• A model's final layers quietly rewrite correct answers to sound more aligned. Early layers guess, middle layers sharpen the reasoning, but late layers drag the polished prediction toward "safer, more aligned" tokens. Confident Decoding skips the damage by decoding from a confident earlier layer — training-free, zero extra memory, under 2% latency, with steady gains on hard reasoning benchmarks.

• Enterprise agents finally have a benchmark built from real work. EnterpriseClawBench reconstructs 852 reproducible tasks with fixtures from actual workplace sessions. The strongest setup — Codex on GPT-5.5 — reaches only 0.663, well short of something you'd trust to ship.

• What stalls terminal-agent training is the data, not the algorithm. Tmax hits 27% on Terminal-Bench 2.0 with a 9B model and a plain outcome-only recipe, matching far larger models. CLI-Universe builds verifiable tasks in parallel. Both point at the training data.

• One model speaks molecules and proteins natively — sequence, structure, language. BioMatrix hits SOTA or near-SOTA on 77 of 80 tasks, so drug and protein work no longer needs a stack of specialist models. The clean-sweep claim needs a check on per-task baseline strength.

• Move passage compute offline and reranking becomes deployable. KaLM-Reranker-V1 decouples query and passage with "fast but not late interaction." The 0.27B Nano version trades blows with 7-12B embedding models and matches industrial rerankers on BEIR.

Also Notable

• Train "Data Processing" Itself as a Skill — DataClaw0 has an agent actively trim and structure raw multimodal streams to serve post-training, rather than passively label.
• Real Phones Force Two Training Environments — real devices are slow, stateful, side-effect-prone, and hard to reset, so the authors combine real and mock apps into a recipe for training open phone agents.
• Computer-Use Skill Learning Usually Assumes a Static, Safe World — SkillHarness handles safe skill learning under adversarial, dynamic risks like prompt injection and pop-ups.
• Bring the MoE Idea to GQA Self-Attention — activate attention heads by token difficulty to cut the quadratic compute of long context.
• Long-Horizon Agents Lock Onto One Reading of the Evidence Too Early — answer-only scoring misses this "process collapse," and this paper diagnoses it directly.
• Use a VLM as a Driving Brain — Does It Generalize Across Geographies? — Robusto-2 uses OOD corner cases in new cities like Lima and New York as the test.
• Entity-Level Membership Inference: Ask an LLM Whether an Entity Was in Training — directly relevant to privacy-leak and copyright-compliance risk assessment.

Read the full edition →