropping to near-chance performance. We investigate these results through a series of analyses across the entire VLM: namely 1) the degradation of vision representations, 2) brittleness to task prompt, and 3) the language model's role in solving the task. We find that the bottleneck in performing these vision-centric tasks lies in this third category; VLMs are not effectively using visual information easily accessible throughout the \textit{entire} model, and they inherit their language biases. Our work helps diagnose the failure modes of open-source VLMs, and presents a series of evaluations useful for future investigations into visual understanding within VLMs.

Vision-language models encode images and text in a joint space, minimizing the distance between corresponding image and text pairs. How are language and images organized in this joint space, and how do the models encode meaning and modality? To investigate this, we train and release sparse autoencoders (SAEs) on the embedding spaces of four vision-language models (CLIP, SigLIP, SigLIP2, and AIMv2). SAEs approximate model embeddings as sparse linear combinations of learned directions, or ``concepts''. We find that, compared to other methods of linear feature learning, SAEs are better at reconstructing the real embeddings, while also able to retain the most sparsity. Retraining SAEs with different seeds or different data diet leads to two findings: the rare, specific concepts captured by the SAEs are liable to change drastically, but we also show that commonly-activating concepts are remarkably stable across runs. Interestingly, while most concepts activate primarily for one modality, we find they are not merely encoding modality per se. Many are almost orthogonal to the subspace that defines modality, and the concept directions do not function as good modality classifiers, suggesting that they encode cross-modal semantics. To quantify this bridging behavior, we introduce the Bridge Score, a metric that identifies concept pairs which are both co-activated across aligned image-text inputs and geometrically aligned in the shared space. This reveals that even single-modality concepts can collaborate to support cross-modal integration. We release interactive demos of the SAEs for all models, allowing researchers to explore the organization of the concept spaces. Overall, our findings uncover a sparse linear structure within VLM embedding spaces that is shaped by modality, yet stitched together through latent bridges—offering new insight into how multimodal meaning is constructed.