Most popular benchmarks for comparing LLMs rely on a limited set of prompt templates, which may not fully capture the LLMs' abilities and can affect the reproducibility of results on leaderboards. Many recent works empirically verify prompt sensitivity and advocate for changes in LLM evaluation. In this paper, we consider the problem of estimating the performance distribution across many prompt variants instead of finding a single prompt to evaluate with. We introduce PromptEval, a method for estimating performance across a large set of prompts borrowing strength across prompts and examples to produce accurate estimates under practical evaluation budgets. The resulting distribution can be used to obtain performance quantiles to construct various robust performance metrics (e.g., top 95% quantile or median). We prove that PromptEval consistently estimates the performance distribution and demonstrate its efficacy empirically on three prominent LLM benchmarks: MMLU, BIG-bench Hard, and LMentry; for example, PromptEval can accurately estimate performance quantiles across 100 prompt templates on MMLU with a budget equivalent to two single-prompt evaluations. Moreover, we show how PromptEval can be useful in LLM-as-a-judge and best prompt identification applications.

Online shopping is a complex multi-task, few-shot learning problem with a wide and evolving range of entities, relations, and tasks. However, existing models and benchmarks are commonly tailored to specific tasks, falling short of capturing the full complexity of online shopping. Large Language Models (LLMs), with their multi-task and few-shot learning abilities, have the potential to profoundly transform online shopping by alleviating task-specific engineering efforts and by providing users with interactive conversations.

We introduce DataComp for Language Models, a testbed for controlled dataset experiments with the goal of improving language models.As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations.Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing atmodel scales ranging from 412M to 7B parameters.As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set.The resulting dataset, DCLM-Baseline, enables training a 7B parameter language model from scratch to 63% 5-shot accuracy on MMLU with 2T training tokens.Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6 percentage point improvement on MMLU while being trained with half the compute.Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation. We release the \dclm benchmark, framework, models, and datasets at https://www.datacomp.ai/dclm/

Large Language Models (LLMs) often fail on multiple-choice questions (MCQs) despite demonstrating correct knowledge in other contexts, such as free-form generation. To investigate the mechanism underlying this knowledge-prediction gap on MCQs and alleviate it, we conduct a probing analysis and find that residual streams in certain layers contain a subspace spanned by two important bases: a \emph{knowledge basis} that encodes the probability of the ground-truth answer for a given MCQ and a \emph{prediction basis} that encodes the probability of the answer choice predicted by the model. We observe that incorrect predictions arise from a misalignment of the model's hidden states along these two bases. Hence, we introduce \textbf{KAPPA} (Knowledge-Aligned Prediction through Projection-based Adjustment), a parameter-free intervention that transforms the hidden states to align the prediction coordinate with the knowledge coordinate within this subspace. Experiments on binary-choice reformulations of Big-Bench-Hard and ARC-Challenge show that KAPPA substantially improves accuracy and consistently outperforms baselines. While optimal subspaces differ across tasks, subspaces generalize to some extent, as supported by cross-dataset experiments. Moreover, KAPPA extends its effectiveness to free-form questions beyond MCQs. Our work provides a new geometric understanding of the knowledge-prediction gap and offers a practical method for better aligning model behavior with its latent knowledge.

User authorization-based access privileges are a key feature in many safety-critical systems, but have not been extensively studied in the large language model (LLM) realm. In this work, drawing inspiration from such access control systems, we introduce sudoLLM, a novel framework that results in multi-role aligned LLMs, i.e., LLMs that account for, and behave in accordance with, user access rights. sudoLLM injects subtle user-based biases into queries and trains an LLM to utilize this bias signal in order to produce sensitive information if and only if the user is authorized. We present empirical results demonstrating that this approach shows substantially improved alignment, generalization, resistance to prefix-based jailbreaking attacks, and ``fails-closed''. The persistent tension between the language modeling objective and safety alignment, which is often exploited to jailbreak LLMs, is somewhat resolved with the aid of the injected bias signal. Our framework is meant as an additional security layer, and complements existing guardrail mechanisms for enhanced end-to-end safety with LLMs.

Large language models learn from their vast pre-training corpora, gaining the ability to solve an ever increasing variety of tasks; yet although researchers work to improve these datasets, there is little effort to understand how efficient the pre-training apparatus is at extracting ideas and knowledge from the data. In this work, we use retrieval augmented generation along with test-time compute as a way to quantify how much dataset value was left behind by the process of pre-training, and how this changes across scale. We demonstrate that pre-training then retrieving from standard and largely open-sourced datasets results in significant accuracy gains in MMLU, Math-500, and SimpleQA, which persist through decontamination. For MMLU we observe that retrieval acts as a ~5x compute multiplier versus pre-training alone. We show that these results can be further improved by leveraging additional compute at test time to parse the retrieved context, demonstrating a 10 percentage point improvement on MMLU for the public LLaMA 3.1 8B model. Overall, our results suggest that today's pre-training methods do not make full use of the information in existing pre-training datasets, leaving significant room for progress.

