COLM 2025 Papers

We present Plancraft, a multi-modal evaluation dataset for LLM agents. Plancraft has both a text-only and multi-modal interface, based on the Minecraft crafting GUI. We include the Minecraft Wiki to evaluate tool use and Retrieval Augmented Generation (RAG), as well as a handcrafted planner and Oracle Retriever, to ablate the different components of a modern agent architecture. To evaluate decision-making, Plancraft also includes a subset of examples that are intentionally unsolvable, providing a realistic challenge that requires the agent not only to complete tasks but also to decide whether they are solvable at all. We benchmark both open-source and closed-source LLMs and compare their performance and efficiency to a handcrafted planner. Overall, we find that LLMs and VLMs struggle with the planning problems that Plancraft introduces, and offer suggestions on how to improve their capabilities.

Language model developers typically filter out high-risk content—such as toxic or copyrighted text—from their pre-training data to prevent models from generating similar outputs. However, removing such data altogether limits models’ ability to recognize and appropriately respond to harmful or sensitive content. In this paper, we introduce Selective Loss to Understand but Not Generate (SLUNG), a pre-training paradigm through which models learn to understand high-risk data without learning to generate it. Instead of uniformly applying the next-token prediction loss, SLUNG selectively avoids incentivizing the generation of high-risk tokens while ensuring they remain within the model's context window. As the model learns to predict low-risk tokens that follow high-risk ones, it is forced to understand the high-risk content. Through our experiments, we show that SLUNG consistently improves models' understanding of high-risk data (e.g., ability to recognize toxic content) without increasing its generation (e.g., toxicity of model responses). Overall, our SLUNG paradigm enables models to benefit from high-risk text that would otherwise be filtered out.

Watermarking has emerged as a promising technique for detecting texts generated by LLMs. Current research has primarily focused on three design criteria -- high quality of the watermarked text, high detectability, and robustness against removal attack. However, the security against spoofing attacks remains relatively understudied. For example, a piggyback attack can maliciously alter the meaning of watermarked text by transforming it into hate speech, while preserving the original watermark, thereby damaging the reputation of the LLM provider. We identify two core challenges that make defending against spoofing difficult: (1) the need for watermarks to be both sensitive to semantic-distorting changes and insensitive to semantic-preserving edits, and (2) the contradiction between the need to detect global semantic shifts and the local, auto-regressive nature of most watermarking schemes. To address these challenges, we propose a semantic-aware watermarking algorithm that post-hoc embeds watermarks into a given target text while preserving its original meaning. Our method introduces a semantic mapping model, which guides the generation of a green-red token list, contrastively trained to be sensitive to semantic-distorting changes and insensitive to semantic-preserving changes. Experiments on two standard benchmarks demonstrate strong robustness against removal attacks and security against spoofing attacks, including sentiment reversal and toxic content insertion, while maintaining high watermark detectability. Our approach offers a significant step toward more secure and semantically aware watermarking for LLMs.

Multimodal Large Language Models (MLLMs), built on powerful language backbones, have enabled Multimodal In-Context Learning (MICL)—adapting to new tasks from a few multimodal demonstrations consisting of images, questions, and answers. Despite showing noticeable improvement on standard vision-language datasets, current MLLMs struggle to leverage visual information in the demonstrations. Specifically, they tend to neglect visual cues and over-rely on textual patterns, leading to mere text imitation rather than genuine multimodal adaptation. This behavior makes MICL still unimodal and largely restricts its practical utility. More importantly, this limitation is often concealed by the improved performance on tasks that do not require understanding the visual context. As a result, how to effectively enhance MICL ability and reliably evaluate the MICL performance remains underexplored. To address these issues, we first introduce Dynamic Attention Reallocation (DARA), an efficient fine-tuning strategy that encourages models to attend to the visual context by rebalancing attention across visual and textual tokens. In addition, we present TrueMICL, an MICL-dedicated dataset with both support and test sets that explicitly requires the integration of multimodal information—particularly visual content—for correct task completion. Extensive experiments demonstrate the effectiveness of our holistic solution, showcasing substantial improvements in the MICL capabilities. Code and datasets are available at [here](https://chenxshuo.github.io/true-micl-colm/).

