Engineering
Thought-Level Beam Search for Reasoning
This paper targets inefficient test-time compute scaling in large reasoning models, where independent sampling wastes resources and naive pruning fails…
Engineering
SKILLER: Language-Level Reinforcement Learning for Reusable Skill Extraction in Small Language Models
SKILLER addresses the model-mismatch problem, where reusable skills designed around powerful frontier models fail when deployed on smaller, cheaper language…
Engineering
Maglev: Sliding Recurrent Memory
Maglev proposes a way for transformers to maintain persistent, token-wise memory while keeping both inference cost bounded and training parallelizable.…
Engineering
LycheeMemory V2: Efficient Long-Term Memory for LLM Agents via Semantic Segment-Level Consolidation
LycheeMemory V2 targets the cost of maintaining long-term memory in long-horizon LLM agents, replacing conventional turn-by-turn memory consolidation with semantic…
Engineering
Knowing When to Quit: Diagnosing and Training LLMs to Abort Futile Reasoning
This paper studies why large language models keep generating long, expensive reasoning chains on problems beyond their real capability, producing…
Engineering
Specification-First Convergence with an AI Coding Agent
This case study reports on using an AI coding agent to carry out a large-scale architectural refactor of a 717,000-line…
Engineering
How Can Rhetoric Reward-Hack AI Reviewers? Dissecting Rhetorical Sensitivity in AI-Based Peer Review
This study investigates whether purely rhetorical changes to a scientific manuscript — with the underlying content held constant — can…
Engineering
Intern-S2-Preview: Scientific Agentic Foundation Model
Intern-S2-Preview is a family of foundation models built for multimodal scientific reasoning and long-horizon agentic tasks. Its training pipeline combines…
Engineering
Extreme Overtraining in Tiny Language Models
This write-up documents a training experiment pushing a 0.9-million-parameter language model to a token-per-parameter ratio of roughly 222,000:1 — far…
Engineering
Cloudflare Introduces Kitesurf, an Agent-First Browser That Runs Inside Workers
Cloudflare introduced Kitesurf, a browser engine built from scratch in Rust and compiled to WebAssembly to run entirely inside Cloudflare…
Engineering
How Cloudflare Built Its Internal AI Engineering Stack on Its Own Platform
Cloudflare detailed the internal AI engineering infrastructure it built for its own workforce of 3,683 employees, using its own products…
Engineering
Circuit Tracing for the Rest of Us: From Probes to Attribution Graphs
Engineer Subhadip Mitra lays out a hierarchical framework for understanding AI model internals, moving from behavioral output analysis through attention…