AI News from BAIR Blog (Berkeley AI Research)

Latest coverage from BAIR Blog (Berkeley AI Research), summarized and scored for signal.

• Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction — Berkeley AI Research introduces ABBEL, a framework that replaces recursive context summarization with supervised belief states for long-horizon LLM agents — closing the performance gap that persists in production systems like Cursor's composer 2.5 even after RL fine-tuning.
• Intelligence is Free, Now What? <br> Data Systems for, of, and by Agents — A Berkeley AI Research perspective argues AI inference costs have fallen 50x per year (median) — from $30/M tokens in early 2023 to under $1 today — and calls for rebuilding data systems around three new realities: agents as the dominant workload, coordinating agent swarms, and agents synthesizing entire data systems themselves.
• 2026 BAIR Graduate Showcase — Berkeley's BAIR Lab announces its 2026 PhD graduating class, with researchers heading to faculty positions, industry labs (including Physical Intelligence), and startups across robotics, LLM reasoning, computer vision, and AI safety.
• Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Gradient-based Planning for World Models at Longer Horizons — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Identifying Interactions at Scale for LLMs — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Information-Driven Design of Imaging Systems — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• RL without TD learning — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• What exactly does word2vec learn? — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Whole-Body Conditioned Egocentric Video Prediction — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Defending against Prompt Injection with Structured Queries (StruQ) and Preference Optimization (SecAlign) — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Repurposing Protein Folding Models for Generation with Latent Diffusion — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Scaling Up Reinforcement Learning for Traffic Smoothing: A 100-AV Highway Deployment — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Virtual Personas for Language Models via an Anthology of Backstories — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Linguistic Bias in ChatGPT: Language Models Reinforce Dialect Discrimination — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• How to Evaluate Jailbreak Methods: A Case Study with the StrongREJECT Benchmark — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Are We Ready for Multi-Image Reasoning? Launching VHs: The Visual Haystacks Benchmark! — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.