Tracking How Verified Agent Experience Becomes a Model Update

A successful agent run can become material for improving the next run. But deciding which attempts qualify for learning is different from deciding what the model should learn from them.

Imagine a coding agent that fixes a defect after several unnecessary edits. The final tests pass, so the run qualifies for learning. One teaching model might emphasize the useful diagnostic step; another might encourage repeating the whole sequence. The accepted record is the same, but the proposed lesson differs.

How ASCENT constructs the update

In ASCENT, Haodong Lu and Dong Gong study agents whose model updates persist across a stream of tasks. Each task receives one scored attempt before its own experience can affect the model.

An accepted attempt supplies context to a frozen copy of the initial model, called the teacher. That teacher predicts next-token probabilities with the benefit of the completed experience. Those probabilities train a small, persistent set of added model parameters, a LoRA adapter, used by the student on later tasks. A rejected attempt does not produce that update.

The teacher's input also excludes turns whose actions were not executed. This filter does not establish that every remaining action was useful. The student still trains at its original response positions, including positions from invalid-action turns; filtering the teacher's context is not the same as deleting all those positions from training.

That gives us three components to identify: the acceptance rule, the selected experience and the model that turns the experience into training targets.

A fixed rule can still feed a changing teacher

The paper also examines teachers refreshed from the changing student. In that arrangement, both the experience being collected and the model interpreting it can depend on earlier student updates.

Our earlier note on defining success focuses on preserving the objective and evaluation authority. ASCENT adds a narrower question: what else changes after an attempt clears that fixed rule?

A passing task score does not authorize retaining every action as future behavior. If the resulting update changes how an agent handles permissions or required checks, those effects need separate evaluation before promotion.

What WTK would need to account for

WTK describes agents and teams through catalog packages compiled for target runtimes. Model and adapter versions belong to the setup that runs the package; a weight update is not simply a new instruction file.

An improvement record could identify which attempts were selected, which teacher version and input produced the training targets, and which candidate model resulted. If the teacher changes, that change should remain visible rather than be attributed only to new experience.

This complements our note on independent promotion evidence. A teacher can generate a candidate's learning signal without its agreement becoming proof that the candidate improved. Humans could govern objectives, update boundaries and promotion conditions while automated work proceeds within that scope.

Where the evidence stops

The October 4 author preprint uses open-weight Qwen3.5 models and simulated environments, with access to full teacher probability distributions. That is different from consuming a hosted model through ordinary agent tools. Its teacher-evolution table and figure give different frozen-reference values whose relationship is unresolved, so we are not carrying forward a numerical conclusion about the best refresh schedule.

WTK has not evaluated this mechanism. The useful question under our research method is whether an update's history can explain both the experience accepted and the component that interpreted it, without confusing either with deployment approval.

Source

Haodong Lu and Dong Gong, UNSW Sydney, ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience, arXiv:2610.05303v1, October 4, 2026. Sections 3.3 and Appendix C.1 describe the teacher and update construction.