Berkeley AI Research (BAIR)
berkeley-ai-research-bair--968badb1·7 events·first seen 1mo agoAliases: Berkeley AI Research (BAIR), BAIR (Berkeley AI Research), Berkeley AI Research
Co-occurring entities
More like this (12)
Recent events (7)
Scaling Up Reinforcement Learning for Traffic Smoothing: A 100-AV Highway Deployment
Berkeley AI Research (BAIR) deployed 100 RL-controlled autonomous vehicles into real rush-hour highway traffic on Interstate 24 near Nashville to dampen stop-and-go waves and reduce fuel consumption. The RL controllers were trained in data-driven simulations built from real highway trajectory data, using only local sensor inputs (speed, lead vehicle speed, gap) to enable decentralized deployment on standard vehicles. Reward design balanced wave smoothing, energy efficiency, safety, comfort, and adherence to human driving norms. The paper documents the sim-to-real transfer challenges encountered during this large-scale field experiment.
What exactly does word2vec learn? A closed-form theory of representation learning dynamics
Researchers from BAIR present a new theoretical paper proving that word2vec's learning dynamics reduce, under mild approximations, to unweighted least-squares matrix factorization, with final representations given by PCA on a specific co-occurrence-derived matrix. The theory solves gradient flow dynamics in closed form, showing that embeddings learn one orthogonal linear subspace (concept) at a time in discrete, rank-incrementing steps. This provides a quantitative, predictive account of the linear representation hypothesis observed in word2vec and, by extension, offers a minimal theoretical foundation for understanding feature learning in modern LLMs.
PEVA: Whole-Body Conditioned Egocentric Video Prediction for Embodied World Models
Researchers from BAIR introduce PEVA (Predicting Ego-centric Video from human Actions), a model that generates first-person video frames conditioned on 48-dimensional whole-body kinematic pose trajectories. The model uses an autoregressive conditional diffusion transformer trained on the Nymeria dataset, which pairs real-world egocentric video with body pose capture. PEVA can generate atomic action videos, simulate counterfactuals, and support long video generation, representing a step toward world models grounded in physically embodied human agents.
Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling
A BAIR blog post surveys recent progress in parallel reasoning for LLMs, covering methods from simple self-consistency and Best-of-N sampling through structured search (Tree of Thoughts, MCTS) to newer adaptive approaches including ParaThinker, GroupThink, and Hogwild! Inference. The core motivation is that sequential reasoning scales linearly with exploration depth, causing latency, context-rot, and compute inefficiency. Adaptive parallel reasoning aims to let models themselves decide when and how to decompose tasks into concurrent threads, rather than imposing fixed parallel structure externally. The post frames this as an emerging inference-time scaling paradigm with implications for agentic and complex reasoning workloads.
Defending against Prompt Injection with Structured Queries (StruQ) and Preference Optimization (SecAlign)
Researchers from BAIR propose two fine-tuning-based defenses against prompt injection attacks: StruQ (Structured Instruction Tuning) and SecAlign (Special Preference Optimization). Both methods use a Secure Front-End with special delimiter tokens to separate trusted prompts from untrusted data, then fine-tune LLMs to ignore injected instructions. SecAlign, which uses DPO-style preference optimization, reduces attack success rates to under 15% against strong optimization-based attacks—more than 4x better than prior SOTA—while preserving model utility on AlpacaEval2.
PLAID: Repurposing Protein Folding Models for Multimodal Protein Generation with Latent Diffusion
PLAID is a generative model that simultaneously produces protein 1D sequences and 3D all-atom structures by learning a diffusion model over the latent space of ESMFold, a protein folding model. It requires only sequence data for training—leveraging databases 2-4 orders of magnitude larger than structure databases—and decodes structure at inference via frozen folding model weights. The approach supports compositional prompting by function and organism, addressing practical drug-design constraints like humanization and solubility. A companion compression model, CHEAP, addresses the high-dimensionality of transformer latent spaces to make the diffusion training tractable.
RL without TD Learning: Divide-and-Conquer Value Learning for Long-Horizon Off-Policy RL
A BAIR blog post introduces a divide-and-conquer paradigm for off-policy reinforcement learning that avoids temporal difference (TD) learning's error accumulation problem by reducing Bellman recursions logarithmically rather than linearly. The approach leverages the triangle inequality structure of goal-conditioned RL to define a transitive Bellman update rule, enabling value learning that scales to long-horizon tasks. The authors claim this is the first practical realization of divide-and-conquer value learning at scale in goal-conditioned RL settings, building on an idea traceable to Kaelbling (1993). The post frames this as a third paradigm alongside TD and Monte Carlo methods, addressing a key gap in scalable off-policy RL.