ENTROPY_COLLAPSE — exploration is dead
Severity: critical (auto-halt candidate) · Key metrics: entropy_mean, cross-checked against reward_mean
What is happening
Policy entropy measures how much probability mass the policy spreads over alternatives at each token — in plain terms, how much it still explores. RL learns from contrast: GRPO compares completions within a group, PPO compares against a baseline; either way, the gradient only exists where sampled outcomes differ.
A near-deterministic policy (entropy ≈ 0) generates essentially the same completion for a prompt every time. Same completions → same rewards → zero advantage → zero gradient. Learning doesn't fail loudly; it silently stops, and compute keeps burning. The characteristic pair on the dashboard is entropy sliding to zero while reward plateaus — the plateau isn't convergence, it's the policy having nothing left to compare.
Why it happens:
- No entropy bonus — nothing in the objective pays for keeping options open, and reward maximization naturally sharpens the distribution.
- Sampling temperature too low during rollouts — exploration was throttled at the source.
- LR too high early — the policy over-commits to the first behavior that scored well before it saw alternatives.
- Reward saturation — if everything already scores maximum, sharpening onto the current behavior is optimal; the collapse is a symptom (see REWARD_STD_DEATH).
What it looks like in the recording
entropy_mean decaying — often roughly exponentially — and then holding near zero,
while reward_mean flattens. For reference: a uniform random policy over a 152k-token
vocabulary has entropy ≈ ln(152k) ≈ 11.9; a healthy mid-training policy typically sits
in low single digits; below ~0.25 the policy is effectively deterministic.
How the detector decides
Fires when the late-window mean of entropy_mean is below collapse_below and
the early-window mean was above healthy_early — a policy that started sharp is a
configuration choice, not a collapse. As supporting evidence it runs a trend test on
late reward: if reward has plateaued too (|z| < 1.96, i.e. no significant trend),
confidence rises from 0.75 to 0.9 — the classic pair is stronger evidence than
entropy alone.
| knob | default | meaning |
|---|---|---|
collapse_below |
0.25 | late entropy below this = collapsed |
healthy_early |
1.0 | early entropy must have been above this |
sustain |
3 | consecutive steps below threshold for the onset estimate |
What to try
- Add or increase an entropy bonus — pay the policy directly for keeping options open.
- Raise the rollout sampling temperature — restores exploration at generation time without touching the objective.
- Check whether reward is saturated — if there is genuinely nothing left to learn on this prompt set, the fix is harder prompts, not more entropy.
- If caught early, lowering LR can slow the sharpening; a fully collapsed policy usually needs a checkpoint restore — the alternatives it would need to explore no longer get sampled at all.
Related signatures
REWARD_STD_DEATH distinguishes "converged at the ceiling" from "collapsed without mastery" — read it together with this one. REPETITION_SPIRAL is often the leading edge: looping text is what a low-entropy policy sounds like.