Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-10 · high
The gain comes from verifying bug fixes by reproduction, not by trusting the agent's own claim, and it holds across three different model backends.
Why now: The paper posted to arXiv in August 2026, and PMCoder's numbers haven't been independently checked yet.
PMCoder resolved 25 more SWE-bench Verified cases by refusing to trust its own "finished" claim, per a study posted to arXiv. That's a 5.0-percentage-point gain over a harness-matched baseline, and it holds across three different backends: Claude Haiku 4.5, DeepSeek-V4-Flash, and an OpenHands port. For teams building coding agents, the backend matters less than whether the harness checks its own work.
The system couples two things agent harnesses usually keep separate: planning and episodic memory. The plan phase now conditions what gets pulled from memory. Memory-derived trajectory stats decide when the agent's stuck and needs a new plan, instead of letting it grind on a broken approach.
The bigger break from convention is verification. PMCoder doesn't accept an agent's self-report that a bug is fixed. It checks whether the reported issue actually reproduces, or stops reproducing, after the patch. On the smaller Verified-500 subset, the gain held at 2.8 percentage points or better across all three backends, per the paper. It doesn't say what the added memory and replanning loop costs in tokens or latency.
Each link below shares sources, entities, or timing with this story.
The leaderboard says first place. The methodology says you should check your own bill. Qwen3.8 Max now ranks first on Artificial Analysis' agentic index, scoring 86.1 on OSWorld-Verified ahead of GPT-5.6 Sol Max at 83.2 and Fable 5 at 85.0, priced at $2.00/M input and $6.00/M...
$3,054 against $38,370. Same benchmark, better score. Praxist (arXiv 2608.25955, submitted August 26) replaces per-attempt agent memory with a typed evidence graph of findings, plus lane-structured frontiers and agendas, so later attempts inherit validated mechanisms rather th...
DeepSeek dropped V4 in mid-June as an open-weight model with a 1-million-token context window, priced at $1.74 per million input tokens, posting near-parity with GPT-5.4 on math and Q&A benchmarks (MindStudio). That's the headline number. The architecture underneath is more in...
DeepSeek released V4 on April 24 and the numbers demand attention. V4-Pro is 1.6 trillion parameters total with 49 billion active, MIT-licensed, native 1M-token context. It scores 80.6% on SWE-bench Verified, putting it within 0.2 points of Claude Opus 4.6. On Terminal-Bench 2...
Someone finally measured the thing benchmarks ignore: are the patches any *better*? Four generations of Claude and DeepSeek models on SWE-bench Lite, measured via CodeQL, CodeScene, CPU execution time, and peak memory (arXiv 2607.18462). Newer models resolve more instances. Bu...
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...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.