By carefully coordinating off-the-shelf models with inference only, we show the DSP framework can achieve surprisingly good results in theorem proving, comparable to the frontier models with RL-based large-scale training.

We propose two techniques to improve the data efficiency of LLM RL fine-tuning: difficulty-targeted online data selection and rollout replay.

A hybrid architecture with linear pre-filling complexity and up-to10x higher throughput on decoding.

We present BEAST, a novel action tokenizer that encodes action sequences into compact discrete or continuous tokens using B-spline.

Novel systematic cognitive reasoning evaluation of VLMs, providing clues for the sources of reasoning bottlenecks, as well as simple and effective solutions.

NaDRO employs unique preference and context-based rewards to effectively train LLMs on noisy data, enabling even smaller models to achieve superior performance in complex decision-making tasks.

We introduce a principled framework for validating LLM-as-a-judge systems under rating indeterminacy, where multiple ratings can be "correct."

We propose Reward Reasoning Models, which leverage additional test-time compute for complex queries where appropriate rewards are not immediately apparent.

We show that the statistical optimality of estimation methods for causal inference depend in a surprising way on the distribution of the treatment noise.

We design the first efficient, near-optimal regret algorithm for contextual dueling bandits using offline oracles, enabling scalable preference-based learning in RLHF and resolving a key open problem in AI alignment.

We provide a theretical analysis on a linear version of LAM, which is recently popularized in embodied AI

We theoretically and empirically analyze the effect of the Pooling Operation in the downstream performance.

We propose EAGLE-3, observing that it can benefit from scaling up the data.

We adopt reinforcement learning to train LLMs to generate quality code with rewards derived from program analysis.

A report of the Big ANN competition in NeurIPS'23

We create a perturbationally-paired single cell imaging and transcriptomics dataset and benchmark and evaluate unimodal, multimodal, and multisample imaging and scRNA-seq models for treatment response modeling; HF data access token in zip.

We propose SAS to simulate larger attention head numbe and hidden size per head for better performance, keeping the original model size.

We propose Flex-Judge, a multimodal judge model trained solely on a small corpus of high-quality text reasoning data.

We introduce a new design principle for LLM matrix optimizers - gradient multi-normalization, unifying previous work, and enabling faster and memory-efficient training of LLMs.

BC-LLM is a novel procedure that integrates Large Language Models into a Bayesian framework for concept discovery to achieve better predictive performance, converge faster to relevant concepts, and provide rigorous uncertainty quantification.

We propose a linear-time efficient and equivariant Transformer for scalable molecular modeling.

We give improved approximation algorithms for fair monotone matroid submodular maximization.

We propose a contextualized position encoding using dynamic Householder matrices in place of static rotary ones, along with a hardware-efficient training algorithm that improves state tracking performance.

We develop a theoretical model of information flow constraints in LLMs that predicts LLM failures on global reasoning tasks.

A foundation model specifically designed for medical time series forecasting

We propose a new method that reduce the sampling steps of image auto-regressive models to only 1

We introduce a new threat model in offensive cybersecurity agent, and provide evaluation framework that can dynamically assess the risk within a fixed compute budget.

We introduce a new comprehensive benchmark, MMTU, designed to evaluate models ability to understand, reason, and manipulate diverse tables.

We propose Pseudo-Zeroth-Order (PseuZO), a framework that estimates Jacobian matrices via model output differentiation and applies EMA for variance reduction, with theoretical convergence guarantees and empirical validation.

We rethink LLMs from the perspective of recommender systems and propose Language System, which exploits the model’s output distribution more efficiently and effectively.

We theoretically study the effect of the pooling operation on adversarial robustness, which allows us to introduce RS-Pool, a new pooling method capable of smoothing the effect of attacks.

SPRINT trains large reasoning models to dynamically spot and run independent subtasks in parallel, reducing sequential tokens by up to 40% on complex tasks like math without sacrificing accuracy compared to sequential reasoning baselines.

We extend DeltaNet by using products of householders as state-transition matrices allowing us to trade-off expressivity and computational complexity.

We design relative positional encoding methods for spiking Transformers

A Unified Principle for Generative Modeling, Representation Learning, and Classification

We find that excessively scaling Chain of Thought (CoT) length can impair the model's reasoning performance in certain domains, and we propose a Thinking-Optimal Scaling strategy to achieve more effective and efficient test-time scaling.

