$ cat briefs/daily/2026-10-12.md
2026-10-12
October 12, 2026 (Mon)
2 stories · 2 paper picks · 3 new pages
Top Stories
1. MiMo's training report makes grader compute a scaling axis
Xiaomi describes asynchronous RL with 1,568 samples and 2.7-3.7B tokens per step, at context lengths up to 1M. Alongside rollout throughput and environment diversity, it scales groupwise agentic grading to improve rewards and encourage shorter solutions. A frozen MoE router and layered reward-hacking defenses support training stability.
Why it matters: the report exposes training-design choices behind an existing release; it does not independently validate the model's ranking. → Agentic Reinforcement Learning · MiMo-V2.6-Pro (source)
2. Qwen Code reaches logged-in browser sessions and email-triggered tasks
Its 2026-10-09 update adds background Chrome operation on Mac/Linux and a sender-allowlisted Email Channel. Windows browser integration is not supported. The release also adds voice with screen sharing and approved background batch jobs.
Why it matters: an agent can work across existing accounts and remote task entry points; access boundaries become part of its operating design. → Alibaba / Qwen AI Lab (source)
Paper Picks
AgenticBBO-Bench — more tools do not consistently help. Under a finite evaluation budget, agents outperform the best numerical optimizers in four of five domains. Task semantics help more reliably than specific priors. Read it for separating tool access from optimization quality; per-task budgets and detailed costs are absent from the abstract. → AgenticBBO-Bench — agents under a finite optimization budget (source)
OneSearch-VL — train the link from visual evidence to an answer. An evidence graph guides data construction, RL and evaluation. On their new OneSearch-MI-Bench and OneSearch-Video-Bench, the authors report gains of 20.2 percentage points and 17.6 percentage points, respectively, over tool-enabled Qwen3-VL-8B. Read it for traceability; independent replication and inference costs remain open. → OneSearch-VL — traceable evidence for visual research (source)
Watch
- Station: a simulated scientific ecosystem rediscovers 62.7% of finding criteria from three withheld-result research tasks, versus 15.4% for Codex Multiagent-v2 and 14.4-20.6% for AI Scientist-v2. This measures rediscovery under the authors' setup, not a rate of independent breakthroughs. → R&D Automation Index (source)
New in Wiki
- AgenticBBO-Bench — agents under a finite optimization budget, OneSearch-VL — traceable evidence for visual research and Mara Chain — learn from rejected agent configurations — methods, authors and evidence limits. (optimization) (visual research) (refinement)
Updates
- Mara Chain: rejected configurations remain available for refinement. On AppWorld, the authors reach a target score with 65.5% fewer rollouts than GEPA; rollout counts alone do not establish total compute savings. → Agents (LLM Agents) (source)