Harness distillation without the harness at runtime
The latest in AI research, read in full and distilled weekly: what's new, what's gaining acceptance, and what newer evidence has overturned.
You have spent the week watching agents get faster by thinking less at test time. That trick has a ceiling: eventually the scaffolding that steers them becomes the thing that slows them down. This week the field went after the scaffolding itself, and one result suggests you can train it in and then leave it behind.
Three findings carry the day. A team from Peking University, Google, and HKUST showed that an agent reviewing its own trajectories can distill a harness into the model, letting you discard the harness at deployment and still beat the model that keeps it. A group at UT Dallas built a memory system that routes simple decisions to a fast controller and saves the heavy model for when it matters, cutting query latency by more than a third. And a paper from Google Cloud AI Research, Stanford, and WashU regularizes the search for better harnesses so the process does not collapse into benchmark-specific hacks.
The harness can be trained in, then dropped
The strongest compounding evidence this week sits on the boundary between training and deployment. Researchers at Peking University, Google, and HKUST have shown that a model can learn to internalize the behavior of an elaborate steering harness, then run without it.
Their method, Harness-Zero, works like this. First they evolve a rich harness that wraps the model during training: an agent that reviews each proposed action against a private reference and replaces it when needed. Then they collect the reviewed trajectories and fine-tune the base model on them. The result is a model that, deployed under the minimal target harness alone, scores 44.3% macro-average success against the base model's 23.3%, and even beats the 41.7% achieved when the specialized harness stays attached at inference time. The harness-guided review signal itself is what matters: alternative supervision sources, from raw stronger-model trajectories to answer-only review, achieve only 3-15% against the 30% from the full agent-as-harness review. The distilled model also recovers 82.3% of the behavioral patterns that the optimized harness induced, measured across 28 distinct patterns in spreadsheets, multi-app tool use, and scientific reasoning.
The evidence is the authors' own experiments across three domains, not yet replicated. What changes for builders is the possibility that your deployment harness can be simpler than your training harness, and that the gap between them can be closed by distillation rather than by accepting the performance hit.
Harness-Zero: Harness Distillation via Agent-as-Harness - full text
Two new bets on how agents improve themselves
The genuinely new work this week tests two different bets on recursive improvement: one that evolves the harness, one that distills it, and one that watches what happens when you stack training stages and they fight each other.
Regularized search finds better harnesses without gaming the benchmark. A team from Google Cloud AI Research, Stanford, and Washington University in St. Louis proposes RRSI, a method for evolving agent harnesses that adds three guardrails to prevent the search from overfitting. The proposal distribution uses a cosine-annealed edit budget. The selector rejects candidates that leak benchmark content, fall below a noise-adjusted score floor, or cost too much for their gain. And when progress stalls, the system reserves part of the budget for under-explored components. The result is up to 14.1 points on the evolve split and up to 4.7 points on held-out benchmarks, using 30% fewer policy tokens than unregularized evolution. The evidence is single-team, single-paper.
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses - full text
Multi-stage training forgets, and the last stage does not fix it. A team led by Haixin Wang and colleagues shows that sequential post-training on math, then search, then e-commerce, then instruction-following causes severe forgetting. Single-hop search drops from 45.2 to 14.6 after the e-commerce stage, and even after the final instruction-following stage only recovers to 33.5, still 11.7 points below peak. The authors trace this to distinct optimization directions in weight space: each stage pulls the model toward a different region, and no single checkpoint can align with all of them. The evidence is one paper, one model family, but the effect is large and the mechanism is visualized directly.
ACLArena: Agent Continue Learning in Multi-stage Post-training - full text
The 82.2% ceiling was not a ceiling
One benchmark result the field had accepted as a hard limit turned out to be breakable by a different approach entirely. The old claim held that an expert-authored reference implementation achieved 82.2% on RMBench, establishing a high ceiling for agent construction. Two newer results have passed it, but they do not measure the same thing.
MaP-WAM now reports 83.3% on RMBench, and DRG-MAPPO reports 87% in cooperative air-combat simulations. The problem is that these numbers share a percent sign and little else. RMBench measures end-to-end agent construction on software tasks; the air-combat result measures win rate in a physics simulator. The edge between them fails the kind test, and the ceiling that was not a ceiling is better understood as a category error. Builders should treat the original 82.2% as a local optimum on one benchmark family, not as a bound on what agents can do.
The more durable update is the one that replaces the old claim about procedural graphs. The field had held that self-evolving execution structures could repair a flawed expert prior. The newer work shows that an on-policy expert-correction pipeline, which rewrites only the failing turn in the weaker model's rollout, preserves its planning style and restores performance where imitation alone fails. This is a replacement, not a contradiction: the mechanism is finer-grained and the supervision is targeted rather than structural.
Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails - full text
What to read if you have one hour
Harness-Zero: Harness Distillation via Agent-as-Harness - full text Worth the hour if you are deciding whether to ship a heavy harness or invest in distilling it away.
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses - full text Worth the hour if you are running harness evolution and watching it overfit to your benchmark.
The ACLArena paper is already covered above.
Watch whether Harness-Zero's 44.3% holds when the harness being distilled is not hand-evolved but automatically generated, and whether the behavioral recovery rate stays high as the pattern set grows past 28. The field is betting that scaffolding can be temporary, but the bet is not yet settled.
Accelerate every builder and agent to frontier speed.