AI Agents

SPLASH switches the attention parallelism layout — tensor, data, or context — while requests are still running, instead of fixing it at launch. The insight is that modern attention with few KV heads decouples where the KV cache lives from how weights are sharded, so most state already sits where the next layout needs it; the handoff happens at a batch boundary with under 0.51% median overhead. It also introduces Decoupled Ownership Parallelism, claiming 27-60% more KV capacity and 1.3-1.73x throughput.

A representation-level study finds that routing a harmful request through a tool call doesn't make models perceive it as less harmful — the harm signal stays encoded and transfers across modes — but it raises the effective threshold at which that perception converts into a refusal. Tool-mediated refusal is also more brittle: weakening the refusal computation breaks it at lower intervention strength than conversational refusal, with benign capability intact. Across several open-weight models, the mechanism analysis is the contribution, not just the behavioral gap.

CRR extracts step-level process rewards from outcome-only RL by exploiting forkable executable environments: at selected decision points it restores state, samples an alternative action, rolls the branch forward under the policy, and replaces the step's advantage with the terminal-return difference. No human process labels, no learned reward model. With a 14B policy it lifts pass@1 on SWE-bench Verified, Live, and rebench, and at equal wall-clock beats extended outcome-only GRPO 41.7% versus 36.7% — fork overhead included. Honest about the limit: it needs cheap, reliable state restoration.

ATTUNER isolates why position-independent caching degrades quality: not the positional mismatch, and not the KV states themselves, which stay faithful — the failure is in attention when the model must select among multiple cached artifacts. Swapping in full-prefill attention scores recovers performance with the cache unchanged. The fix is low-rank adapters on the query projections, trained by distilling full-prefill distributions, touching under 0.05% of parameters and needing no recomputation at inference. On Qwen3-4B/8B across seven skill, document, and memory benchmarks, 3.73x speedup.

ROFT asks whether an agent can improve by training only on explanations of its own experience. The loop: attempt a task, observe feedback, write a retrospective, fine-tune with next-token loss on the explanation tokens alone — no teacher, no reward update, no verifier. With Qwen3.5-4B it reaches 49.2% and 26.8% on held-out SWE-bench Verified and Pro after 20 updates, versus GRPO's 48.0% and 25.3% after 40. Strikingly, it solves tasks where all 64 base attempts failed, and behavioral analysis shows it indirectly assigns credit to actions.

The claim: most of what gets built as agent "memory" is better served by documentation. Instead of a vector store accumulating opaque state, write durable, retrievable docs the agent reads like any other reader — versioned, inspectable, correctable. The pitch resonates with anyone who has debugged memory-induced drift and found no way to audit what the agent thinks it knows. It's an opinion piece, not an eval, and the line between "documentation" and "a well-designed memory store" blurs under pressure.

Vendor guidance accompanying the Opus 5.5 release, aimed at Claude Code and Claude users: how to prompt, when to lean on extended reasoning, and workflow patterns for the new model. Useful if you operate Claude Code in anger, but it's first-party advice with the usual incentives — treat the usage recommendations as a starting hypothesis, not measured results. No independent benchmarks here, and the HN thread will likely surface more honest tradeoffs than the post itself.

Willison argues that pay-per-use APIs and hosted platforms need hard budget caps as the default — spend $X, return errors, full stop — not soft warning emails. The motivating case is coding and personal agents that cheaply spin up services calling paid APIs or billable compute, where a runaway loop overnight becomes a four-figure surprise. He grants businesses may prefer uptime, but makes opt-out dangerous mode an explicit checkbox. It's a product-policy argument, light on how providers would actually enforce it.