Reasoning under uncertainty in humans and language models
Authors: Gonçalo Guiomar, Elia Torre, Victoria Shavina, Pehuen Moure, Mario Giulianelli, Shih-Chii Liu, Valerio Mante
Organizations: Equal contribution · ETH AI Center, Zürich, Switzerland · Institute for Neuroinformatics, University of Zürich & ETH Zürich, Switzerland · ETH Zürich, Switzerland · University College London, London, United Kingdom.
Do language models make decisions under uncertainty like humans do? And if so, what role does extended reasoning play in the underlying decision process? We answer this question by introducing an active probabilistic reasoning task that cleanly separates sampling (actively acquiring evidence) from inference (integrating evidence towards a decision). Benchmarking humans and a broad set of contemporary LLMs against an optimal reference policy reveals a consistent pattern: extended reasoning is the key determinant of strong performance, driving large gains in inference, while yielding only modest improvements in active sampling. To explain these differences, we fit a behavioral model that captures systematic deviations from optimal Bayesian behavior through interpretable parameter families, placing humans and models in a shared low-dimensional cognitive space. The resulting fits show how reasoning shifts models toward human-like regimes of evidence accumulation and belief-to-choice mapping, and yield testable predictions about the latent dynamics that might drive each decision. Probing residual-stream activations of an open-weight reasoning model, we find that the geometry of internal representations tracks these predicted dynamics, linking behavior to representational correlates of the fitted latent dynamics.
Figures & tables
Figure 1: From task performance to latent cognitive variables. A: Task. An active probabilistic reasoning task in which agents sequentially sample from up to four buttons (A–D), each revealing a binary outcome (RED/GREEN). One button is biased toward RED, while the others are unbiased. During T sampling rounds, agents sample evidence among the buttons available. During inference round, agents indicate which button they believe is biased. An equivalent text-based version is used for LLMs (App. B.1 ). B: Behavior. Comparison of human and LLM behavior by overall performance , sampling/inference quality , and invalid choices . These metrics reveal a broad spectrum of performance, with extended reasoning improving overall success via enhanced inference , while gains in sampling remain limited. C: Mechanisms. Behavioral model that captures deviations from optimal Bayesian inference using interpretable parameters: Memory ( β ), Strategy ( m,κf ), Choice Bias ( ω,γ ), and Occlusion Awareness ( θs,θf ). D: Cognitive space. These parameters define a shared low-dimensional space for humans and LLMs. Reasoning shifts LLMs toward human-like inference strategies, and tightens the gap in sampling strategies.
Figure 2: Extended reasoning enables LLMs to match or exceed humans in inference, but not in sampling. Models are grouped into Reasoning , CoT , and Base (App. B.2 ); model-name colors indicate undisclosed (black), mixture-of-experts (orange), or dense (purple) architectures. A: Task performance. Mean success rate across trial lengths T∈{2,…,15} . Humans are shown in green, the optimal dynamic-programming agent in blue, and extended-reasoning conditions as gray overlays. The vertical line marks chance (25%). B: Sampling and inference quality. Performance losses relative to the optimal agent (lower is better). Sampling loss compares the optimal reference with MAP inference applied to each agent’s sampled evidence; inference loss compares this counterfactual MAP performance with actual success. Error bars show bootstrap standard deviations ( B=1,000 ). Asterisks indicate significance of differences from human performance (black: baseline; gray: extended reasoning; n.s.: not significant). C: Invalid choices. Fraction of invalid sampling and final choices, including occluded, malformed, or missing responses. Extended reasoning primarily improves inference, with smaller gains in sampling and fewer invalid choices.
Figure 3: The behavioral model reproduces behavior, and each parameter has a distinct signature. A: Behavioral model. The posterior pt is integrated into memory ht with a kernel set by β (light orange, dashed: stubborn ; dark orange: forgetful ), which drives the sampling ( πts ) and final ( πf ) policies. B: Posterior dynamics. Posterior of the chosen cue in real (solid) and model-simulated (dashed) games of 15, 10 and 5 rounds. C: Parameter signatures. Single-parameter perturbations of the model fitted to the optimal agent (black; 15-round games). D: Loss increase in sampling (left) and inference (right) for these perturbations. E: Model vs. data loss. Sampling ( ▲ ) and inference ( ∙ ) loss of all 38 agents (open: extended reasoning ).
