Answers are grounded in this paper and cite their sources.
Investigates how fine-tuning and in-context demonstration affect language models, suggesting new insights into behavioral training methods.
Practitioners reach for fine-tuning or in-context demonstration to make a language model behave a certain way — check premises, abstain on a false assumption, commit only when warranted. This paper asks a prior question: what can such an intervention actually install? Framing every intervention as sculpting the model's route-landscape (an instruction is the terrain; data is the flow), the central finding is a division of labour. Fine-tuning on instruction-style data and in-context demonstration install an instruction — a disposition that reshapes routing — and, in our tests, they do not install correctness: the behavioural-instruction route moves dispositions without moving out-of-distribution reasoning accuracy, and forcing an incompatible response template destroys the model's answer-delivery behaviour while leaving its knowledge largely in place. The boundary is stated at exactly that scope — behavioural- instruction installation, on the templates and tasks tested — not as a universal negative. Three results organise the paper. (1) At 70B scale the fine-tuning outcome is governed by how well an instruction's response template fits the model's natural output format, not by training-set size: a short premise-check template preserves general capability at the matched dose (full-benchmark MMLU 82.3% versus a vanilla 84.5%, n = 14,042) while a verbose template collapses it (MMLU 8.2%) — a cliff, not a slope. The collapse is behavioural: under format-invariant forced-choice log-probability scoring the collapsed model still scores 78.3% (versus vanilla 81.4%), so the incompatible template overwrites answer-delivery while leaving the underlying knowledge in place. (2) In-context premise-checking, measured without truncation, gives no accuracy gain on any of five substrates; what it installs is a task-dependent caution disposition (raising abstention and false-premise rejection). (3) A cautionary methodological finding: an apparent +10 percentage-point accuracy lift can be manufactured entirely by a short generation cap that truncates open reasoning before it commits — the gain is pure commit-rate, not reasoning. The caution disposition's value is made quantitative with a closed-form crossover in false-premise density, offered as a descriptive model: its benefit is bounded by saturation (modern instruct models already reject false premises at 89–100%), so its net effect is a capacity-gated cost of over-rejecting valid items — indiscriminate on small models, discriminative on strong ones. Series position. Part of the Friction Theory paper-series; the competing-routes / route-landscape substrate is developed in the series' foundational papers. Data and code. The training pools, in-context exemplar sets, the false/valid-premise evaluation set, the deterministic scorer, the LoRA adapters, and the raw generations are released with the publication.
No takes yet. Share an insight, caveat, or question.
Tomas Pødenphant Lund (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: