$ cat briefs/daily/2026-10-28.md
2026-10-28
October 28, 2026 (Wed)
3 stories · 2 paper picks · 5 new pages
Top Stories
1. Genie 4 puts a reasoning loop around interactive worlds
Google's preview combines live scene steering with native audio. Gemini observes frames and checks progress toward multi-step goals. Access is currently limited to trusted testers; consumer and developer access is forthcoming. Google still reports drift and physical inaccuracies.
Why it matters: simulation gains an explicit planning loop, while visual plausibility remains insufficient evidence of reliable physics. → Genie 4 (source)
2. LFM3-1.7B targets local agents with a small text model
Liquid's open-weight model offers 256K context. Its table reports 45.60 on LiveClaw, but 43.40 on SWE-bench Verified with OpenCode, below Qwen3.5-9B's 51.20 on the latter. These are vendor scores, not independent measurements.
Why it matters: local-agent options improve, but selecting a model still depends on the task and harness. → LFM3-1.7B (source)
3. Anthropic reports gains in drug-development paperwork
In its Enveda case study, Claude flagged all 11 previously identified issues plus 7 additional issues across 1,061 pages, reportedly reducing human review time 89%. Partner technical reports were not independently reviewed.
Why it matters: the reported benefit concerns a concrete review workflow; it does not establish clinical efficacy or reliability on other document sets. → Anthropic (source)
Paper Picks
DynaFlow: train the coordinator while keeping sub-agents fixed. Authors report 88.08% accuracy in Forced mode, 10.90 percentage points above the strongest evaluated baseline across six benchmarks using Qwen3.5-27B. Read it for executable collaboration; matched total inference cost remains unestablished by the abstract. → DynaFlow: learning executable collaboration (source)
iPrefBench: distinguish information inferable from history from information that requires asking. Across 7 agents, action completion reaches 91.6%, but full personalized success reaches 47.7%; these maxima need not belong to the same agent. Read it for evaluating necessary clarification. → iPrefBench: knowing when to infer and when to ask (source)