dair-ai weekly picks: Declarative Attention tops the week's AI papers
dair_ai · x · 2026-09-06
dair-ai published its Top AI Papers of the Week (Aug 31 - Sep 6), featuring CORAL, WikiSkill, SKILL.state, E-Commerce Bench, Declarative Attention, Harness-of-Harness, and AI Research Preference Models.
Top pick Declarative Attention targets a core inefficiency: models read their entire KV cache for every generated token while only attending to a tiny slice of it—ask about one detail in a million-token conversation and nearly all cache reads are wasted. The paper proposes declaratively retrieving only the relevant attention spans to cut long-context inference cost.
Other entries span agent skill learning (WikiSkill, SKILL.state), an e-commerce benchmark (E-Commerce Bench), tool/harness orchestration (Harness-of-Harness), and preference models for AI research.
More from Research
- ByteX: ByteDance's vector search engine scales to 7,000 clusters and 300PB — _reachsumit · 2026-09-07
- Feldar AI is building a writing benchmark probing model naming and sentence quirks — almmaasoglu · 2026-09-06
- "Huge implications": binaries are now basically editable code — sergeyarl · 2026-09-06
- AlphaEvolve and SCALE: papers on coding agents for scientific and legal reasoning — ponguru · 2026-09-06
- Paper: Indirect prompt injection can compromise real-world LLM apps — ponguru · 2026-09-06
- Startup Mostik bridges models in latent space, tops ARC-AGI 3 leaderboard — emilahlback · 2026-09-06