Source
Herrmann, Csordás and Schmidhuber (2025), LIT-489 — ARXIV-2503.13431 — read as NOTE-238.
The problem with the obvious metric
Next-token loss conflates difficulty with computation, and it does so at both ends:
- Uniform random tokens give the highest possible loss and require the model to do nothing, because nothing it computes can help.
- A memorized licence gives near-zero loss and also requires nothing, the program having been written into the weights already.
So both extremes are uninteresting, the interesting cases land in the middle, and an intermediate loss is indistinguishable from a mixture of trivial and impossible tokens. Reaching for perplexity to answer “is the model doing something here” returns an answer to a different question.
What to do instead
The replacement is LIT-489’s. It names this conflation and proposes the PHi layer against it, then shows the metric separating in-context automaton learning from memorization and noise on a Transformer and an LSTM, and tracking analytically computed automaton complexity within every next-token-loss bin.
Split the model between two sequence layers and insert a PHi layer:
- A posterior encoder mapping the hidden state to a latent.
- A decoder reconstructing the hidden state from that latent.
- A learned autoregressive prior predicting the latent from its own past. This is the part that distinguishes it from an ordinary variational bottleneck, and it is what makes the quantity mean something.
The metric is KL(posterior ‖ prior) per step — the nats of hidden-state
information the past did not predict, which is the incremental description
length of the in-context program. Average it over a sequence to score the
sequence.
Train it jointly with the ordinary next-token loss so the bottleneck stays under pressure to keep what the upper layers need, or freeze a pretrained model and train only the PHi layer.
Put it in the upper layers. Inserting the bottleneck early causes posterior collapse — next-token loss degrades badly and the metric goes to nearly zero, which reads as “no computation” and means “instrument broken”. In a 3B Llama the usable band was layers 18–24. Sweep and check next-token loss is undisturbed before trusting any number.
What it buys
- The separation next-token loss cannot make. Across memorized subsequences, memorized automata, fresh in-context automaton learning and uniform noise: only the in-context task scores high, on a 12-layer Transformer and a 2-layer LSTM.
- Complexity tracking with the confound controlled. Binning tokens by next-token loss and stratifying by the analytically computed description length of the generating automaton, higher complexity gives higher PHi loss in every bin. This is the result the practice rests on, because the complexity axis is external rather than judged.
- It transplants. A single layer trained into a frozen Llama-3.2-3B reproduces the separation, while next-token loss shows no consistent pattern across the same texts.
- Maths difficulty correlates with PHi loss on step-by-step solutions, partial correlation significant at every layer tested.
Conditions
This measures, it does not select. The paper’s most quotable downstream result — higher-PHi-loss rationales are likelier to be correct — is reported without the comparison that would make it actionable. Choosing by lower next-token loss alone already scores 71% on the same GSM-8K pairs, the paper says so, and no number is given for PHi as a standalone selector. What is shown is incremental value under partial correlation, plus performance on an adversarial subset where the next-token baseline is wrong by construction. Use this to ask what a model is doing; do not reach for it as a verifier on this evidence.
You have to modify the model. A bottleneck is inserted and trained. In the joint-training setting the model is also shaped by the metric, so it is not an observational probe of a network you already have.
Half the evidence rests on the authors’ taxonomy. Which of trivial tasks, licences, shuffled tokens, unseen literature and unseen code count as “interesting” is their judgement, and the metric agreeing with it is suggestive rather than probative. The automaton experiments, where complexity is a formula, are what keep the claim from being circular — and they are the smaller-scale half.
One model family for the LLM results, and the usable layer band is reported for that model only, discovered by sweeping.
Cost is unreported. 10,000 training steps for the layer, no wall-clock and no comparison, which matters if you were considering this as standing instrumentation rather than a one-off study.