Small Foundation Models of Human Cognition and Behaviour

arXiv:2608.052242026-08-04

参数量只有几亿到十亿的小模型,在同类实验上就能追平700亿参数大模型对人类选择行为的预测能力

研究团队在Psych-101数据集上微调了14个语言模型,参数规模从1.35亿到140亿不等,涵盖四个模型家族,用来检验模型规模是否真的重要,以及这些模型究竟是理解了任务结构还是只在利用统计捷径。该数据集包含160个心理学实验中1070万条逐次试验的人类选择数据。结果显示,在与训练数据同类型的实验上,模型规模几乎不影响表现,6亿到10亿参数的模型就能追平复现的700亿参数基准模型,但在完全陌生的新实验上,更大的模型明显泛化得更好;通过逐步拆解提示词的不同部分,团队还证实模型确实用到了实验中实际呈现的刺激和反馈内容,而不只是死记硬背过去的选择记录。

METAL MEDIA 解读图

小型认知微调模型的三步验证流程

证据状态已报告实测结果

1. 规模扫描实验在Psych-101(160个实验、1070万条试验)上微调1.35亿至140亿参数的14个模型,比较分布内与分布外表现
2. 提示词通道拆解依次移除任务说明、刺激、反馈、选择历史,在27个实验中检验模型实际依赖哪些信息
3. 试验顺序打乱测试在独立试验任务(THINGS)和顺序依赖任务(延迟选择)中打乱试验顺序,检验模型是否respect每个任务的结构
结论:任务适应性信息使用模型依据每个实验实际呈现的刺激与反馈内容作出预测,并按任务结构调整使用方式,而非单纯依赖记忆或格式模板

这是 METAL MEDIA 制作的解读图，并非论文作者提供的原图。

他们做了什么

研究团队在Psych-101(来自160个心理学实验、共1070万条逐次试验选择数据的数据集)上微调了来自Llama、Qwen3、SmolLM、OLMo四个模型家族的14个模型,参数量从1.35亿到140亿不等。
在与训练数据同类型的实验上(分布内),模型规模几乎不重要,6亿到10亿参数的模型就能追平他们复现的700亿参数Centaur模型;但在完全陌生的新实验上(分布外),更大的模型明显泛化得更好。
为弄清模型到底依赖哪些信息,研究者把提示词拆成任务说明、实验中呈现的刺激、结果反馈、选择历史四个部分,逐一移除,在27个实验中进行测试。
只遮盖刺激和反馈的具体内容(保留格式)后,模型已学到信息的75.7%被摧毁,表现跌到低于随机猜测的水平,说明模型并非仅凭记住过去的选择模式就能做出预测。
打乱试验顺序的测试显示:在各次试验相互独立的任务上,打乱顺序几乎不影响预测;而在后一次试验内容取决于前一次回答的任务上,打乱顺序会显著改变预测,说明模型会依据每个任务的实际结构调整信息使用方式。

Figure 1: Adapter rank against model size on Psych-101. Mean negative log-likelihood over the 38 of 46 Psych-101 tasks for which 11 publish a domain-specific cognitive model, plotted against parameter count, one panel per adapter rank (r=4–64). Trend lines are fitted only within groups matched on generation and context window, so Olmotaur has none and the Smoltaur fit covers the SmolLM2 models only.