We propose GUI-Actor, a VLM-based, coordinate-free GUI grounding method with an attention-based action head and verifier, achieving state-of-the-art results and strong generalization.

Rigor in AI remains largely understood in terms of methodological rigor. This common, yet narrow conceptualization of rigor has contributed to many responsible AI concerns. Here, we argue for a broader view of what rigorous AI research should entail.

We provide a scalable bandit architecture for prompt tuning of decision transformers for increased downstream performance.

We propose Point-RFT, a multimodal framework using visually grounded Chain-of-Thought reasoning with two-stage finetuning, which exhibits superior generalization capability and potentials in complex real-world scenarios.

We investigate last-iterate convergence of Regret Matching$^+$ variants in games satisfying the weak Minty variation inequality.

We introduce the Generalized Induction Model (GIM), a retrieval-based in-context module that enhances interpretable next-token prediction in language modeling and fMRI response prediction.

To resolve long video understanding task, we propose Deep Video Discovery agent to iterative reasoning and gather information from video content via an agentic search and tool use strategy..

We propose Hierarchical Diffusion Language Models, which is a discrete diffusion with a general time-varying next semantic scale prediction process for language modeling.

This paper introduces an approach for training o1-like RAG models that retrieve and reason over relevant information step by step before generating the final answer.

A reinforcement-learning post-training framework teaches LLM assistants to reason about contextual integrity, slashing inappropriate information disclosure while helping users complete their tasks.

We extend a new RL-based gradient-informed finetuning method to the task of reward finetuning/alignment for 3D native diffusion models.

We propose a cognitive-science-inspired framework and benchmark to systematically evaluate learning abilities of large language models.

We improve and speed up training-free conditional sampling from diffusion models using an approximation to the Newton method

We argue that there are important gaps between what unlearning methods can do and claims about/hopes for what these methods can do for broader impact.

We present an optimistic query routing protocol in clustering-based approximate maximum inner product search that reaches the same accuracy as SoTA routers by probing 50% fewer data points.

We perform large-scale strong MIAs on pre-trained LLMs to clarify the extent of actual privacy risk MIAs pose in this setting.

We show that transformers achieve length generalization when training on shorter main task and longer auxiliary tasks together.

SCFG-based framework characterizes uncertainty in LLM formal reasoning, enabling selective verification that reduces errors while minimally abstaining.

We present the first parameter-free last-iterate convergence of Counterfactual Regret Minimization algorithms.

Achieve Pluralistic Value Alignment through Counterfactual Reasoning

We propose a 3D full-body pose and cooking videos dataset along with multimodal behavior understanding benchmarks.

a genomic sequence model that unifies long-context reasoning and sequence interpretation with state-of-the-art performance.

We propose CADMorph, an inference-time editing method for parametric CAD models using the signal of geometry changes via a Plan–Generate–Verify Loop over pretrained priors, namely a LDM and an LLM, bypassing the need of non-existent editing data.

We propose a general reinforcement learning framework tailored for interleaved multimodal tasks by permutating image sequences to simulate varied positional relationships and explore more spatial and positional diversity

We introduce rStar-Coder to train advanced code reasoning LLMs, with our 14B model achieving comparable performance to QWQ-32B.

We only need one example for RLVR on LLMs to achieve significant improvement on math tasks

We show that Schedule-Free methods effectively navigate the river structure of the loss landscape, enabling scalable language model training without decay schedules or extra memory.

RetrievalAttention improves decoding speed and reduces GPU memory usage in Transformer-based LLMs by using pre-built, attention-aware KV vector indexes stored in CPU memory, achieving significant efficiency gains without compromising accuracy.

We develop trust region methods for stochastic optimal control to improve sampling from unnormalized densities, transition path sampling, and diffusion model finetuning.

We propose a value gradient matching formulation for reward finetuning/alignment for flow matching models with the theory of optimal control, and empirically verify our method on the popular text-to-image flow matching model StableDiffusion3

We argue that conclusions drawn about relative system safety or attack method efficacy via AI red teaming are often not supported by evidence provided by attack success rate (ASR) comparisons.