PaperlayerView exact revision on arXiv

Source-linked AI summary

Small Foundation Models of Human Cognition and Behaviour

Nick Oh, Fernand Gobet

arXiv:2608.05224v3cs.AIcs.CY

TL;DR

The paper asks whether cognitively fine-tuned language models require very large scale and whether they use task structure rather than statistical shortcuts. It trains and diagnoses fourteen models on Psych-101, finding that scale matters little in-distribution but improves out-of-distribution transfer, while stimulus and feedback content accounts for most retained information. These models can therefore estimate noise ceilings within covered paradigms, but their scope remains bounded by training coverage and methodological limits.

  • Problem

    It remains unclear whether general-purpose cognitive proxies require very large models and whether their predictions reflect task structure or statistical shortcuts.

  • Method

    The study trains fourteen models from 135M to 14B parameters across four families, varies capacity and data, and applies prompt-channel ablations plus trial-order permutations.

  • Results

    Scale barely matters in-distribution but improves out-of-distribution transfer, while masking stimulus and feedback content removes 75.7% of learned information and pushes models below chance.

  • Takeaways & Limitations

    Small cognitively fine-tuned models can estimate noise ceilings for in-distribution psychological paradigms when predictions use the information participants were shown and told.

  • Takeaways & Limitations

    The conclusions are bounded by fixed paradigm coverage, low-rank adaptation, and dense decoder-only transformer architectures.

Abstract

from arXiv · show

Large language models fine-tuned on human behavioural data have emerged as general-purpose cognitive proxies, but the scale this requires, and whether these models process task structure or exploit statistical shortcuts, remain open questions. We train fourteen models from 135M to 14B parameters across four architecture families on Psych-101, a dataset of 10.7 million trial-level choices from 160 experiments. For in-distribution simulations, scale barely matters. The models fall within a narrow band, as though against a ceiling, and 0.6B to 1B parameters suffice to match a 70B baseline on held-out participants. Out-of-distribution, that band opens into a markedly steeper scaling gradient, with larger models clearly advantaged in generalisation to novel task structure. To determine what information these models use, we run two diagnostics. We progressively strip four prompt channels -- task instructions, experimental stimuli, outcome feedback, and choice history -- across 27 experiments, and permute trial order. Masking the content of stimuli and feedback destroys 75.7% of learned information and pushes models below chance, demonstrating that choice history alone does not account for performance. Permutation reveals invariance on tasks with independent trials but sensitivity where trial order is determined by prior responses. Small cognitively fine-tuned models therefore show promise as noise ceiling estimators for psychological experiments, though their scope remains bounded by the paradigms seen in training.

1 Introduction

This paper tests how model scale and prompt information shape cognitively fine-tuned language models. It finds that scale matters little within trained paradigms but improves transfer beyond them, while predictions depend chiefly on task-relevant content rather than shortcuts.

  • Scale: Fourteen models spanning 135M–14B parameters vary scale, adapter rank, and training data to separate sources of predictive capacity.The models cover four architecture families and use Psych-101 behavioral data.
  • Scale: In-distribution, increasing adapter rank reduces the size needed to match reproduced Centaur-70B from 8B to 0.6B parameters.The result indicates that model scale is not the main constraint within trained paradigms.
  • Scale: Out-of-distribution, larger models generalize better, indicating that additional parameters primarily improve transfer beyond Psych-101.The scale advantage reappears when models encounter novel task structure.
  • Information use: Structural ablations remove instructions, stimuli, feedback, or choice history, while order permutations test whether trial sequence carries task-relevant information.The diagnostics target whether models process task structure or exploit statistical shortcuts.
  • Scope: The study concludes that these models can estimate behavioral noise ceilings, but only across paradigms represented in their training data.They are not presented as mechanistic theories of cognition.

2 Related work

The paper places cognitively fine-tuned language models within a broader shift from handcrafted cognitive models toward data-driven behavioral benchmarks. Prior work established noise ceilings and domain-specific neural proxies, while newer foundation models broaden the target domains and data sources.