Table 1: Core comparison of LLM-based cognitive and behavioural foundation models.
Model	Base LLM	Post-training	Training data	Domain
Centaur (11)	Llama-3.1-70B	SFT (masked CE on response tokens); rank-stabilised QLoRA (r=α=8), 4-bit quantised	Psych-101: 160 expts, 60,092 partic., 10.7M choices	Decision, memory, learning, planning
1021	Qwen2.5-7B-Instruct	SFT / Centaur-style SFT / GRPO compared; LoRA (r=α=32)	choices13k: 13,102 train / 1,462 test risky choice problems	Decision
Be.FM (95)	Llama-3.1-{8,70}B-Instruct	SFT with LoRA (all layers); 8-bit quantised2	AER: 2,703 papers.; MobLab: 68,779 subj., 82,057 obs.; Big Five: 17,667 subj.	Behavioural science, economic game, personality
Socrates (47)	Llama-3-8B-Instruct Qwen2.5-14B-Instruct	SFT / SFT + oracle reasoning traces / contrastive DPO compared; full fine-tuning	SocSci210: 210 TESS3 expts, 400,491 partic., 2.9M individual responses	Social sciences (economics, psychology, political science)
HumanLLM (55)	Qwen2.5-{3,7}B-Instruct Qwen3-8B4 Llama-3.1-8B-Instruct Phi-3-mini-128k-instruct	SFT (masked non-response tokens); full fine-tuning; 1:1 weight merge with base (LM-Cocktail)	Cognitive Genome: Reddit 2.8M, Twitter 673K, Blogger 368K, Amazon 1.7M; 1.2M train samples	Social intelligence (personalised behaviour)
GeCCo (66)	Llama-3.1-70B-Instruct DeepSeek-R1-Distill-Llama-3.1-70B Qwen2.5-72B-Instruct	No fine-tuning; in-context learning with iterative BIC-based refinement (10×5 runs)	Behavioural data from 4 cognitive domains (in-context, not for training)	Decision, learning, planning, working memory
Ours Llama-Centaur Qwentaur Smoltaur Olmotaur	Llama-3.2-{1,3}B Llama-3.1-8B Qwen3-{0.6,1.7,4,8,14}B-Base SmolLM2-{135,360}M SmolLM2-1.7B SmolLM3-3B-Base OLMo-2-0425-1B OLMo-3-1025-7B	SFT (masked CE on response tokens, Centaur-style); rank-stabilised LoRA (r=α∈{4,8,16,32,64})	Psych-101: 160 expts, 60,092 partic., 10.7M choices	Decision, memory, learning, planning
1No named model; methods (SFT v. RL) comparison only. 28-bit quantisation applies to 70B variant only; 8B is unquantised. 3TESS: NSF’s Time-sharing Experiments for the Social Sciences, a repository of peer-reviewed social science experiments conducted on nationally representative samples. 455 states “Qwen3-8B” without Instruct suffix; base/instruct status unspecified.

Figure 2: In-distribution versus out-of-distribution scaling at rank 16. (a) Psych-101, mean NLL over the 38 tasks with a reported cognitive model. (b) Psych-201-RT, all 18 held-out experiments; no cognitive baseline is published for Psych-201. Trend lines are fitted only within groups matched on generation and context window.

Table 2: Training hyperparameters for LLM-based cognitive and behavioural models. Adaptation strategy refers to the parameter-efficient or full fine-tuning method applied to the base LLM. All our variants share identical hyperparameters except for per-device batch size, gradient accumulation steps, and the set of adapter ranks trained. Entries marked “Not specified” indicate that the original paper did not report the corresponding detail.
Model	Adaptation	Optimiser	Epochs	Learning rate	Eff. batch size (PD×GA×d)	Scheduler
Centaur (11)	QLoRA (r=α=8), all linear layers	8-bit AdamW	11	5×10−5	1×32×1	Cosine (WU 100 steps)
102	LoRA r=α=32, all linear layers, dropout 0.05	AdamW	SFT: 6; RL: 3	SFT: 10−5; RL: 3×10−6	SFT: PD×8×1; RL: PD×8×4	SFT: fixed; RL: cosine
Be.FM (95)	LoRA, all layers	Not specified	3	10−4	1×8×d	Cosine (WU 0.1)
Socrates (47)	Full fine-tuning (no LoRA)	Not specified	1	SFT: 10−5; DPO: 10−6	PD×GA×8=256	Cosine (WU 0.05)
HumanLLM (55)	Full fine-tuning (no LoRA)	Not specified	3	5×10−6	PD×GA×8=64	Cosine (WU 0.5)
GeCCo (66)	N/A (no training)2	N/A	N/A	N/A	N/A	N/A
Qwentaur-0.6B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	2×16×1	Linear warmup (100 steps)
Qwentaur-1.7B	LoRA (r=α∈{8,16}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Qwentaur-4B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Qwentaur-8B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Qwentaur-14B	LoRA (r=α∈{4,16,64}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Llama-Centaur-1B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	2×16×1	Linear warmup (100 steps)
Llama-Centaur-3B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Llama-Centaur-8B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Smoltaur-0.1B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	8×4×1	Linear warmup (100 steps)
Smoltaur-0.4B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	4×8×1	Linear warmup (100 steps)
Smoltaur-1.7B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	2×16×1	Linear warmup (100 steps)
Smoltaur-3B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Olmotaur-1B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	2×16×1	Linear warmup (100 steps)
Olmotaur-7B	LoRA (r=α∈{4,8,16,32,64}), all linear layers	AdamW	1	5×10−5	1×32×1	Linear warmup (100 steps)
Abbreviations: PD = per-device batch size; GA = gradient accumulation steps; d = number of GPU devices; WU = warmup ratio. 1The published training script (https://github.com/marcelbinz/Llama-3.1-Centaur-70B/blob/main/scripts/cluster_train.sh) specifies 5 epochs. 11 report training for 1 epoch, suggesting early checkpoint selection. 2GeCCo (66) uses in-context learning with iterative BIC-based feedback over 10 sampling iterations × 5 independent runs. Models are fitted to held-out data with SciPy minimize (20 random restarts) and evaluated by BIC. All post-training methods use supervised fine-tuning unless otherwise noted; 102 additionally compare RL (GRPO with 12 candidate completions per step, max 1024 tokens; reward =1−\|oB−pB\| + format bonus up to 0.5, no standard-deviation normalisation); Socrates (47) additionally compares contrastive DPO (preference pairs constructed by varying the demographic persona under the same experimental condition and outcome question).

