Substrates Encode Experience, Not Information: An Encoding-through-Loading Framework with Cross-Substrate Tests in Language Models

Paper 4B · Pødenphant Lund, T. (2026q) · Preprint · Live on Zenodo

Substrates encode the processing-friction generated by operating on input-friction, not the information they receive. Inputs that do not open a race in the substrate leave no trace. Inputs that open a race the substrate must resolve leave a trace structured by the resolution. Eight experiments on six LLM substrates (Qwen2.5 1.5B/7B/32B base, Qwen2.5-7B-Instruct, Llama-3.3-70B-Instruct, Qwen3-235B, DeepSeek-V3) recover a substrate-graded expertise-reversal U-curve, per-token CR peaks at 1-shot strategy-crossover, elaborated demonstrations reducing rather than raising friction, a 22-percentage-point format-violation collapse, and trigger-specific placement-patterns in the response-token sequence.

DOI (concept)	10.5281/zenodo.20059861
Status	Preprint live as v1, 2026-05-25
Target venue	Learning and Instruction / Cognitive Science / TMLR
Cite-letter	2026q
Author	Tomas Pødenphant Lund [ORCID]

TL;DR

A folk model of teaching treats information transfer as the central act: deliver the content, and the learner stores it. This paper develops the alternative the empirical record has been pointing at for decades. Substrates encode the processing-friction generated when they operate on input-friction, not the information they were given. An input that opens no race in the substrate (because it matches the existing commitment, or because the substrate lacks the capacity to entertain alternatives) leaves no trace. An input that opens a race the substrate must resolve leaves a trace structured by the resolution. Information is the occasion for encoding; what gets encoded is the friction-signature of the race.

The mechanism is operationalised through Friction Theory's competing-routes (CR) signal: the per-token count of viable continuations above a probability threshold, freely readable from any logprob-returning API. CR ≥ 4 events are routes-resolution costs, concurrent strategies the substrate is holding open at that position. The framework predicts: (i) substrate-graded patterns (capacity-limited substrates cannot show the dip), (ii) per-token signatures (race friction is observable in logprobs, not only in accuracy), (iii) counter-intuitive elaboration effects (clearer demonstrations reduce friction by closing the strategy-race), (iv) format-violation reactance distinct from strategy-ambiguity, and (v) placement-patterns specific to the trigger class.

Eight experiments on the chemistry composition task Y = N × X, simple by human standards but in the regime where capable LLMs commit and less capable ones do not:

Substrate-graded U-curve in expertise reversal across model size: flat at 1.5B (substrate-limit, no race possible), monotone gain at 7B (novice tier), classical inverted-U at 70B (75% → 52% → 61% across 0/1/3-shot ICL — expert tier shows the reversal).
Per-token CR peaks at 1-shot: 100 test queries on Llama-3.3-70B confirm mean CR = 1.114 at 1-shot vs. 1.052 at 0-shot and 1.073 at 3-shot. The friction signature is visible in the logprobs distribution, not only in the accuracy collapse.
Elaborated demonstrations reduce friction: the naive prediction was that elaborated demos add load. Falsified. Elaborated demonstrations produce lower CR (1.077 vs 1.138) and +16 percentage points accuracy. Commitment-cost reading: clearer signals enable earlier commitment and close the sustained strategy-race.
Format-violation reactance: when the system message specifies one output format (<result>) and demonstrations violate it (<answer>), accuracy collapses 70% → 48% (zero-shot control: 78%). The substrate follows the system instruction in all conditions but pays a substantial friction cost actively rejecting the demonstration format throughout the response. CR rise survives RLHF-filtering on the capable substrate.
Within-session repetition produces no encoding gain: language models without persistent-memory architectures pay friction within a generation but do not propagate it across calls. The persistence layer is the cross-substrate variable.
Random nonsense costs only 13pp across 1600× volume increase: noise that habituates to non-signal does not cost regardless of token count. Semantically-plausible elaboration costs 20pp at 60× less volume. Unreducible prediction-error, not raw token count, is the cost-driver.
Sweet-spot capacity-graded peaks collapse to overload-floor when task-pressure exceeds substrate headroom: 32B-base events_CR≥3 = 5.87 at 2 facts collapses to 6.4–7.3 across all conditions with 0% accuracy when pushed past the substrate's resolution-bandwidth.
Exploratory placement-patterns in the response-token sequence: value-commit-position spikes (pos 5, CR=2.02, 62% of responses) for reactance/strategy-ambiguity; structure-decision spikes (pos 0 + pos 9) for overexplanation/closing-uncertainty. If confirmed under pre-registered replication, this would enable trigger-classification from the first ~10 tokens of a response. Hypothesis for future test, not an established diagnostic.

