Google DeepMindJun 12, 2025 Scaling Laws for Reasoning with Chain-of-Thought
J. Wei, A. Chowdhery, M. Chen
A systematic study of how test-time reasoning compute scales with model size and chain-of-thought length across diverse tasks.
Key Findings
- Test-time compute improves accuracy with predictable scaling behavior
- Reasoning gains are largest on math and coding benchmarks
- Smaller models can match larger ones given more inference compute
Stanford UniversityMay 28, 2025 Towards General-Purpose Multimodal Agents
L. Chen, S. Gu, R. Bommasani
A survey and empirical evaluation of agent architectures that combine vision, language and tool use for real-world tasks.
Key Findings
- Unified multimodal backbones outperform pipeline approaches
- Tool-use generalization remains a key bottleneck
- Benchmarks lag real-world agent difficulty
World Models for Embodied AI and Robotics
Y. LeCun, D. Hafner
Proposes learned world models that predict action consequences, enabling sample-efficient robot learning and planning.
Key Findings
- World models reduce real-world samples by an order of magnitude
- Latent dynamics enable long-horizon planning
- Sim-to-real transfer improves with predictive depth
A Unified Benchmark for Evaluating Frontier LLMs
D. Hendrycks, C. Burns
Introduces a standardized harness covering reasoning, coding, knowledge and safety across 40 tasks with contamination controls.
Key Findings
- Public benchmark saturation is widespread
- Contamination controls materially change rankings
- Human-verified evaluation remains essential
Scaling Diffusion Transformers for Video Generation
W. Peebles, T. Brooks
Describes the architecture behind text-to-video systems, scaling diffusion transformers to longer, higher-fidelity clips.
Key Findings
- Patch-based diffusion transformers scale well for video
- Temporal coherence improves with joint space-time attention
- Compute is the primary constraint on clip length
Constitutional AI and Scalable Oversight
Y. Bai, J. Kaplan
Advances methods for aligning models with human values using AI feedback and scalable oversight of superhuman outputs.
Key Findings
- AI feedback reduces reliance on human labels
- Scalable oversight catches subtle failures
- Constitutional approaches improve robustness
Retrieval-Augmented Generation at Production Scale
P. Lewis, S. Reddy
Lessons from deploying RAG systems to billions of queries, covering indexing, retrieval quality and hallucination control.
Key Findings
- Hybrid retrieval beats dense-only in production
- Citation grounding reduces hallucination
- Latency budgets shape architecture choices
Reinforcement Learning from AI Feedback at Scale
R. Zheng, J. Schulman
Compares RLHF and RLAIF for aligning large models, showing AI feedback can scale alignment without more human data.
Key Findings
- RLAIF matches RLHF on most axes
- Reward hacking remains a persistent risk
- Constitutional constraints reduce over-optimization