Figure 4: Structural ablation results across eight models and 27 experiments with well-defined response options. Five experiments with mixed or continuous formats are excluded. (a) Mean information retention R=(ln⁡k−ℒc)/(ln⁡k−ℒorig) by condition, averaged over experiments; R=1: no loss, R=0: chance. Circles: Llama-Centaur; squares: Qwentaur; colour intensity scales with model size; error bars: SEM. (b) Per-experiment fraction of learned information lost, δ=1−R, averaged across models. Columns sorted by history-only δ; colour strip indicates task type. Experiments with δ>1 under the history-only condition are omitted; see Figure 18 (Appendix F.2) for the complete set.

Table 3: Data pipeline and loss configuration for LLM-based cognitive and behavioural models. Precision refers to the numerical format used during training: quantised formats (4-bit NF4, 8-bit) apply to base model weights while LoRA adapters and forward/backward computation use bf16. Loss masking describes which tokens contribute to the training loss; masked approaches restrict gradient updates to human response tokens, avoiding optimisation on task instructions and context. Data format describes the input representation seen by the model during training. Data synthesis indicates whether and how any computer-assisted tools were used to construct or augment the training corpus.
Model	Precision	Weight decay	Loss masking	Data format	Data synthesis
Centaur (11)	4-bit NF4	0.01	Human response tokens only	NL trial-by-trial prompts (∼32K tokens)	Template-based prompt construction1
102	Not specified	Not specified	SFT: standard; Centaur-style. GRPO	JSON aggregated choice proportions per problem (empirical % rounded to nearest integer, e.g. {"A": 29, "B": 71})	Reformatted from choices13k empirical choice frequencies (problem-level rather than individual-participant-level prediction)
Be.FM (95)	8-bit (70B base, bitsandbytes); bf16 (8B)	Not specified	Standard (Alpaca template)	Alpaca template {instruction, input, output}	GPT-4o for research workflow extraction
Socrates (47)	Not specified	0.1	SFT: Response token only. DPO	{persona, stimuli, outcome, response}	o4-mini-high (dataset agent2); GPT-4o-mini (reasoning traces3)
HumanLLM (55)	Not specified	Not specified	Non-response positions masked	ShareGPT format	Llama-3.3-70B (extraction); GPT-4o (quality validation)
GeCCo (66)	N/A	N/A	N/A	NL prompt + Python function template	N/A
Ours	bf16	0.01	Human response tokens only (Centaur-style)	NL trial-by-trial prompts (∼32K tokens)	Psych-101 dataset (unmodified)
1Each experiment is converted into natural-language trial-by-trial prompts via author-written scripts that map structured experimental data (participant responses, stimuli, feedback) to verbalised narratives; for an example, see https://github.com/marcelbinz/Psych-201/blob/main/binz2022heuristics/generate_prompts.py. 2An LLM-based agent (o4-mini-high) parses raw TESS datasets into structured {persona, stimuli, response} tuples; see Appendix A of 47 for details. 3Given an experimental prompt and the corresponding human response, GPT-4o-mini generates reasoning traces explaining the human decision from a social scientist’s perspective; see Appendix E of 47.

Figure 5: Per-participant order variance across models and experiments. (a) THINGS odd-one-out (exchangeable; primary test). (b) Intertemporal choice (adaptive staircase; negative control). Left: violin and box plots; fine-tuned (solid) vs. base (hatched); y-axis scales differ. Right: ECDFs; solid = fine-tuned, dashed = base; steeper curves near zero indicate greater order invariance.

