$ cat briefs/daily/2026-10-20.md
2026-10-20
October 20, 2026 (Tue)
3 stories · 2 paper picks · 3 new pages
Top Stories
1. OpenAI measures coding beyond familiar languages
On MultiLangBench, Astra reaches 87.4% mean test pass rate across 20 languages at maximum effort. Separately, GPT-6.1 Sol reaches 86.6% on COBOL migration at low effort, versus GPT-5.6 Sol's 5.4%, at 61% of its API cost. Migration requires every hidden test to preserve observable outputs byte-for-byte.
Why it matters: building partial functionality and preserving existing behavior need different success criteria; these vendor evaluations make that distinction explicit. → Eval Harness Configuration (source)
2. GRAFT removes the parameters that learned to cheat
MATS research on Anthropic's Alignment Science blog routes detected reward-hacking updates into an adapter removed after training. In Qwen3-8B Countdown-Code, hacking falls to nearly zero while correct solutions remain nearly 20pp above the base model. RL experiments use small models, single-turn tasks and heuristic monitors.
Why it matters: routing gradients may let a weak monitor help without rewarding evasion of its blind spots; realistic multi-turn RL remains untested. → AI Alignment (source)
3. MiniMax M3.1 Flash ships with explicit token rates
The API release lists $0.15/M input · $0.60/M output · $0.015/M cached input. MiniMax reports DeepSWE v1.1 73.7%, but the announcement prose supplies no harness settings. Weights are promised later, without a release date.
Why it matters: developers can evaluate a priced agent model now; downloadable deployment remains a promise. → MiniMax M3.1 Flash (source)
Paper Picks
SciΠvot: on 248 selected research-intervention cases, the best model/effort configuration with project context reaches 63.5%, versus 73.4% for fresh researchers. Read it for judgment under context: these selected failures do not estimate error rates across all research decisions. → SciΠvot (source)
DuoDiscover: agents build cheap proxy evaluators and continually recheck their fidelity. Authors report 1.92× gains for vLLM and 1.38× for SQLite over base code. Read it for optimizing the evaluation budget; total agent cost and verification overhead are absent from the abstract. → DuoDiscover (source)
Watch
- GPT-6 Luna reaches 40.3% on MultiLangBench at $0.77 per sample, versus $3.48 for GPT-5.6 Sol; the figures are estimated API costs. → GPT-6 Luna (source)
New in Wiki
- MiniMax M3.1 Flash and both paper notes above add specifications, methods and evidence limits. (model) (judgment) (optimization)
Updates
- Agentic Reinforcement Learning links GRAFT's anchor-task requirement: learning without the removed adapter needs a separate structurally similar environment. (source)