Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-03-22 · source-backed
Equips code agents with structured memory built from historical commits, distilling intent-to-code mappings with self-refinement via verification feedback. Using DeepSeek-V3.2 as backbone, boosts SWE-bench Verified from 68.4% to 77.8% — new SOTA. Co-evolution with project history beats stateless generation. Source
Each link below shares sources, entities, or timing with this story.
The standard multi-model coding pipeline uses a reasoning model to plan, then a code specialist to generate. A new paper flips the pattern — let the specialist generate freely, then have the reasoning model review — and hits 90.2% pass@1, outperforming GPT-4o at 87.2% and O1 P...
Agent Lightning v1.0 (arXiv 2608.17528) inverts the standard agentic RL architecture, and the inversion is the whole point. Normally the training engine owns the environment loop. It drives the agent, collects trajectories, computes rewards. Which means your training setup and...
The Agent Lightning result (arXiv 2608.17528) is a 9B model gaining 14.6 points from 3,500 lines of training code and modest compute. The point isn't that a 9B beats anything, it's the cost curve: 6K examples is a dataset a small team can build. Task-specific agentic RL on an...
While the capability stories pile up, here's the counterweight. As SWE-bench Verified scores cluster near saturation on July leaderboards, an enhanced analysis (SWE-Bench+, on the AIware 2026 benchmark track) found 60.83% of commonly resolved issues contain solution leakage ri...
It segments trajectories at perplexity spikes to keep code and log syntax intact, uses proxy attention weights to score how relevant each historical block is to current reasoning, and runs a dynamic rolling window that can recall context it previously dropped. On SWE-Bench-Ver...
Same weights. Same tasks. Same time budget. One change to the harness, and fail-to-pass on SWE-bench Verified went from 28% to 49%. The setup in arXiv 2608.26218 is clean enough to be worth describing precisely. 169 SWE-bench Verified tasks, a 20,480-token context window, a fi...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.