Table 4: Infrastructure, packages, and computational cost for LLM-based cognitive and behavioural models. Hardware refers to GPU resources used during training (or in-context generation for GeCCo). Training time reports wall-clock duration as stated in each paper; missing entries indicate the information was not reported. Our models were each trained for 1 epoch on Psych-101 on a single A100 80GB, with rank-stabilised LoRA at r=α∈{4,8,16,32,64}.
Model	Training Framework	Key packages	Hardware	Train time	Max seq. len.	Inference
Centaur (11)	unsloth	unsloth	1× A100 80GB	∼5 days	∼32,768	Not specified
102	Not specified	vLLM (inference)	RL: 4× H100; SFT: 1× A100	RL: ∼80 h; SFT: ∼5 h	1,024 (RL); 30 (SFT inf.)	T=0.7, top-p=0.95, top-k=0.5
Be.FM (95)	LlamaFactory	LlamaFactory, bitsandbytes	Not specified	Not specified	Not specified	Not specified
Socrates (47)	LlamaFactory	LlamaFactory	8× A100 80GB	4–24 h	4,096 (inf.)	T=0.6, top-p=0.9
HumanLLM (55)	LlamaFactory	LlamaFactory, DeepSpeed Zero, vLLM	8× A100 40GB	∼120 h (8B)	8,192	T=0.7
GeCCo (66)	N/A1	SciPy (minimize); Python exec()	4× A100 40GB	≤8 h per domain	N/A	T: 0.1–0.22
Qwentaur-0.6B	unsloth	unsloth	1× A100 80GB	4 h	∼32,768	—
Qwentaur-1.7B	unsloth	unsloth	1× A100 80GB	7 h	∼32,768	—
Qwentaur-4B	unsloth	unsloth	1× A100 80GB	18 h	∼32,768	—
Qwentaur-8B	unsloth	unsloth	1× A100 80GB	1 d	∼32,768	—
Qwentaur-14B	unsloth	unsloth	1× A100 80GB	1 d 15 h	∼32,768	—
Llama-Centaur-1B	unsloth	unsloth	1× A100 80GB	7 h	∼32,768	—
Llama-Centaur-3B	unsloth	unsloth	1× A100 80GB	12 h	∼32,768	—
Llama-Centaur-8B	unsloth	unsloth	1× A100 80GB	21 h	∼32,768	—
Smoltaur-0.1B	unsloth	unsloth	1× A100 80GB	3 h	∼8,192	—
Smoltaur-0.4B	unsloth	unsloth	1× A100 80GB	4 h	∼8,192	—
Smoltaur-1.7B	unsloth	unsloth	1× A100 80GB	7 h	∼8,192	—
Smoltaur-3B	unsloth	unsloth	1× A100 80GB	1 d 4 h	∼32,768	—
Olmotaur-1B	unsloth	unsloth	1× A100 80GB	5 h	∼4,096	—
Olmotaur-7B	unsloth	unsloth	1× A100 80GB	2 d 1 h	∼32,768	—
Abbreviations: T = sampling temperature; inf. = inference; d = days; h = hours. 1No training framework required; GeCCo generates cognitive models via in-context prompting. SciPy is used for parameter fitting of the generated models; Python exec() executes the LLM-generated code. 2Temperature varies by LLM: Llama 0.2, Qwen 0.15, DeepSeek-R1 0.1. All our models were trained on a single NVIDIA A100 80GB GPU. Training times scale approximately linearly with parameter count within each model family.

Figure 6: Adapter rank against model size on Psych-101. Mean negative log-likelihood over the 38 of 46 Psych-101 tasks for which 11 publish a domain-specific cognitive model baseline, against parameter count, with colour intensity encoding LoRA rank and right-hand panels showing one rank at a time. Trend lines and shaded envelopes are fitted only within groups matched on model generation and context window, so Olmotaur has none and the Smoltaur fit covers the SmolLM2 models only. Marker shape indicates context window (circle ≥ 32k, square < 32k); colour indicates family.

