What would you like to explore?
Browse the full archive →Trending papers
Most discussed on X · last 14 days
Measuring AI Agents as Teammates with Collaborative Intelligence
This paper introduces the Collaborative Intelligence Benchmark to assess AI agents as individual teammates, revealing that even capable models often recognize collaboration problems but fail to resolve them.
Exo: An Architecture for Self-Improving Systems at Runtime
This paper introduces Exo, a guarded system that lets an AI improve the routines guiding its work as it goes, often making repeated tasks cheaper without changing the model itself.
Contextual Embeddings Beyond the Gold Chunk
This paper teaches document embeddings to find useful evidence across context-dependent passages, rather than treating one “gold” chunk as the only right answer, improving retrieval without extra steps at search time.
Hinted Scaling Laws
This paper shows that gradually hinting models through difficult benchmarks turns near-zero scores into a useful forecasting signal, cutting prediction error from 21 to 4.3 percentage points.
“There may be no single “correct” ranking of weak models: the ranking can depend on how capability is measured. On Humanity’s Last Exam, models’ ranking agreement between average accuracy and pass@k (success at least once across k attempts) fell from Kendall’s τ = 0.62 at k = 5 to 0.05 at k = 50.”
Hinted Scaling LawsDaily Publications
Last window Claude cracked the Switch 2. This window Anthropic reportedly barred Claude Code from recreating existing video games. Original games and games in existing genres are still allowed. Dexerto's post on the rule drew 12,111 likes, and another report drew 15,993. Then someone showed Claude r...