Stall detection: hardcoded heuristics vs LLM judgment #33

Closed
opened 2026-08-17 19:48:52 +02:00 by SirStone · 1 comment
Owner

Parent: #29

Question

How should "training is stuck" be detected?

Resolved by scope clarification: the training runner does not detect stalls. It outputs structured per-round stats; an external observer (human or LLM) reads them and judges. No heuristics baked in.

The real question becomes: what derived/aggregate stats should the runner include in its output to make stall detection easy for an external reader? Rolling averages? Trend indicators? Or just raw per-round data and let the reader compute?

This merges into #32 (structured output format).

Parent: #29 ## Question ~~How should "training is stuck" be detected?~~ **Resolved by scope clarification:** the training runner does not detect stalls. It outputs structured per-round stats; an external observer (human or LLM) reads them and judges. No heuristics baked in. The real question becomes: **what derived/aggregate stats should the runner include in its output** to make stall detection easy for an external reader? Rolling averages? Trend indicators? Or just raw per-round data and let the reader compute? This merges into #32 (structured output format).
SirStone added the wayfinder:grilling label 2026-08-17 19:48:52 +02:00
Author
Owner

Merged into #32 — stall detection is an external concern; the runner just needs rich enough output.

Merged into #32 — stall detection is an external concern; the runner just needs rich enough output.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: SirStone/SirRoboGarage#33