Table 5: Fraction of available information captured above chance, (ln⁡k−NLL)/ln⁡k, by task type for finetuned models (bf16). A value of 0 indicates chance-level performance; 1 indicates perfect prediction. Restricted to 34 experiments (of 46) with both a cognitive model baseline and a well-defined discrete response space (ln⁡k>0); 12 experiments are excluded. Finetuned families: Qwentaur, Llama-Centaur, Smoltaur, Olmotaur. Base columns: Q-8B/Q-14B (Qwen3), L-8B (Llama-3.1). Subscript r denotes our reproducing evaluation of the original Centaur model distributed by 11, evaluated under identical python library and CUDA versions as our small foundation models for fair comparison. Subscript p denotes values published by 11. 70Bp is shown for reference but excluded from best/second-best marking, as it was evaluated under different software conditions. Bold+underline marks the best model, underline the second-best.
	Qwentaur	Llama-Centaur	Smoltaur	Olmotaur	Base	11
Task type	0.6B	1.7B	4B	8B	14B	1B	3B	8B	0.1B	0.4B	1.7B	3B	1B	7B	Q-8B	Q-14B	L-8B	70Bp	70Br
Decision (8)	0.53	0.53	0.54	0.54	0.54	0.53	0.53	0.54	0.41	0.46	0.48	0.54	0.47	0.53	0.34	0.35	0.33	0.56	0.52
MDP (5)	0.28	0.30	0.31	0.31	0.32	0.28	0.30	0.31	0.18	0.20	0.26	0.30	0.20	0.30	0.15	0.17	0.19	0.32	0.31
Bandit (12)	0.48	0.48	0.49	0.49	0.50	0.46	0.48	0.49	0.36	0.41	0.44	0.49	0.42	0.48	0.35	0.36	0.33	0.46	0.45
Memory (4)	0.57	0.57	0.58	0.58	0.58	0.56	0.57	0.58	0.46	0.51	0.55	0.57	0.53	0.57	0.42	0.47	0.43	0.58	0.58
Misc. (2)	0.57	0.57	0.58	0.58	0.58	0.56	0.57	0.58	0.45	0.53	0.55	0.58	0.55	0.57	0.41	0.42	0.40	0.57	0.57
Sup. learn. (3)	0.30	0.29	0.31	0.31	0.31	0.29	0.31	0.31	0.18	0.23	0.25	0.30	0.20	0.31	0.24	0.24	0.23	0.30	0.30
Mean (34)	0.46	0.46	0.47	0.48	0.48	0.45	0.47	0.48	0.35	0.39	0.43	0.47	0.40	0.47	0.32	0.34	0.32	0.47	0.46

Figure 7: Adapter rank and training-set size by family. (Left) Mean NLL against LoRA rank at full data. (Right) Mean NLL against the fraction of Psych-101 used for training, at r=16. Rows are the four model families; the dotted line is the cognitive-model baseline and the diamonds are the reproduced and reported Centaur-70B values. Subsets are nested and experiment-stratified, so every fraction covers all 160 experiments and reducing data quantity does not reduce paradigm coverage.

Table 9: Fraction of available information captured above chance, (ln⁡k−NLL)/ln⁡k, by task type on Psych-201 (out-of-distribution). A value of 0 indicates chance-level performance; 1 indicates perfect prediction. Restricted to 15 experiments (of 18) with a well-defined discrete response space (ln⁡k>0); 3 experiments with continuous or mixed responses are excluded. Bold+underline marks the best model, underline the second-best.
	Qwentaur	Llama-Centaur	Smoltaur	Olmotaur	Base	11
Task type	0.6B	1.7B	4B	8B	14B	1B	3B	8B	0.1B	0.4B	1.7B	3B	1B	7B	Q-8B	Q-14B	L-8B	Centaur-70B
Decision (2)	0.30	0.32	0.35	0.34	0.37	0.27	0.31	0.33	0.20	0.21	0.24	0.34	0.21	0.31	0.28	0.29	0.28	0.33
MDP (2)	0.56	0.56	0.58	0.58	0.58	0.48	0.53	0.57	0.39	0.46	0.52	0.57	0.42	0.57	0.51	0.53	0.53	0.50
Bandit (4)	0.32	0.35	0.37	0.36	0.38	0.33	0.33	0.36	0.24	0.23	0.32	0.34	0.28	0.34	0.22	0.25	0.25	0.39
Misc. (7)	0.28	0.31	0.45	0.46	0.49	0.14	0.30	0.40	0.03	0.11	0.18	0.36	0.09	0.35	0.37	0.41	0.29	0.52
Mean (15)	0.33	0.35	0.43	0.43	0.45	0.26	0.34	0.40	0.16	0.20	0.27	0.38	0.20	0.37	0.34	0.37	0.31	0.46

Figure 8: Reproducibility of Centaur-70B (4-bit) evaluation across Psych-101. Reproduced vs. reported NLL for Centaur-70B (4-bit) across 46 Psych-101 experiments (r=0.994, mean Δ=+0.024, median Δ=+0.006).