  • From handcrafted models to behavioral benchmarks: Traditional cognitive models are handcrafted, whereas neural networks trained on human behavior provide data-driven reference points for interpretable theories.This reframes neural models as benchmarks rather than replacements for process-level theories.
  • Domain-specific precedents: A 4-in-a-row model trained on 82.8 million moves from 1.2 million players established a noise ceiling that exposed failures in heuristic-search theory.The failures were not detectable from raw behavioral data alone.
  • Domain-specific precedents: A noise ceiling estimates the best achievable prediction performance after accounting for irreducible stochasticity in human responses.This provides a benchmark for judging how much behavior is predictably structured.
  • Domain-specific precedents: Other studies used neural networks to model planning and risky choice, recovering behavioral regularities and features associated with established theories.Examples include deliberation patterns resembling hippocampal replays and Prospect Theory’s S-shaped utility function.
  • Foundation models: Recent foundation-model work fine-tunes large language models directly on human behavioral data to support general-purpose cognitive proxies.These systems differ in training data, adaptation strategy, and prediction granularity.
  • Foundation models: Newer systems extend the paradigm from controlled experiments to behavioral literature, social surveys, and social-media corpora.This broadens both the data foundation and the behavioral domains targeted.

3 Methods

The study trains fourteen base language models on Psych-101 while independently varying model size, adapter capacity, and training-set size. Its comparisons account for model families, fine-tuning choices, context limits, and confounds between generation and parameter count.

  • Data: Psych-101 contains 10.7 million trial-level choices from over 60,000 participants across 160 experiments, represented as natural-language participant sessions.The experiments span decision-making, bandit, MDP, memory, and supervised-learning tasks.
  • Models: Fourteen non-instruct base models from Qwen3, Llama-3, SmolLM, and OLMo are fine-tuned into the Qwentaur, Llama-Centaur, Smoltaur, and Olmotaur families.Model sizes range from 135M to 14B parameters.
  • Training: Supervised fine-tuning uses response-token cross-entropy with rank-stabilised LoRA applied across linear layers.The loss is masked to human response tokens rather than task instructions or context.
  • Training: The study uses base rather than post-trained models because prior work reported reduced behavioral alignment for post-trained variants.The cited comparison also found no same-size advantage for instruction-tuned models over base models in brain-response alignment.
  • Capacity and data: Model scale, adapter rank, and training-set size are varied independently, with ranks 4–64 and experiment-stratified data fractions from 1/16 to full data.Stratification preserves representation of all 160 experiments across training-set sizes.
  • Comparability: Context windows cause unequal trial truncation, and within SmolLM and OLMo generations, larger models also have newer generations and larger windows.The analysis therefore focuses mainly on Qwentaur and Llama-Centaur comparisons.

4 Experiments

Fourteen cognitively fine-tuned models show little scale dependence for held-out participants but stronger size advantages when transferring to unseen experiments. Prompt diagnostics indicate that models rely mainly on stimulus and feedback content while adapting to task-specific sequential structure and exchangeability.

  • 4.1 Experiments: 0.6B to 1.7B models can match larger models when adapter rank increases, with 0.6B matching the 70B baseline at rank 32 and above.At rank 8, models from 3B to 8B match or beat the 70B reproduction; adapter capacity substitutes for parameters.
  • 4.1 Experiments: Psych-101 performance varies little with scale: eight matched models span 0.509–0.537 mean NLL, a 0.028-nat band containing the Centaur-70B reproduction.The benchmark evaluates held-out participants in-distribution, with lower negative log-likelihood indicating better performance.
  • 4.1 Experiments: Out-of-distribution performance separates models more strongly: Psych-201-RT scores span 0.788–1.033, a 0.244 spread, and larger models retain a substantial advantage after fine-tuning.After fine-tuning, the tenfold-size advantage is 0.104 nats for Qwen and 0.172 for Llama on the out-of-distribution benchmark, versus 0.014 and 0.026 in-distribution.
  • 4 Experiments: The models are best interpreted as empirical noise-ceiling estimators for paradigms represented in Psych-101, not as mechanistic theories of cognition.Their predictions provide a bound on how much behavior is predictable, but their reliability beyond the training paradigms is unvalidated.
  • 4.2 Experiments: 75.7% of learned information is lost when stimulus and feedback content is masked, whereas instructions account for 12.5% and the template for 11.7%.Information retention falls from 1.00 in the original condition to 0.84 after instruction ablation, −0.12 after content masking, and −0.26 for history-only prompts.
  • 4.2 Experiments: On 18 of 27 experiments (67%), learned information degrades monotonically across ablations, while order permutation reveals exchangeability on independent trials and order sensitivity when prior responses generate later offers.The ablations identify which prompt channels models use, while permutation tests whether trial order follows the task’s experimental structure.

5 Discussion

The models are best understood as predictive instruments rather than mechanistic theories: they estimate in-distribution noise ceilings and indicate where interpretable theories may fall short.