Figure 4: Behavioral parameters characterize sampling and inference in humans and language models. Each point shows the 3-fold CV fit average of one fitted agent (fit quality in App. J.5 ). LLMs are colored by their position in the benchmark in Fig. 2 , humans are shown in green and the optimal agent in blue; filled markers denote base and open markers extended reasoning , with paired conditions connected by light gray lines. A : Memory β versus task performance, spanning stubborn ( β<0 ) to forgetful ( β>0 ) evidence integration. B : Sampling strategy scale m versus performance. m controls the linear temporal sampling gain κs(t) . C : Choice inertia γ versus performance, capturing perseveration on the most recent valid sampling choice. D : Inference sharpness κf versus performance, controlling how decisively memory is mapped to the final inference policy. E : Bias spread summarizes systematic choice preference strength as the mean range of sampling and inference bias vectors. F : Occlusion awareness θ versus performance, capturing avoidance of invalid or occluded choices. G–J : Sampling & inference loss . Inference loss against inference sharpness κf ( G ) and memory β ( I ); sampling loss against the sampling gain m ( H ) and memory β ( J ). Inference loss decreases with κf , whereas sampling loss tends to rise for both stubborn and forgetful integration. Dots mark pooled humans (green) and the optimal agent (blue; its κf=10.5 lies outside G ).
Figure 5: Activation geometry is consistent with latent dynamics predicted by the behavioral model. A: Optimal Bayesian posterior. Evidence sequence (top) and Bayesian posterior pt (bottom) for an example game. B–C: Latent-variable dynamics and centroid distances for gpt oss 20b under high - ( B ) and low-reasoning ( C ). For each condition, we show the memory state ht (top), policy πt (middle), and softmax-normalized centroid-distance vector computed from the model’s activations (bottom; see App. I ). Under high reasoning , ht tracks the posterior switch from C to A ( β≈0 ) and πt becomes sharply peaked ( κf≫1 ); under low reasoning , ht remains stubborn ( β<0 ) and πt stays diffuse ( κf≈0 ). The activation-derived distance vectors recapitulate these dynamics. Dashed boxes mark the example rounds shown in panels E–F . D: Choice decodability across layers. Cross-validated R2 of ridge probes predicting the agent’s choice from residual-stream activations ( high in purple, low in teal). Swapped controls (SC; green/yellow) train probes on one condition and evaluate on the other; SC R2 collapses to near zero. E–F: Activation trajectories of the example game projected into the layer-1 PCA basis under high ( E ) and low ( F ) reasoning . Filled circles mark the per-choice centroids; the trajectory is colored by the closest centroid at each round. Under high reasoning , the trajectory begins near C and migrates toward A ; under low reasoning , it remains anchored near C throughout the game. Pairwise PC views are shown in App. Fig. 11 .
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Predictors
Success rate
β
κf
m
Bias spread
Model size
0.088
0.116
0.147
0.059
0.084
Architecture
0.416
0.002
0.555
0.172
0.004
Training regime
0.862
0.068
0.589
0.305
0.319
Model size + architecture
0.428
0.154
0.561
0.172
0.173
Training regime + architecture
0.906
0.105
0.756
0.338
0.348
Model size + training regime
0.889
0.214
0.666
0.333
0.433
Appendix
Table 1: Training regime, architecture, and model size as predictors of behavior. R2 values from linear regressions predicting task success and fitted behavioral parameters across 19 open-source models, using model size, architecture (dense vs. MoE), and training regime (Base, CoT, Reasoning) as covariates.
Model
Reasoning
N
LLM loss
Human loss
Diff
95% CI
p (raw)
Sig.
gpt-5-mini
base
5400
0.079
0.015
-0.064
[-0.083, -0.045]
<10−4
grok-3-mini
base
5800
0.062
0.015
-0.047
[-0.066, -0.029]
<10−4
gemini-2.5-flash
base
2800
0.074
0.015
-0.060
[-0.083, -0.037]
<10−4
gemini-2.5-pro
base
2800
0.071
0.015
-0.056
[-0.079, -0.033]
<10−4
gpt-oss-120b
base
2800
0.076
0.015
-0.062
[-0.085, -0.039]
<10−4
gpt-5
base
2800
0.073
0.015
-0.058
[-0.081, -0.035]
<10−4
Appendix
Table 2: Sampling loss comparison between average human performance and all tested models (lower loss indicates better sampling). The human reference pool is composed of 46 participants (4,600 games); mean human sampling loss is 0.0148. Sampling loss is 0.68−MAPsuccess , using the fixed 68% Optimal Agent (DP) reference. N is the number of LLM games. Diff =Humanloss−LLMloss ; negative values favor humans. The 95% CIs are percentile intervals from 10,000 independent trial-level bootstrap resamples of both groups (seed 42). The p -values are two-sided permutation estimates using 100,000 random assignments and (b+1)/(100,000+1) , where b counts equally or more extreme assignments. Sig.: ∗p<0.05 , ∗∗p<0.01 , ∗∗∗p<0.001 ; n.s., not significant.