Sliding-window attention offers a hardware-efficient solution to the memory and throughput challenges of Large Language Models (LLMs) in long-context scenarios. Existing methods typically employ a single window length across all attention heads and input sizes. However, this uniform approach fails to capture the heterogeneous attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose *Mixture of Attention Spans* (MoA), which automatically tailors distinct sliding-window length configurations to different heads and layers. MoA constructs and navigates a search space of various window lengths and their scaling rules relative to input sizes. It profiles the model, evaluates potential configurations, and pinpoints the optimal length configurations for each head. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer inputs, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by 3.9× with the same average sliding-window length, boosting retrieval accuracy by 1.5-7.1× over the uniform-window baseline across Vicuna-{7B,13B}, and Llama3-{8B,70B} models. Moreover, MoA narrows the performance gap with full attention, reducing the maximum relative performance drop from 9%-36% to within 5% across three long-context understanding benchmarks. MoA achieves a 1.2-1.4× GPU memory reduction, boosting decode throughput by 6.6-8.2× and 1.7-1.9× over FlashAttention2 and vLLM, with minimal performance impact. Our code is available at https://github.com/thu-nics/MoA.

Language model (LM) benchmarking faces several challenges: comprehensive evaluations are costly, benchmarks often fail to measure the intended capabilities, and evaluation quality can degrade due to labeling errors and benchmark saturation. Although various strategies have been proposed to mitigate these issues, they tend to address individual aspects in isolation, neglecting broader questions about overall evaluation quality. Here, we introduce Fluid Benchmarking, a new evaluation approach that advances LM benchmarking across multiple dimensions. Inspired by psychometrics, Fluid Benchmarking is based on the insight that the relative value of benchmark items depends on an LM's capability level, suggesting that evaluation should adapt to each LM. Methodologically, Fluid Benchmarking estimates an item response model based on existing LM evaluation results and uses the inferred quantities to select evaluation items dynamically, similar to computerized adaptive testing in education. In our experiments, we compare Fluid Benchmarking against the common practice of random item sampling as well as more sophisticated baselines, including alternative methods grounded in item response theory. We examine four dimensions—efficiency, validity, variance, and saturation—and find that Fluid Benchmarking achieves superior performance in all of them (e.g., higher validity and less variance on MMLU with fifty times fewer items). Our analysis shows that the two components of Fluid Benchmarking have distinct effects: item response theory, used to map performance into a latent ability space, increases validity, while dynamic item selection reduces variance. Overall, our results suggest that LM benchmarking can be substantially improved by moving beyond static evaluation.

Large Language Models (LLMs) exhibit significant disparities in performance across languages, primarily benefiting high-resource languages while marginalizing underrepresented ones. Continual Pretraining (CPT) has emerged as a promising approach to address this imbalance, although the relative effectiveness of monolingual, bilingual, and code-augmented data strategies remains unclear. This study systematically evaluates 36 CPT configurations involving three multilingual base models, across 30+ languages categorized as altruistic, selfish, and stagnant, spanning various resource levels. Our findings reveal three major insights: (1) Bilingual CPT improves multilingual classification but often causes language mixing issues during generation. (2) Including programming code data during CPT consistently enhances multilingual classification accuracy and language modeling capabilities, particularly benefiting low-resource languages, but introduces a trade-off by slightly degrading generation quality. (3) Contrary to prior work, we observe substantial deviations from language classifications according to their impact on cross-lingual transfer: Languages classified as altruistic often negatively affect related languages, selfish languages show conditional and configuration-dependent behavior, and stagnant languages demonstrate surprising adaptability under certain CPT conditions. These nuanced interactions emphasize the complexity of multilingual representation learning, underscoring the importance of systematic studies on generalizable language classification to inform future multilingual CPT strategies.

