So far: models read the direction of evidence reliably and its strength badly, and an undercutter on its own raises a model’s “yes.” Belief seems to need a commitment store.
Description
Instead of asking a model what it believes, this project reads open-weight models directly: from the probabilities they give to their answers and from their hidden states. A language model conflates the linguistic, the conceptual and the epistemic, and the work is to pull out the epistemic patterns tangled up in the other two, and to learn how much structure has to be layered on top.
It sits in the middle ground between philosophical discussion of whether belief applies to models at all and technical evaluations that rest on a single measure such as next-token probability. Its findings fed the turn back to commitment in the other projects. The project began in September 2026 and has been quiet since mid-October.
Method
Three reads. The probability of a relation word placed between evidence and claim, a placebo-matched change in the model’s yes-or-no answer, and a read at the answer slot.
Ladders of models. The reads ran across Qwen, Gemma and OLMo ladders, with each phase pre-registered and frozen before any spend.
Hidden states and the literature. A probing study of models’ hidden states went alongside a reading of four recent papers on belief in language models, with an AI referee reviewing the synthesis over three rounds.
Diary
Two entries so far.
Locating belief
September – October 2026
Experiment log September 21–28
The implicit-credence program
A research detour with its own agenda. A language model conflates the linguistic, the conceptual and the epistemic, and the work is to understand the epistemic patterns tangled up in the other two, and to learn how much structure has to be layered on top. On the second question, I think we’re at a stage like NLP before transformers.
The method was to read open-weight models directly from their probabilities rather than asking them, across Qwen, Gemma and OLMo ladders, with each phase pre-registered and frozen before any spend. The headline was direction yes, numeric strength no, verbal strength coarse: models tell support from attack but don’t rank how strong it is, and an undercutter on its own raises a model’s “yes.” On September 28 I wrote it up as “the wall”: credences are relative to priors, and a model’s weights are not a representation of priors. So I turned back to the doxastic project.
Dev diary October 12–15
What belief needs
A reading of four recent papers on belief in LLMs went alongside a probing study of models’ hidden states, with an AI referee reviewing the synthesis over three rounds. Out of it came my current thesis: belief requires a commitment store, meaning something that persists, governs assertion and inference, and is revised by procedure. Within a conversation the context window is the only candidate store, and it has only a raw write path. On this view an agent with a disciplined commitment store is an ally of LLM belief, and “memory plus tools” on its own is not enough.
To test that directly, my newest work is an in-context store-maintenance battery: noticing what a new item bears on, retracting what it overturns, and carrying the consequences through. It uses hand-written village scenarios with rebutters and undercutters rewritten to Pollock’s standard, audited cold, and run through smoke tests, pilots and length controls. Its second pass is being planned now.
Findings
Direction, not strength.Direction yes, numeric strength no, verbal strength coarse. Models tell support from attack but don’t rank how strong it is. “Supports” works like Carnap’s classificatory concept: weak support is still support.
Relevance gets read as bearing. An undercutter for a claim, a reason to distrust the evidence for it, on its own raised a model’s “yes” by 0.3 to 4.1 nats in one study, and irrelevant on-topic premises earned 0.1 to 0.3 of a real reason’s credit. A model trained the way modern models are trained becomes more likely to state a claim when anything relevant to it enters its context, for or against.
The wall. Credences are relative to priors, and a model’s weights are not a representation of priors.
Belief needs a commitment store. Belief requires something that persists, governs assertion and inference, and is revised by procedure. Within a conversation the context window is the only candidate store, and it has only a raw write path. On this view an agent with a disciplined commitment store is an ally of LLM belief, and “memory plus tools” on its own is not enough. Strength tests whether a model can maintain such a store in context.