Table 12: Impact of cognitive fine-tuning across different benchmarks. Each cell shows Δ = fine-tuned − base, where fine-tuned models are trained with LoRA r=16 on the full dataset. Significance is assessed with a two-sided z-test; group means use Stouffer’s method to combine per-task z-scores. EQ-Bench is on a separate scale and excluded from means.
	Qwentaur	Llama-Centaur
Benchmark	0.6B	1.7B	4B	8B	14B	1B	3B	8B
MetaBench
ARC	-0.03	-0.01	-0.01	+0.01	-0.01	+0.01	-0.05	+0.01
GSM8K	-0.25∗∗∗	-0.15∗∗∗	-0.05	-0.17∗∗∗	-0.12∗∗∗	-0.05∗	+0.00	-0.10∗
HellaSwag	-0.02	-0.09	-0.02	-0.03	-0.03	-0.04	-0.03	-0.06
MMLU	-0.09	-0.05	-0.06	+0.02	+0.02	-0.04	-0.05	+0.01
TruthfulQA	-0.01	+0.03	-0.05	+0.05	+0.03	-0.01	+0.04	+0.05
Winogrande	-0.02	-0.02	-0.02	+0.03	-0.02	-0.03	-0.04	-0.01
Mean (6)	-0.07∗∗∗	-0.05∗	-0.03	-0.01	-0.02	-0.03	-0.02	-0.02
Ethics
CM	+0.01	+0.00	+0.07∗∗∗	+0.01	+0.04∗∗∗	-0.02∗	-0.03∗∗	+0.00
Deontology	+0.02∗	-0.03∗	-0.02	+0.03∗	-0.04∗∗∗	+0.00	+0.05∗∗∗	+0.02∗
Justice	+0.06∗∗∗	+0.00	-0.01	+0.17∗∗∗	-0.08∗∗∗	+0.00	+0.05∗∗∗	+0.03∗
Utilitarian	+0.00	+0.02∗	+0.09∗∗∗	+0.03∗∗∗	+0.09∗∗∗	+0.00	-0.04∗∗∗	-0.01
Virtue	+0.56∗∗∗	+0.02∗	-0.01	+0.03∗∗∗	-0.05∗∗∗	-0.01	+0.20∗∗∗	-0.16∗∗∗
Mean (5)	+0.13∗∗∗	+0.00	+0.02∗∗∗	+0.05∗∗∗	-0.01	-0.01	+0.04∗∗∗	-0.02∗∗∗
Cog. & Lang.
LogiQA	+0.00	-0.01	-0.02	-0.02	-0.05	+0.00	-0.02	+0.00
PIQA	-0.02	-0.01	+0.01	+0.02	+0.01	-0.01	+0.00	+0.00
Social IQA	+0.00	+0.01	+0.03∗	+0.00	+0.01	+0.01	+0.01	+0.01
CoQA (F1)	-0.11∗∗∗	-0.03	-0.02	-0.01	+0.00	-0.07∗∗	-0.05∗	-0.03
LAMBADA (OAI)	+0.00	-0.01	+0.00	+0.00	+0.01	+0.01	+0.02∗	+0.01
LAMBADA (Std)	+0.02	+0.01	+0.03∗∗	+0.02∗	+0.02	+0.02∗	+0.03∗∗∗	+0.02∗
EQ-Bench	-44.7∗∗∗	+5.1	+5.9	-21.3∗∗∗	-0.3	+16.0∗∗∗	+22.9∗∗∗	+7.4
Mean (6)	-0.02	-0.01	+0.01	+0.00	+0.00	-0.01	+0.00	+0.00
ACP (Planning)
App (B)	-0.10	-0.10	-0.09	-0.02	-0.05	-0.29∗∗∗	+0.08	+0.05
Areach (B)	-0.04	-0.22∗∗∗	-0.24∗∗∗	-0.24∗∗∗	+0.01	-0.37∗∗∗	+0.14∗	+0.02
Just (B)	+0.05	-0.06	-0.06	-0.12∗	-0.01	-0.34∗∗∗	-0.13∗	-0.06
Land (B)	-0.12∗∗	-0.44∗∗∗	-0.15∗∗	-0.21∗∗∗	+0.10	-0.11∗∗	-0.13∗∗	-0.04

Figure 9: Specificity of cognitive fine-tuning: comparison with non-cognitive control models. Mean NLL on Psych-101 for cognitively fine-tuned models (Llama-Centaur, Qwentaur; LoRA r=16, full training data) and size-matched non-cognitive and cognitive controls (Hermes, Nemotron, Be.FM). Models are grouped by parameter count. Cognitively fine-tuned models consistently outperform non-cognitive controls at every scale, confirming that the improvement is specific to the behavioural signal in Psych-101 and not an artefact of fine-tuning per se.