Reasoning and strategic behavior in social interactions is a hallmark of intelligence. This form of reasoning is significantly more sophisticated than isolated planning or reasoning tasks in static settings (e.g., math problem solving). In this paper, we present Strategic Planning, Interaction, and Negotiation (SPIN-Bench), a new multi-domain evaluation designed to measure the intelligence of strategic planning and social reasoning. While many existing benchmarks focus on narrow planning or single-agent reasoning, SPIN-Bench combines classical PDDL tasks, competitive board games, cooperative card games, and multi-agent negotiation scenarios in one unified framework. The framework includes both a benchmark as well as an arena to simulate and evaluate the variety of social settings to test reasoning and strategic behavior of AI agents. We formulate the benchmark SPIN-Bench by systematically varying action spaces, state complexity, and the number of interacting agents to simulate a variety of social settings where success depends on not only methodical and step-wise decision making, but also conceptual inference of other (adversarial or cooperative) participants. Our experiments reveal that while contemporary LLMs handle basic fact retrieval and short-range planning reasonably well, they encounter significant performance bottlenecks in tasks requiring deep multi-hop reasoning over large state spaces and socially adept coordination under uncertainty. We envision SPIN-Bench as a catalyst for future research on robust multi-agent planning, social reasoning, and human--AI teaming.

LLMs have demonstrated remarkable performance across various tasks but face challenges related to unintentionally generating outputs containing sensitive information. A straightforward approach to address this issue is to retrain the model after excluding the problematic data. However, this approach incurs prohibitively high computational costs. To overcome this limitation, machine unlearning has emerged as a promising solution that can effectively remove sensitive information without the need to retrain the model from scratch. Recently, FILA has been proposed as a parameter-efficient unlearning method by integrating LoRA adapters. Specifically, it calculates the Fisher information to identify parameters associated with the forget set and assigns them to LoRA adapters for updates. Despite its innovative approach, FILA still requires access to all model parameters and does not adequately account for fundamental assumptions underlying Fisher information, leading to inaccuracies in importance estimation. To address these limitations, we propose VILA, a novel unlearning framework that explicitly considers the assumptions overlooked in FILA, thereby enhancing the accuracy of parameter identification for the forget set. Moreover, VILA significantly reduces computational costs by enabling parameter identification without accessing the entire model. Our method achieves up to 100× higher parameter efficiency and 40× faster training speed compared to FILA, and sets new state-of-the-art performance on benchmarks including TOFU, WMDP, and MUSE. Our code is available at https://github.com/kyj93790/VILA.

The rapid deployment of Large Language Models (LLMs) highlights the need for efficient low-bit post-training quantization (PTQ) due to their high memory costs. A key challenge in weight quantization is the presence of outliers, which inflate quantization ranges and lead to large errors. While a number of outlier suppression techniques have been proposed, they either: fail to effectively shrink the quantization range, or incur (relatively) high bit overhead. In this paper, we present ICQuant, a novel framework that leverages outlier statistics to design an efficient index coding scheme for outlier-aware weight-only quantization. Compared to existing outlier suppression techniques requiring $\approx 1$ bit overhead to halve the quantization range, ICQuant requires only $\approx 0.25$ bits; a significant saving in low bit regimes (e.g., 2-3 bits). ICQuant can be used on top of any existing quantizers to eliminate outliers, improving the quantization quality. Using just 2.3 bits per weight and simple scalar quantizers, \ours improves the zero-shot accuracy of the 2-bit Llama3-70B model by up to 130\% and 150\% relative to QTIP (Tseng et al. (2024b)) and QuIP\# (Tseng et al. (2024a) respectively; and it achieves comparable performance to the best-known fine-tuned quantizer (Malinovskii et al. (2024)) without any fine-tuning.

Modern large language models often encode sensitive, harmful, or copyrighted knowledge, raising the need for post-hoc unlearning—the ability to remove specific domains of knowledge from a model without full retraining. A major bottleneck in current unlearning pipelines is constructing effective forget sets—datasets that approximate the target domain and guide the model to forget it. In this work, we introduce a scalable, automated approach to generate high-quality forget sets using language models themselves. Our method synthesizes textbook-style data through a structured prompting pipeline, requiring only a domain name as input. Through experiments on unlearning biosecurity, cybersecurity, and Harry Potter novels, we show that our synthetic datasets consistently outperform the baseline synthetic alternatives and are comparable to the expert-curated ones. Additionally, ablation studies reveal that the multi-step generation pipeline significantly boosts data diversity, which in turn improves unlearning utility. Overall, our findings suggest that synthetic datasets offer a promising path toward practical, scalable unlearning for a wide range of emerging domains without the need for manual intervention. We release our code and dataset at [https://github.com/xyzhu123/Synthetic_Textbook](https://github.com/xyzhu123/Synthetic_Textbook).