Multimodal large language models (MLLMs) have shown strong capabilities but remain limited to fixed modality pairs and require costly fine-tuning with large aligned datasets. Building fully omni-capable models that can integrate text, images, audio, and video remains impractical and lacks robust reasoning support. In this paper, we propose an Agent-Omni framework that coordinates existing foundation models through a master-agent system, enabling flexible multimodal reasoning without retraining. The master agent interprets user intent, delegates subtasks to modality-specific agents, and integrates their outputs into coherent responses. Extensive experiments across text, image, audio, video, and omni benchmarks show that Agent-Omni consistently achieves state-of-the-art performance, particularly on tasks requiring complex cross-modal reasoning. Its agent-based design enables seamless integration of specialized foundation models, ensuring adaptability to diverse inputs while maintaining transparency and interpretability. In addition, the framework is modular and easily extensible, allowing future improvements as stronger models become available.

We introduce Context Tuning, a simple and effective method to significantly enhance few-shot adaptation of language models (LLMs) without fine-tuning model parameters. While prompt-based adaptation techniques have demonstrated the effectiveness of lightweight adaptation methods for LLMs, they typically initialize a trainable prompt or prefix with irrelevant tokens for the task at hand. In contrast, Context Tuning initializes the trainable prompt or prefix with task-specific demonstration examples, leveraging the model's inherent In-Context Learning (ICL) ability to extract relevant information for improved few-shot learning performance. Extensive evaluations on benchmarks such as CrossFit, UnifiedQA, MMLU, BIG-Bench Hard, and ARC demonstrate that Context Tuning outperforms traditional prompt-based adaptation methods and achieves competitive accuracy to Test-Time Training with significantly higher training efficiency.

Full fine-tuning of Large Language Models (LLMs) is notoriously memory-intensive, primarily because conventional optimizers such as SGD or Adam assume access to exact gradients derived from cached activations. Existing solutions either alter the model architecture (e.g., reversible networks) or trade memory for computation (e.g., activation checkpointing), but the optimizer itself remains untouched. In this work, we introduce GradLite, a backward-friendly optimizer that relaxes the requirement of exact gradients, enabling efficient training even when intermediate activations are aggressively discarded or approximated. GradLite leverages two key techniques: (i) low-rank Jacobian approximation, which reduces the dimensionality of backpropagated error signals, and (ii) error-feedback correction, which accumulates and compensates approximation errors across iterations to preserve convergence guarantees. We provide a theoretical analysis showing that GradLite maintains unbiased gradient estimates with bounded variance, ensuring convergence rates comparable to Adam. Empirically, GradLite reduces optimizer-state and activation memory consumption by up to 50\% without architectural changes, and achieves on-par or superior downstream performance on reasoning (MMLU, GSM8K), multilingual, and dialogue benchmarks compared to checkpointing and optimizer-centric baselines (LoMo, GaLore).

Modern language models are evaluated on large benchmarks, which are difficult to make sense of, especially for model selection. Looking at the raw evaluation numbers themselves using a model-centric lens, we propose SimBA, a three phase framework to Simplify Benchmark Analysis. The three phases of SimBA are: stalk, where we conduct dataset & model comparisons, prowl, where we discover a representative subset, and pounce, where we use the representative subset to predict performance on a held-out set of models. Applying SimBA to three popular LM benchmarks: HELM, MMLU, and BigBenchLite reveals that across all three benchmarks, datasets and models relate strongly to one another (stalk). We develop an representative set discovery algorithm which covers a benchmark using raw evaluation scores alone. Using our algorithm, we find that with 6.25% (1/16), 1.7% (1/58), and 28.4% (21/74) of the datasets for HELM, MMLU, and BigBenchLite respectively, we achieve coverage levels of at least 95% (prowl). Additionally, using just these representative subsets, we can both preserve model ranks and predict performance on a held-out set of models with near zero mean-squared error (pounce). Taken together, SimBA can help model developers improve efficiency during model training and dataset creators validate whether their newly created dataset differs from existing datasets in a benchmark. Our code is open source, available at https://github.com/nishantsubramani/simba.