Table 13: Prior ablation conditions expressed in the four-channel notation. ∙ = present, — = removed, Imin = reduced to a minimal action-space definition, ⋅~ = content replaced by generic placeholders with formatting preserved, ⊗ = replaced by an instruction that contradicts the task, ⋅π = trial order permuted.
Source	Condition	𝑰	𝑺	𝑭	𝑪	Closest condition in ours
11	Original	∙	∙	∙	∙	Original
94	No psychological task	Imin	—	—	∙	History-only
	Zero-shot prediction	∙	∙	—	—	no analogue
57	Instruction free	—	∙	∙	∙	Instruction-ablated
	Misleading instruction	⊗	∙	∙	∙	no analogue
	Context free	—	—	—	∙	Choice-only
Ours	Instruction-ablated	—	∙	∙	∙	—
	Content-masked	Imin	S~	F~	∙	new
	History-only	Imin	—	—	∙	—
	Choice-only	—	—	—	∙	—
	Order-permuted	∙	Sπ	Fπ	Cπ	new

Figure 10: Impact of cognitive fine-tuning on MetaBench performance (Δ = fine-tuned − base). Each bar shows the change in accuracy for a matched base–fine-tuned pair on six standard LM benchmarks (ARC, GSM8K, HellaSwag, MMLU, TruthfulQA, Winogrande) plus their mean. Error bars show pooled standard errors. Positive values (green region) indicate improvement; negative values (red region) indicate degradation. Colour intensity scales with model size within each family.

Table 15: Per-experiment NLL under sequential ablation conditions for Centaur-70B (11) on Psych-101 (in-distribution). Columns correspond to progressive prompt degradation: orig retains the full prompt; inst removes task instructions; cont additionally masks stimulus values and feedback; hist further removes trial structure, leaving only the response history. The ln⁡(k) column shows the random-guessing baseline where k is the number of per-trial response options. Experiments with mixed or continuous response formats have no well-defined k and are shown as – .
Experiment	Type	orig	inst	cont	hist	ln⁡(k)
Gardening task (28)	Decision	0.49	0.49	0.69	0.69	0.69
Columbia card task (32)	Decision	0.21	0.24	0.28	0.29	0.69
Experiential-symbolic task (34)	Decision	0.46	0.46	0.85	1.43	—
Multi-attribute DM (42)	Decision	0.06	0.08	0.74	0.74	0.69
Risky choice (50)	Decision	0.43	0.64	0.82	0.89	—
choices13k (63)	Decision	0.43	0.44	0.52	0.56	0.69
CPC18 (64)	Decision	0.35	0.35	0.38	0.41	0.69
Decisions from description (93)	Decision	0.59	0.61	0.69	0.68	0.69
Two-step task (48)	MDP	0.48	0.53	1.18	1.23	0.69
Two-step task (49)	MDP	0.53	0.54	1.15	1.16	0.69
Virtual subway network (84)	MDP	1.16	1.46	1.41	1.30	1.61
Multi-task RL (83)	MDP	0.57	0.66	0.77	0.82	1.10
Two-step task (104)	MDP	0.51	0.52	0.97	1.03	0.69
Drifting four-armed bandit (6)	Bandit	0.71	0.85	0.84	0.86	1.39
Horizon task (27)	Bandit	0.40	0.39	0.53	0.55	0.69
Two-armed bandit (35)	Bandit	0.30	0.36	0.39	0.46	0.69
Prob. instrumental learning (54)	Bandit	0.50	0.49	2.07	2.01	0.69
Horizon task (69)	Bandit	0.58	0.59	0.63	0.68	0.69
Structured bandit (72)	Bandit	0.64	0.68	0.84	1.00	2.08
Horizon task (77)	Bandit	0.35	0.36	0.53	0.61	0.69
Iowa gambling task (80)	Bandit	0.91	1.08	0.97	0.97	1.39
Horizon task (88)	Bandit	0.15	0.15	0.34	0.43	0.69
Horizon task (89)	Bandit	0.48	0.48	0.56	0.62	0.69
Spatially correlated MAB (91)	Bandit	1.82	1.95	2.45	2.62	3.40
Decisions from experience (93)	Bandit	0.47	0.95	0.54	0.57	—
Changing bandit (96)	Bandit	0.45	0.45	1.28	0.98	0.69
Cond. assoc. learning (19)	Memory	0.52	0.55	1.03	1.06	1.10
Shepard categorization (5)	Sup. learn.	0.54	0.58	0.70	0.70	0.69
Multiple-cue judgment (20)	Sup. learn.	1.14	1.19	1.94	1.94	2.20
Medin categorization (56)	Sup. learn.	0.50	0.58	0.79	0.90	—

Figure 11: Impact of cognitive fine-tuning on Ethics benchmark performance (Δ = fine-tuned − base). Each bar shows the change in accuracy for a matched base–fine-tuned pair across five ethical reasoning tasks (commonsense morality, deontology, justice, utilitarianism, virtue) plus their mean. Error bars show pooled standard errors. Positive values (green region) indicate improvement; negative values (red region) indicate degradation. Colour intensity scales with model size within each family.