  • These models do not constitute cognitive theories because they posit no architecture, mechanism, or account of how choices are generated.Their generality instead comes from learned statistics and scale rather than designed-in assumptions.
  • Noise-ceiling estimation maps where behaviour contains systematic structure and where current interpretable theories leave predictive gaps.The model–theory gap identifies paradigms where theory-building has more room to improve.
  • The paper frames this role as scientific regret minimisation rather than a claim that the models recover cognitive mechanisms.Used only for prediction, a model returns a number; used as a ceiling, it locates theory shortfalls.
  • Fine-tuning may move models toward internal cognitive structure, given prior reports of recovered memory, prefrontal-like specialisation, and planning mechanisms.Whether that occurs here remains a question for mechanistic and representational analysis.

6 Limitations and future work

The study’s conclusions are constrained by how behavioural data are sampled and adapted: paradigm breadth is fixed in the main ablation, while low-rank adaptation and dense decoder-only architectures limit what is tested.

  • Data structure: Participant-level subsetting isolates data quantity while holding coverage of all 160 experiments fixed, so the study measures depth rather than breadth.The authors argue that experiment-level subsetting would better estimate how many distinct paradigms a foundation model needs.
  • Data structure: Psych-201 is 3.5 times the size of Psych-101 but still nests participants within experiments, preventing cross-experiment person–task decomposition.This structure limits conclusions about individual and task contributions.
  • Training and architecture: Low-rank adaptation can substitute for parameters over the tested range, but full fine-tuning is needed to distinguish adaptation limits from a ceiling on learnable signal.The data ablation also shows diminishing returns, so these limits may occur at different points.
  • Training and architecture: All tested families are dense decoder-only transformers, leaving open whether mixture-of-experts, state-space, or hybrid designs shift the observed plateau.SFT also optimises imitation of the training distribution at a measurable cost to distant capabilities.

7 Conclusion

Fine-tuning matters more than scale for prediction within trained paradigms, whereas larger models improve transfer beyond them; small models can estimate in-distribution noise ceilings, but coverage limits their reach.

  • Across fourteen models from 135M to 14B, fine-tuning matters far more than model size for Psych-101 accuracy.
  • Beyond the training paradigms, larger models transfer better without saturation, making scale relevant to out-of-distribution generalisation.
  • Masking stimulus and feedback content collapses prediction, while trial-order sensitivity appears only when the design makes order informative.
  • Small cognitively fine-tuned models are adequate for estimating noise ceilings within seen paradigms, with paradigm coverage rather than capacity as the binding constraint.

Ethics Statement

The models predict laboratory task responses using existing behavioural data and are intended as scientific instruments, but their utility remains bounded by training coverage and untested extrapolation.

  • The models are trained on anonymised trial-level responses from published psychological experiments, with no new human data collected or participants re-identified.
  • Their intended use is as scientific instruments for predicting laboratory task responses, although models capable of predicting choices could in principle be misused.
  • Utility is bounded by training-data paradigm coverage, and reliable extrapolation to conditions producing more extreme outcomes than Psych-101 remains untested.

Reproducibility Statement

The paper documents its models, training configuration, data pipeline, and computational infrastructure to support direct methodological comparison and reproducibility.

  • Reproducibility Statement: Tables 1–4 compare model components, hyperparameters, data and loss configuration, and computational infrastructure.The comparison covers base LLMs, post-training methods, training data, target domains, optimisation settings, numerical precision, loss masking, input format, hardware, training time, and inference settings.
  • Reproducibility Statement: All model variants share identical hyperparameters except for per-device batch size, gradient accumulation steps, and adapter ranks.Entries marked “Not specified” indicate details not reported by the original paper.
  • Reproducibility Statement: The data pipeline records numerical precision, loss masking, input representation, and whether computer-assisted tools were used to construct or augment training data.Quantised formats apply to base weights, while LoRA adapters and forward/backward computation use bf16.
  • Reproducibility Statement: The models were each trained for 1 epoch on Psych-101 using a single NVIDIA A100 80GB GPU.Infrastructure, packages, and computational cost are reported alongside the training setup.

B Adapter rank and training-set size

The study separates adapter capacity from model size and training-data quantity, using rank sweeps and experiment-stratified subsets to compare their effects on Psych-101 performance.

  • B Adapter rank and training-set size: Adapter rank is swept over r ∈ {4, 8, 16, 32, 64} at full data, while training-set size is varied from 1/16 to the full dataset at fixed r=16.The training subsets are nested and experiment-stratified, preserving coverage of all 160 experiments at every fraction.
  • B Adapter rank and training-set size: A small model at high adapter rank can overlap a much larger model at low rank, making the substitution between adapter capacity and parameter count visible.Figure 7 separates rank and training-data effects by model family.
  • B Adapter rank and training-set size: Figure 6 plots mean negative log-likelihood against parameter count, with colour intensity encoding LoRA rank.Trend lines are fitted only within groups matched on model generation and context window.
  • B Adapter rank and training-set size: Figure 7 compares mean NLL against LoRA rank at full data and against training-data fraction at r=16 across the four model families.The dotted line denotes the cognitive-model baseline, while diamonds denote reproduced and reported Centaur-70B values.