Theoretical positioning. The recursive race account is positioned alongside (not against) Cognitive Load Theory (Sweller, Ayres & Kalyuga 2011): CLT specifies a phenomenological description of load whose internal structure has not been mechanistically derived; the recursive race account specifies the mechanism beneath it. The Tetzlaff, Simonsmeier & Peters (2025) meta-analysis reports asymmetric expertise-reversal magnitudes (d = +0.505 for novices, d = −0.428 for experts). The directional asymmetry is consistent with a recursive-race mechanism, not with a symmetric load curve. The framework also positions race-resolution as a substrate-level operational specification beneath four of the sixteen mediators in the Multiple Moderated Mediations (Triple-M) framework recently proposed for the prior-knowledge × learning relationship (Simonsmeier & Schneider 2025).

Substrate-universality with persistence-layer specificity. The race mechanism is substrate-universal: brains, slime moulds, transformer LLMs all instantiate it. The persistence layer differs: biological substrates accumulate cross-session encoding-traces (memory, sensitisation, burnout, traumatic spectrum); language models without persistent-memory architectures pay friction within a generation but do not propagate it across calls. Amplitude is evolutionarily tuned in biology (selection for survival-relevant memory) and RLHF-tuned in language models (selection for helpfulness/harmlessness), making cross-substrate amplitude-comparison structurally incommensurable while placement remains comparable.

The framework recovers Dewey's learning-by-doing, mirror-friction, Shackleton-Jones's Emotional Context Theory, Transfer-Appropriate Processing, the peak-end heuristic, and the Yerkes-Dodson curve as observation-windows on one substrate-level process. Direct human-side empirical confirmation of the trigger-class mapping is future work.

Key findings

Substrate-graded U-curve — on the chemistry composition task, Llama-3.3-70B shows the classical 1-shot expertise reversal dip (75% → 52% → 61% across 0/1/3-shot), Qwen2.5-7B shows monotone gain (novice tier), Qwen2-1.5B stays flat at substrate-floor (no race possible).
Per-token CR peaks at 1-shot — CR = 1.114 at 1-shot vs 1.052 at 0-shot and 1.073 at 3-shot on Llama-3.3-70B, confirming that the friction signature is visible in logprobs, not just in accuracy.
Elaborated demonstrations reduce friction — counter-intuitive direction: showing the procedure (rather than only the answer) produces lower CR (1.077 vs 1.138) and +16pp accuracy. Commitment-cost reading: clearer signals close the strategy-race earlier.
Format-violation reactance — <result> vs <answer> format mismatch produces 22-pp accuracy collapse + measurable CR rise that survives RLHF-filtering.
Within-session repetition no encoding gain — LLMs without persistent-memory architectures pay friction within a generation but do not propagate it across calls.
Volume is not the cost-driver — random nonsense costs only 13pp across 1600× volume increase; semantically-plausible elaboration costs 20pp at 60× less volume. Unreducible prediction-error is what costs, not raw token count.
Sweet-spot peaks collapse to overload-floor — substrate-headroom-exceedance pushes 32B-base from events_CR≥3 = 5.87 at sweet spot to 6.4–7.3 across all conditions with 0% accuracy.
Placement-patterns specific to trigger class — exploratory: reactance/strategy-ambiguity spikes at pos 5 (value-commit position, CR = 2.02, 62% of responses); overexplanation/closing-uncertainty spikes at pos 0 + pos 9 (structure-decision positions). Hypothesis for future test.

The encoding-through-loading framework

The folk model of learning treats information as the thing that gets encoded: present the content, the learner stores it, retrieval recovers it. The empirical record across cognitive science, instructional psychology, motor-skill psychology, and now LLM in-context learning has been pointing at a different story for decades. What gets encoded is not the input but the substrate's response to the input: specifically, the friction-signature generated when the substrate must run competing routes to resolve the input.

Three triggers open the race-load:

Reactance. The substrate is given an explicit format-contract by a system message and a demonstration that violates the contract. The substrate may continue to follow the contract while running an active rejection of the demonstration throughout the response (Brehm 1966; Wegner 1994 ironic process). Friction signature visible across the full response, not localised.
Pseudo-complexity attribution. Apparent simplicity in a stimulus is treated as a signal that something hidden must be searched for. The substrate that "knew how" is the one most disrupted by being shown how, because the demonstration is processed as if there were a more complex hidden pattern to fit it into.
Capacity-overload. Task-pressure exceeds substrate headroom; sweet-spot peaks collapse to overload-floor. The "limit" is bandwidth, not storage.

Two additional triggers known from cognitive psychology, expectation-violation and personal relevance, have human-side empirical support but await direct LLM-substrate confirmation in future experiments.

The recursive race mechanism

Strategy selection (how to approach a problem) is itself a race resolved by competition between candidate strategies, much as token selection is a race resolved by competition between candidate tokens. The same race-mechanism applies recursively: strategy-level race shapes execution-level race, and the resulting friction signature is observable per-token.