研究结果

在训练数据同类型的实验范围内(分布内),6亿到10亿参数的模型在负对数似然指标上追平了复现的700亿参数Centaur-70B模型,可比的8个模型全部聚集在仅0.028 nats的狭窄区间内。
在训练时未见过的全新实验上(分布外,Psych-201-RT),同样这8个模型的表现差距扩大到0.244 nats,且模型越大表现越好。
遮盖刺激和反馈的具体内容后,模型已学到信息的75.7%被摧毁,表现跌至随机猜测水平以下,而仅移除任务说明造成的信息损失只有12.5%。
在27个测试实验中,有18个(67%)呈现出随着信息被逐步移除而持续变差的单调趋势。
在试验彼此独立的THINGS相似性判断任务上,打乱试验顺序对预测几乎没有影响;而在后一次试验取决于前一次回答的延迟选择任务上,打乱顺序会显著改变预测结果。

Figure 12: Impact of cognitive fine-tuning on cognitive and language benchmark performance (Δ = fine-tuned − base). Each bar shows the change in accuracy for a matched base–fine-tuned pair across six tasks (LogiQA, PIQA, Social IQA, CoQA, LAMBADA-OpenAI, LAMBADA-Standard) plus their mean. Error bars show pooled standard errors. Positive values (green region) indicate improvement; negative values (red region) indicate degradation. Colour intensity scales with model size within each family.

可应用场景

用规模更小的微调模型替代巨型模型,作为预测人类选择行为的低成本研究工具
将这类小模型用作噪声上限基准,评估现有心理学理论在特定实验中还有多大改进空间
把逐个拆除提示词信息通道的方法用于检验其他行为预测模型是否真正利用了任务相关信息,而非捷径

Figure 18: Complete per-experiment ablation heatmaps for 26 experiments with a well-defined chance baseline. Five additional experiments with mixed or continuous response formats are excluded throughout because no single k defines a chance baseline. The probabilistic instrumental learning task (54) is also omitted. Each cell shows the fraction of learned information lost, δ=(ℒc−ℒorig)/(ln⁡k−ℒorig), averaged across all eight models. Columns represent individual experiments sorted by δ under the history-only condition; rows represent ablation conditions. (Top) 20 experiments with mean δ≤1 under the history-only condition (i.e. performance remains at or above chance). (Bottom) 6 experiments where ablation degrades performance below chance (δ>1); note the separate colour scale. The colour strip below each panel indicates task type (see legend).

局限与待验证事项

结果仅在与Psych-101中160种实验类型相似的任务上得到验证,对完全新颖类型实验的泛化能力尚未证实。
所有模型都只用低秩适配器(LoRA)微调,尚未测试全参数微调或专家混合、状态空间模型等其他架构在小规模下的表现。
结构性拆解测试仅覆盖27个实验,部分任务类型(如刺激本身就是需要记住的内容的记忆任务)因内容与反应无法分离而被排除在分析之外。
顺序可交换性的打乱测试只在两个性质相反的实验(独立试验的THINGS任务和顺序依赖的延迟选择任务)上进行。
认知微调被观察到会降低数学推理、形式化规划等通用基准上的表现,而对伦理推理基准的影响在不同模型间并不一致,作者对此未做进一步解读。

为什么重要

以往认为要用AI预测人类行为需要像700亿参数那样的巨型模型和大量计算资源,这项研究表明,只要不超出训练时见过的任务范围,规模小得多的模型也能达到相近效果。对心理学实验设计者和认知科学研究者来说,这类小型微调模型有望被用作衡量特定实验中现有理论还有多大改进空间的参照基准。

本文术语

Psych-101 · 汇集160个心理学实验、共1070万条逐次试验人类选择数据的数据集
分布内/分布外 · 分布内指与训练数据同类型的实验,分布外指训练时未见过的全新实验
LoRA(适配器) · 一种轻量级微调方法,不重新训练整个模型,只训练一个小的附加组件
噪声上限(noise ceiling) · 由于人类行为本身存在不可消除的随机性,任何模型能达到的预测准确率理论上限
可交换性(exchangeability) · 各次试验彼此独立、顺序不携带任何预测信息的性质

论文原文摘要（英文）

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. In-distribution, 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.

作者 · Nick Oh

在 arXiv 阅读

METAL MEDIA 最新报道

图片来源： Nick Oh et al., arXiv:2608.05224, CC BY 4.0