Visual Language Models (VLMs) have become foundational models for document understanding tasks, widely used in the processing of complex multimodal documents across domains such as finance, law, and academia. However, documents often contain noise-like information, such as watermarks, which inevitably leads us to inquire: Do watermarks degrade the performance of VLMs in document understanding? To address this, we propose a novel evaluation framework to investigate the effect of visible watermarks on VLMs performance. We takes into account various factors, including different types of document data, the positions of watermarks within documents and variations in watermark content. Our experimental results reveal that VLMs performance can be significantly compromised by watermarks, with performance drop rates reaching up to 36\%. We discover that \emph{scattered} watermarks cause stronger interference than centralized ones, and that \emph{semantic contents} in watermarks creates greater disruption than simple visual occlusion. Through attention mechanism analysis and embedding similarity examination, we find that the performance drops are mainly attributed to that watermarks 1) force widespread attention redistribution, and 2) alter semantic representation in the embedding space. Our research not only highlights significant challenges in deploying VLMs for document understanding, but also provides insights towards developing robust inference mechanisms on watermarked documents.

Existing LLM agent systems typically select actions from a fixed and predefined set at every step. While effective in closed, narrowly scoped environments, this approach presents two major challenges for real-world, open-ended scenarios: (1) it significantly restricts the planning and acting capabilities of LLM agents, and (2) it requires substantial human effort to enumerate and implement all possible actions, which is impractical in complex environments with a vast number of potential actions. To address these limitations, we propose an LLM agent framework that enables the dynamic creation and composition of actions in an online manner. In this framework, the agent interacts with its environment by generating and executing programs written in a general-purpose programming language. Furthermore, generated actions are accumulated over time for future reuse. Our extensive experiments across multiple benchmarks demonstrate that this framework significantly improves flexibility and outperforms prior methods that rely on a fixed action set. Notably, it enables LLM agents to adapt and recover in scenarios where predefined actions are insufficient or fail due to unforeseen edge cases.

To succeed in common digital tasks such as web navigation, agents must carry out a variety of specialized tasks such as searching for products or planning a travel route. To tackle these tasks, agents can bootstrap themselves by learning task-specific skills online through interaction with the web environment. In this work, we demonstrate that programs are an effective representation for skills. We propose agent skill induction (ASI), which allows agents to adapt themselves by inducing, verifying, and utilizing program-based skills on the fly. We start with an evaluation on the WebArena agent benchmark and show that ASI outperforms the static baseline agent and its text-skill counterpart by 23.5% and 11.3% in success rate, mainly thanks to the programmatic verification guarantee during the induction phase. ASI also improves efficiency by reducing 10.7–15.3% of the steps over baselines, by composing primitive actions (e.g., click) into higher-level skills (e.g., search product). We then highlight the efficacy of ASI in remaining efficient and accurate under scaled-up web activities. Finally, we examine the generalizability of induced skills when transferring between websites, and find that ASI can effectively reuse common skills, while also updating incompatible skills to versatile website changes.

Mixture-of-Experts (MoE) Large Language Models (LLMs) suffer from severely sub-optimal expert pathways—our study reveals that naive expert selection learned from pretraining leaves a surprising 10-20% accuracy gap for improvement. Motivated by this observation, we develop a novel class of test-time optimization methods to re-weight or “re-mixing” the experts in different layers jointly for each test sample. Since the test sample’s ground truth is unknown, we propose to optimize a surrogate objective defined by the sample’s “successful neighbors” from a reference set of samples. We introduce three surrogates and algorithms based on mode-finding, kernel regression, and the average loss of similar reference samples/tasks. To reduce the cost of optimizing whole pathways, we apply our algorithms merely to the core experts’ mixing weights in critical layers, which enjoy similar performance but save significant computation. This leads to “Critical-Layer, Core-Expert, Collaborative Pathway Optimization (C3PO)”. We apply C3PO to two recent MoE LLMs and examine it on six widely-used benchmarks. It consistently improves the base model by 7-15% in accuracy and outperforms widely used test-time learning baselines, e.g., in-context learning and prompt/prefix tuning, by a large margin. Moreover, C3PO enables MoE LLMs with 1-3B active parameters to outperform LLMs of 7-9B parameters, hence improving MoE’s advantages on efficiency. Our thorough ablation study further sheds novel insights on achieving test-time improvement on MoE.