C Full per-experiment results

The paper reports per-experiment and task-type performance using raw and normalised NLL, while documenting exclusions caused by non-discrete, mixed, degenerate, or state-dependent response spaces and evaluating unseen experiments separately.

  • C Full per-experiment results: Continuous responses, mixed response types, degenerate action spaces, and state-dependent action spaces prevent a single meaningful ln(k) normalisation in excluded experiments.Examples include free-form probability judgements, sessions combining binary and numeric responses, go/no-go responses with ln(1)=0, and navigation tasks whose valid actions vary by state.
  • C Full per-experiment results: Table 5 reports fraction of available information captured above chance as (ln k − NLL) / ln k for 34 of 46 Psych-101 experiments.A value of 0 indicates chance-level performance and 1 indicates perfect prediction; 12 experiments are excluded because no well-defined discrete response space or baseline is available.
  • C Full per-experiment results: Tables 6–8 provide base-versus-finetuned NLL, per-experiment NLL across four fine-tuned families, and comparisons with behavioural and non-cognitive controls.The tables use bf16 inference where specified, and Table 8 groups comparisons by base-model family.
  • C Full per-experiment results: Psych-201-RT tests generalisation to entirely unseen experiments, with normalised analysis restricted to 15 of 18 experiments having discrete response options and a chance-level baseline.The dataset spans decision-making, Markov decision processes, multi-armed bandits, and miscellaneous tasks.

D Specificity of cognitive fine-tuning

The study tests whether Psych-101 gains are specific to cognitive fine-tuning rather than generic or unrelated behavioural post-training. Cognitively fine-tuned models outperform matched controls, while prediction targets distinguish their behavioural value.

  • D Specificity of cognitive fine-tuning: Non-cognitive controls do not improve over their base models, extending the 70B finding downward across the tested scales.This indicates that generic supervised fine-tuning does not incidentally produce models that predict human behaviour in these evaluations.
  • D Specificity of cognitive fine-tuning: Be.FM gains modestly on overlapping task domains, whereas Socrates performs worse than its base model on several tasks despite broader behavioural training.Socrates targets population-level distributional alignment, while Psych-101 requires within-participant sequential prediction conditioned on stimuli, feedback, and prior responses.
  • D Specificity of cognitive fine-tuning: The results partition behavioural foundation models by prediction target rather than placing them on a single continuum from worse to better.The relevant distinction is between population-level response distributions and trial-by-trial prediction of an individual participant’s choices.
  • D Specificity of cognitive fine-tuning: Cognitively fine-tuned models consistently outperform non-cognitive controls at every tested parameter scale.The comparison uses Llama-Centaur and Qwentaur against size-matched Hermes, Nemotron, and Be.FM controls.

E Retention of general capabilities

Cognitive fine-tuning selectively reorganises general capabilities rather than uniformly degrading them. Mathematical reasoning and formal planning decline, while cognitive and language benchmarks remain comparatively stable, and ethical effects are inconsistent.

  • Cognitive fine-tuning produces selective reorganisation rather than uniform degradation.
  • −0.25 to −0.12: GSM8K degrades significantly in six of eight models, with smaller losses at larger Qwentaur scales.
  • ACPBench group means also degrade significantly in six of eight models, indicating a similar pattern for formal planning.
  • ±0.02: cognitive and language group means remain within this range at every scale.
  • Ethics effects are inconsistent: four models improve significantly, one declines significantly by −0.02, and three show null effects.

F.2.1 Non-monotonic experiments

Most structurally ablated experiments degrade monotonically, while the non-monotonic cases largely reflect floor effects, normalisation artefacts, or partial information being misleading. These reversals therefore do not overturn the aggregate ablation hierarchy.

  • 67%: 18 of 27 experiments with defined chance baselines degrade strictly monotonically across the four ablation conditions.
  • Five apparent inversions are non-substantive because both compared conditions are already at or below chance.
  • <0.03 retention units: four of those inversions fall within the spread across models.
  • Four genuine reversals show that partial information can perform worse than either less or no task content, consistent with misleading placeholders or mismatched task interpretation.
Loading 2608.05224v3…