Substrates Encode Experience, Not Information

Paper 4B · Pødenphant Lund, T. (2026) · Preprint · v6 live on Zenodo

An Encoding-through-Loading Framework with Cross-Substrate Tests in Language Models

The expertise reversal effect is well established in instructional psychology but lacks a process-level mechanism. The paper proposes a recursive strategy-choice race: a substrate able to hold multiple strategies opens a competition between them whenever incoming information does not match its existing commitment; the competition costs computation while unresolved, and a substrate near its competence threshold pays most. Language models are used as a measurement substrate in which the competition is readable per token as a competing-routes signature. They are positioned as a model system for a cognitive mechanism whose human-side confirmation remains future work.

Results

• Substrate-graded expertise reversal. On the composition task Y = N × X, across Qwen2-1.5B, Qwen2.5-7B and Llama-3.3-70B and six shot counts: flat at the floor at 1.5B (4–6%), a classic in-context scaffold at 7B (31% → 43%), and a U-curve at 70B (73% → 50% → 61% at 0/1/3 shots, replicated at 75% / 52% in Experiment 2). The capacity-graded direction is known — larger models are more sensitive to in-context information (Wei et al. 2023; Shi et al. 2024). The contribution is that the harm arises from a correct, non-contradictory demonstration, the defining signature of expertise reversal rather than of prior-conflict.
• Task-specific, not a law. A second, structurally different task shows correct demonstrations harming again, but on a different substrate and through a different token-level signature. The capacity-ordering of the effect is therefore task-specific.
• Per-token signature. On Llama-3.3-70B, mean CR peaks at the strategy-crossover: 1.114 at 1-shot against 1.052 at 0-shot and 1.073 at 3-shot.
• Clarity separated from length. An elaborated demonstration gives +16 pp accuracy at lower mean CR (1.077 vs 1.138) than a minimal one on the same substrate. Cognitive load theory predicts the opposite; friction tracks race-state, not token count.
• Volume does not behave like the cost-driver. On Qwen2.5-7B-Instruct, 1,600 non-signal filler items (about 15,300 tokens) cost 13.3 pp of accuracy (40.0% to 26.7%), while nine plausible task-relevant elaboration sentences (about 150 tokens) cost 20.0 pp (63.3% to 43.3%) in a separate run on the same 30 queries. The runs start from different baselines and the 20-point drop is not significant on its own (exact McNemar, p = 0.11), so this suggests that volume and race-opening come apart; it does not demonstrate it. This is not yet a test of the friction-signature against depth of analysis, which would need a manipulation holding race-opening fixed and varying depth.
• Format-violation reactance. A system message specifying one format with a demonstration in another drops accuracy to 48%, against 70% for a matching demonstration and 78% for none (paired McNemar p = 0.007 on match vs mismatch). The result is carried by accuracy; the accompanying CR difference runs in the predicted direction but is inconclusive at n = 50.

Clarity is not the same axis as assistance (§7.3b)

Paired texts specify an action either clearly or in a buried, action-ambiguous form, matched so that Flesch reading-ease and per-token surprisal are decorrelated from the manipulation. A reader-substrate’s competing-routes over three candidate actions is read at the commit point. Clarity lowers action-selection entropy sharply (is-clear coefficient t ≈ −7 to −10 across scorers from 355M to 1.5B) while readability and surprisal carry no independent weight. On a found corpus of 17 expert-authored plain-language rewrites, the after version raises action-clarity in 15, and in a subset the clarity gain runs opposite to the readability change.

Up the capability ladder the clarity benefit does not reverse on genuinely ambiguous items (a clear-minus-buried gain in P(correct) of +0.24 at 355M and +0.18 at 1.5B; the n = 8 extraction-scored frontier cell is not leaned on for magnitude). The entropy reduction persists where accuracy ceilings, consistent with a capacity cost rather than a knowledge gap. Redundant assistance and under-clarity are therefore two signs of one mechanism, set by whether the race was open or closed before the support arrived: assistance that re-opens a closed race is capability-graded harm, support that closes an open race is a benefit that does not reverse with capability. Whether the dissociation holds for human learners is a prediction routed to Paper 16, not a result of these measurements.

Limits of this measurement: the hard-item ladder is a small constructed set (n = 8) offered as an existence proof of direction; the found-corpus replication (n = 17) carries the ecological weight; the frontier rung is scored by LLM-judge extraction rather than forced-choice logprobs; and a lexical-binding confound, in which a clear text binds the action verb more tightly, is only partly controlled.

Scope

The winning-route amplification argument is a formal inference from the gradient of cross-entropy loss on a softmax distribution, and applies to any substrate trained that way; the paper claims no novelty for it and credits neural collapse (Papyan, Han & Donoho 2020) and cross-entropy overconfidence (Guo et al. 2017). The encoding-through-loading account describes a within-session mechanism on the LLM substrate. Its within-generation signature is structurally comparable across the substrates measured, but its temporal scope is substrate-specific: stateless inference lacks the persistence layer that lets a trace accumulate across calls.

Companion papers

• Paper 16 — the human-side development, where the clarity-versus-assistance prediction is routed.
• Paper 20 — the same shift from content to action, applied to rule-sets.
• Paper 0 (BFT) — the behavioural framework.

Cite