Scaling laws predict that the performance of large language models improves with increasing model size and data size. In practice, pre-training has been relying on massive web crawls, using almost all data sources publicly available on the internet so far. However, this pool of natural data does not grow at the same rate as the compute supply. Furthermore, the availability of high-quality texts is even more limited: data filtering pipelines often remove up to 99% of the initial web scrapes to achieve state-of-the-art. To address the "data wall" of pre-training scaling, our work explores ways to transform and recycle data discarded in existing filtering processes. We propose REWIRE, REcycling the Web with guIded REwrite, a method to enrich low-quality documents so that they could become useful for training. This in turn allows us to increase the representation of synthetic data in the final pre-training set. Experiments at 1B, 3B and 7B scales of the DCLM benchmark show that mixing high-quality raw texts and our rewritten texts lead to 1.0, 1.3 and 2.5 percentage points improvement respectively across 22 diverse tasks, compared to training on only filtered web data. Training on the raw-synthetic data mix is also more effective than having access to 2x web data. Through further analysis, we demonstrate that about 82% of the mixed in texts come from transforming lower-quality documents that would otherwise be discarded. REWIRE also outperforms related approaches of generating synthetic data, including Wikipedia-style paraphrasing, question-answer synthesizing and knowledge extraction. These results suggest that recycling web texts holds the potential for being a simple and effective approach for scaling pre-training data. We make our high-quality synthetic data publicly available at https://huggingface.co/datasets/facebook/recycling_the_web.

The assumption across nearly all language model (LM) tokenization schemes is that tokens should be subwords, i.e., contained within word boundaries. While providing a seemingly reasonable inductive bias, is this common practice limiting the potential of modern LMs? Whitespace is not a reliable delimiter of meaning, as evidenced by multi-word expressions (e.g., "by the way"), crosslingual variation in the number of words needed to express a concept (e.g., "spacesuit helmet" in German is "raumanzughelm"), and languages that do not use whitespace at all (e.g., Chinese). To explore the potential of tokenization beyond subwords, we introduce a "superword" tokenizer, SuperBPE, which incorporates a simple pretokenization curriculum into the byte-pair encoding (BPE) algorithm to first learn subwords, then superwords that bridge whitespace. This brings dramatic improvements in encoding efficiency: when fixing the vocabulary size to 200k, SuperBPE encodes a fixed piece of text with up to 33% fewer tokens than BPE on average. In experiments, we pretrain 8B transformer LMs from scratch while fixing the model size, vocabulary size, and train compute, varying *only* the algorithm for learning the vocabulary. Our model trained with SuperBPE achieves an average +4.0% absolute improvement over the BPE baseline across 30 downstream tasks (including +8.2% on MMLU), while simultaneously requiring 27% less compute at inference time. In analysis, we find that SuperBPE results in segmentations of text that are more uniform in per-token difficulty. Qualitatively, this may be because SuperBPE tokens often capture common multi-word expressions that function semantically as a single unit. SuperBPE is a straightforward, local modification to tokenization that improves both encoding efficiency and downstream performance, yielding better language models overall.

Personalized preference alignment for large language models (LLMs), the process of tailoring LLMs to individual users' preferences, is an emerging research direction spanning the area of NLP and personalization. In this survey, we present an analysis of works on personalized alignment and modeling for LLMs. We introduce a taxonomy of preference alignment techniques, including training time, inference time, and heuristic-driven methods. We provide analysis and discussion on the strengths and limitations of each group of techniques and then cover evaluation, benchmarks, as well as open problems in the field.

In-context learning is a remarkable capability of transformers, referring to their ability to adapt to specific tasks based on a short history or context. Previous research has found that task-specific information is locally encoded within models, though their emergence and functionality remain unclear due to opaque pre-training processes. In this work, we investigate the formation of task vectors in a controlled setting, using models trained from scratch on synthetic datasets. Our findings confirm that task vectors naturally emerge under certain conditions, but the tasks may be relatively weakly and/or non-locally encoded within the model. To promote strong task vectors encoded at a prescribed location within the model, we propose an auxiliary training mechanism based on a task vector prompting loss (TVP-loss). This method eliminates the need to search for task-correlated encodings within the trained model and demonstrably improves robustness and generalization.