Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-09-21 · source-backed
SWE-Proof turns 500 real issues into formally verified tasks by writing a specification for the new code, axiomatizing the functions it calls, and admitting an instance only after mechanical and adversarial gates agree. Across two frontier models, a quarter to a half of patches passing held-out tests admit counterexamples, and a correct formal specification lifts resolution from 85% to 95% for Opus 4.8. The catch kills the easy application: models can't write the specifications. Those that must gain nothing over an unaided baseline, only 62% of their specs pass the audit, and specification faithfulness fails on 89% of unresolved instances against 47% of resolved ones. The bottleneck moved from patch to spec.
Each link below shares sources, entities, or timing with this story.
xAI launched Grok 4.5 and Grok Build on July 8, trained partly on Cursor developer-session data. The numbers are loud: 83.3% on Terminal-Bench 2.1, 64.7% on SWE-Bench Pro, priced at $2/$6 per million tokens. On a single coding task that works out to roughly $2.49 versus $11.80...
This is the paper of the week. arXiv 2607.28871 introduces BSG-VA, which replays every validation command an agent runs across three code states: the original buggy code (B), the candidate patch (S), and the gold developer fix (G). If a test passes in all three states, it neve...
For two years the technique was accumulation. Longer system prompts, longer CLAUDE.md, more numbered do/don't lists, more "always verify your work" imperatives. Anthropic's context-engineering guidance for Claude 5 models inverts it, with an 80% deletion figure attached. The s...
A paper from Xiao Yu, Baolin Peng, and Ruize Xu makes a claim that seems obvious once stated and is genuinely new as a training methodology: modern agents are inseparable from their inference harnesses, so training them in stripped-down RL sandboxes produces a train/serve mism...
Alibaba released Qwen3.6-27B on April 22. Dense architecture. Open weights. 77.2% on SWE-bench Verified, within 3.7 points of Claude Opus 4.6. On SkillsBench, it scores 48.2% versus its own 397B MoE predecessor's 30.0%. That's a 77% improvement with 14.8x fewer parameters. Let...
This one changed how I'm spending my week. Anthropic's July 24 context-engineering post says they removed over 80% of Claude Code's system prompt for Opus 5 and Fable 5 with no measurable loss on coding evals. They call it "unhobbling" — stripping guardrails and rules that new...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.