Three mechanism-level components, all tested empirically:

Recursive race — strategy selection is a race, token selection is a race, schema-identification is a race; the races run concurrently and constrain each other rather than running in strict hierarchy.
Commitment cost — substrates that hold multiple committed strategies pay a sustain-friction cost; capable substrates feel this cost more, because they have more strategies to sustain.
Reactance to format-violation — explicit contracts that the demonstration violates produce a CR signature across the response, not at a single decision point.

What looks like deliberation in capable systems facing simple inputs is the visible signature of multiple concurrent races converging on resolution at once. No additional special-purpose machinery is required to generate strategy-selection, schema-application, or error-detection-by-mismatch.

Position relative to allied frameworks

Cognitive Load Theory (Sweller, Ayres & Kalyuga 2011) describes the phenomenology of load and the major instructional effects (worked-example, redundancy, modality, expertise reversal). CLT operates at the level of aggregated load and does not generate per-token or per-trial mechanistic predictions. The recursive race account specifies the mechanism beneath the load construct: where CLT predicts continuous and roughly symmetric expertise-reversal effects, the Tetzlaff et al. (2025) meta-analysis reports asymmetric magnitudes (d = +0.505 for novices vs d = −0.428 for experts) consistent with a directional mechanism.

Triple-M framework (Simonsmeier & Schneider 2025) proposes Multiple Moderated Mediations for the prior-knowledge × learning relationship, with sixteen identified mediators. The recursive race account is positioned as a substrate-level operational specification beneath four of those mediators (interference, strategy change, skill automatization, transfer), and connects to the Prior Knowledge Paradox (Simonsmeier et al. 2021) of near-zero average correlation with knowledge gains across heterogeneous mechanisms: the average is near-zero because the mechanisms differ in direction.

Reinvestment / explicit-monitoring (Masters 1992; Beilock & Carr 2001) specifies a mechanism for pressure-induced performance failure in motor skill. The race account is broader (addressing the wider phenomenon of demonstrations harming the capable, not only pressure-induced disruption), and we treat reinvestment as a sibling account at the motor-substrate level.

Connections to other papers in the series

Paper 1 (FT) — operationalises friction as competing routes (CR); 4B uses the same operationalisation and applies it to demonstration-driven race signatures.
Paper 2B (ICL/FT memory) — tests the cross-session encoding distinction. 4B is the within-session companion: same encoding-through-loading mechanism on inference-time (4B) vs training-time (P2B) scope. Cross-citation runs both ways.
Paper 2 (Capacity scaling) — cloze-vs-application asymmetry across model sizes. 4B's substrate-graded patterns extend the same capacity-substrate story to the demonstration-induced race regime.
Paper 13 (Operational FT) — behaviour-as-manifestation thesis. 4B's response-token placement-patterns are operationally compatible with Paper 13's race-opening / recursive-resolution / manifested-behaviour framing.
Paper 4 (in preparation) — training-time analogue: SFT paraphrase-distribution-match closes the strategy-race at the weight level; 4B closes it at the context-state level.

Implications

For instructional design. The same scaffolding that helps novices harms experts because the substrate's race-load is what determines what gets encoded, not the surface information. Practical heuristic: design for the receiver's expected race-load, not for sender-side completeness. A field example from the author's own work: a curriculum on malnutrition in elderly care (originally specified as nine clinical signs plus the underlying disease catalogue, with the recipients already saturated by information from other channels) was redesigned around the single action the relevant guidelines already prescribed (monthly weighing, react on ≥1 kg loss). The detailed material remained available on the back-end. Training compressed from 15–20 minutes of information to roughly 3 minutes of action. Completeness is a sender-side virtue; learnability is a receiver-side property; the two are not the same axis.

For prompt engineering. Elaborated demonstrations close the strategy-race; minimal demonstrations open it. For capable models on simple tasks, fewer-but-clearer demos beat more-but-ambiguous ones. The 1-shot dip on 70B-class models is a structural feature, not a quirk.

For LLM evaluation. The CR signature distinguishes route-competition from reactance: format-violation produces a distributed CR rise; strategy-ambiguity localises at the strategy-crossover. Per-token CR analysis can identify which mechanism is active in a given evaluation.

For substrate theory. The race-mechanism is substrate-universal; the persistence layer is substrate-specific. Brains encode across sessions; LLMs without persistent-memory architectures do not. This is not a deficiency of LLMs as cognitive models. It is the operational dissection that lets us see what is mechanism (race-resolution) and what is implementation (cross-session persistence).

Read the paper

The full paper is on Zenodo (concept DOI 10.5281/zenodo.20059861):

Pødenphant Lund, T. (2026q). Substrates Encode Experience, Not Information: An Encoding-through-Loading Framework with Cross-Substrate Tests in Language Models. Zenodo. https://doi.org/10.5281/zenodo.20059861

Read on Zenodo → · Plain English version · Dansk version