Berkeley AI Research (BAIR)

Restricting LLM fine-tuning to the bottom singular subspace nearly eliminates catastrophic forgetting, even when protecting 99% of weights, while maintaining strong target-task performance.

Larger minibatch optimal transport couplings in flow matching, enabled by a novel entropic approximation, significantly boost training stability and generation performance.

Humanoid locomotion learned from scratch and transferred to hardware in 10x less time, thanks to a novel off-policy flow RL approach.

This work addresses the adverse interactions between sparsity and data repetition, and presents evidence for the core mechanisms of overfitting and its potential remediation, and suggests promising avenues for future methods to reduce over-specialization in model parameters by disrupting memorization patterns.

Continuous-time reinforcement learning can effectively tackle non-Markovian Hawkes processes, outperforming discrete methods in optimization tasks.

Retaining the right contextual information can boost long-context modeling performance by over 20% compared to traditional methods.

A novel retraction-free optimization method achieves global convergence on the Stiefel manifold, drastically improving LoRA fine-tuning efficiency for large language models.

FourTune slashes memory overhead by 2.25x while matching the performance of full-precision fine-tuning in diffusion models.

Muon achieves rapid convergence in matrix factorization by avoiding slow saddle dynamics and enabling high learning rates, aligning weights in just two steps.

Solving Blackwell approachability problems via Gradient Equilibrium oracles reveals a deep connection between these two seemingly distinct optimization frameworks.

Asynchronous OPD can boost training throughput significantly while managing the challenges of stale data, transforming the efficiency of large language model fine-tuning.

Transforming sparse rewards into dense feedback can accelerate RL training by significantly enhancing policy learning efficiency without compromising optimality.

QGF achieves superior performance in reinforcement learning by optimizing policies solely at test time, sidestepping the instability of traditional training methods.

Transformers can provably internalize chain-of-thought reasoning, matching the sample efficiency of explicit CoT while eliminating its inference overhead.

Forget hand-crafted examples: this system automatically generates worked examples tailored to student errors by mining common code patterns.

COMPASS outperforms traditional multilingual fine-tuning by effectively leveraging semantic gaps to enhance cross-lingual transfer and model adaptability.

Kernel launch overhead is a bigger bottleneck than you think: GPUOS achieves up to 15.3x speedup by fusing operations at runtime.

Unlock 36% better video depth estimation and 20% better camera pose estimation by simply letting your model learn from its own unlabeled video predictions.

Verifier-free evolution can now match or exceed the performance of verifier-based methods, while slashing API costs by 3x and boosting throughput by 10x, thanks to a clever model orchestration strategy.

MoEs can be pruned more effectively by considering cross-layer redundancy, leading to significant performance gains compared to uniform pruning strategies.

Scaling prompt learning by 17x without sacrificing accuracy is now possible, unlocking efficient self-improvement for LLM agents.

Forget temperature scaling: JUCAL calibrates aleatoric and epistemic uncertainty in classifier ensembles, achieving SOTA results with significantly smaller ensembles and lower inference costs.