Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-24 · high
The system never attempts the task itself; it just judges one alternative move and passes the preference back as advice the actor can ignore.
Why now: The finding surfaces on August 24, when a cheaper, weaker check already looks sufficient for catching bad agent plans without a bigger evaluator model.
A comparator model improved three different AI agents in every one of nine tested settings, without ever solving the task itself. The system, called COTA, is weaker at the underlying task than the actors it corrects, per the COTA paper.
Agent builders who want to catch a bad plan mid-task have usually needed a second model that's better at the job. COTA didn't need that edge and still helped in every setting.
COTA doesn't attempt WebShop, ALFWorld or tau^3-Retail tasks on its own. It samples one alternative to the actor's proposed next step and judges only which one leads to a better outcome. It hands that preference back as advice, and the actor stays in charge and can ignore it.
Training used pairwise comparisons. Two rollouts branch from the same prefix, one following the actor's move and one following the alternative. A label marks which rollout works out better.
The paper doesn't say by how much COTA trails the actors it corrects, only that the gap holds across all nine actor and benchmark combinations.
For builders adding a runtime check, COTA's design says the check only needs to compare two outcomes, not equal the actor's skill at the task.
Each link below shares sources, entities, or timing with this story.
SkillPyramid extends Voyager-style libraries with a hierarchical topology plus a self-evolution loop, raising average reward 38% and cutting execution steps 27.7% across ALFWorld, WebShop, and ScienceWorld. The takeaway: don't append every successful trajectory as a new flat s...
SkillForge (arXiv 2608.24747) notes that skill-extraction approaches like SkillRL never verify whether a stored skill still works against the current environment, so the bank grows monotonically while quality rots. It makes skill usage explicit during interaction so RL optimiz...
It treats the executable runtime, context construction, tool mediation, action validation, execution recovery, as the thing to learn. A separate harness engineer converts batches of target-agent failures into validated executable patches, with same-batch reruns of the frozen t...
Policy-Aware Training Scaffolding converts rollout groups from the latest policy into evidence cards and adjusts agent context per task, removing guidance as the agent improves. The scaffold is discarded at deployment — the skills exist to shape training, not to ship. 18.6% ov...
LLM as post-hoc critic for step-level Q-values. +7.7% WebShop, +13.8% ALFWorld over GRPO. Third paper in the online RL-for-agents cluster this week. arXiv:2603.08754
Instead of cloning full teacher demonstrations that mismatch the contexts a student hits at test time, spend a fixed teacher-labeling budget on short continuation rollouts that branch from the student's own trajectories (arXiv). On HotpotQA, ALFWorld, and Terminal-Bench-Dev, b...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.