$ cat briefs/daily/2026-10-26.md
2026-10-26
October 26, 2026 (Mon)
1 research story · 1 paper pick · 2 new pages
Top Stories
1. Heartbeat Thinking moves agent work into listening time
New speech starts fresh reasoning invocations while existing reasoning and tool calls finish. On FDBench v3, the authors report 16 percentage points more tool coverage and 36.4% lower mean final-response latency than a SHANKS-style windowed baseline. Final task-quality effects vary by benchmark. First submitted 2026-10-19.
Why it matters: preserving computation across speech updates may reduce waiting, but the abstract leaves additional compute cost and controls for irreversible tool actions unspecified. → Heartbeat Thinking (source)
Paper Picks
EvalResearchBench: agents build and freeze cheap evaluators under fixed budgets. The best agree with target rankings on about 75% of model pairs, below a 91% ceiling from disagreement among targets. Read it for testing the feedback instrument before optimizing against it; these figures measure ranking agreement, not task accuracy. First submitted 2026-10-03, revised 2026-10-16. → EvalResearchBench (source)
Watch
- EvalResearchBench's best evaluator on development targets is not the best on sealed targets. Whether ranking fidelity survives repeated optimization remains open. → Eval Harness Configuration (source)
New in Wiki
- Two research notes add methods, author lists and evidence limits: Heartbeat Thinking and EvalResearchBench. (voice) (evaluation)
Updates
- Agents (LLM Agents) connects scheduling during listening with agent-built evaluation, keeping their outcome checks separate. (voice) (evaluation)