Model
Reasoning
N
LLM loss
Human loss
Diff
95% CI
p (raw)
Sig.
gpt-5-mini
base
5400
0.114
0.058
-0.056
[-0.073, -0.039]
<10−4
grok-3-mini
base
5800
0.017
0.058
+0.040
[+0.026, +0.054]
<10−4
gemini-2.5-flash
base
2800
0.228
0.058
-0.171
[-0.194, -0.148]
<10−4
gemini-2.5-pro
base
2800
0.188
0.058
-0.130
[-0.153, -0.107]
<10−4
gpt-oss-120b
base
2800
0.088
0.058
-0.031
[-0.051, -0.011]
0.001900
**
gpt-5
base
2800
0.156
0.058
-0.098
[-0.121, -0.077]
<10−4
Appendix
Table 3: Inference loss comparison between average human performance and all tested models (lower loss indicates better inference). The human reference pool is composed of 46 participants (4,600 games); mean human inference loss is 0.0576. Inference loss is MAPsuccess−actualsuccess . N is the number of LLM games. Diff =Humanloss−LLMloss ; negative values favor humans. The 95% CIs are percentile intervals from 10,000 independent trial-level bootstrap resamples of both groups. The p -values are two-sided permutation estimates using 100,000 random assignments and (b+1)/(100,000+1) , where b counts equally or more extreme assignments. Sig.: ∗p<0.05 , ∗∗p<0.01 , ∗∗∗p<0.001 ; n.s., not significant.
Figure 6: Reasoning token budget does not explain the selective improvement in inference. Average number of reasoning tokens produced by gpt oss 20b per round, shown separately for sampling and inference phases under low- and high-reasoning conditions, across all game lengths ( T∈{2,…,15} ). Under high reasoning, sampling rounds consume substantially more tokens ( ∼8,000 – 12,000 ) than inference decisions ( ∼2,000 – 5,000 ). If extended reasoning improved performance primarily by providing a larger compute budget, the phase receiving more tokens should benefit most. Instead, the opposite holds: reasoning reduces inference loss by 10.9 pp but sampling loss by only 3.0 pp. This dissociation is difficult to reconcile with a pure compute-budget explanation, and instead suggests that reasoning selectively enables the kind of computation relevant to evidence integration—consistent with the behavioral model’s account of improved inference via more accurate memory ( β≈0 ) and sharper belief-to-choice mapping (larger κf ).
H
2
3
4
5
6
7
8
9
10
11
12
13
14
15
JH⋆
41.7
48.5
54.2
58.9
62.8
66.3
69.3
72.0
74.4
76.6
78.5
80.3
81.9
83.4
Appendix
Table 4: Exact DP ceiling JH⋆ (expected success, %) by horizon for K=4 , αB=0.9 , αU=0.5 , q=0.5 .
Figure 7: Evolution of task performance across rounds. Mean success rate (correct identification of the biased button) as a function of the number of sampling rounds T∈{2,…,15} . Human performance is shown in green, pooled across participants; LLMs are shown as individual traces, colored by overall success rate (orange: lower; purple: higher). Higher-performing models exhibit a clear increase in success with additional evidence, whereas lower-performing models remain comparatively flat, often close to chance (25%). We quantify this lift-off as a predicted improvement exceeding 5 percentage points (pp) between T=2 and T=15 . Claude haiku 3.5 is the lowest-ranked model meeting this criterion ( Δ=17.5 pp, 95% CI [11.6,23.3] pp, pHolm=0.0002 ). Across the eight configurations evaluated under extended reasoning, all exhibit lift-off, with a mean improvement of 40.0 pp, whereas only one Base configuration ( gpt 4.1 mini ) meets this criterion. Extended-reasoning gains exceed those of CoT and Base configurations by 29.4 pp ( pHolm=0.0015 ) and 37.1 pp ( pHolm=0.00003 ), respectively. These results suggest that extended reasoning improves the ability to leverage additional evidence for final inference decisions.
Figure 8: Individual human performance. Each row corresponds to one of the 46 participants who completed 100 games (trial lengths uniformly sampled from T∈{2,…,15} ). Participants are ordered by mean success rate. Mean success rate (correct identification of the biased option at the final decision); the vertical line marks chance performance (25%). The cohort exhibits substantial heterogeneity, spanning ≈37% to ≈77% success, with most participants concentrated around ∼55 – 67% .
Figure 9: Behavioral metrics are recapitulated by the fitted behavioral model. A–C : Empirical behavioral metrics computed directly from human and LLM trajectories. A : Overall task performance (success rate) across agents, as in Fig. 2 . B : Decomposition of performance loss into sampling versus inference deficits, measured by comparing each agent to counterfactual and optimal reference policies (loss w.r.t. optimal sampling on the left; loss w.r.t. optimal inference on the right). C : Invalid-choice rates during sampling and final report, as in Fig. 2 . D–F : The same metrics computed from trajectories generated by the behavioral model of Fig. 3 after fitting (β,m,γ,κf,ω,θ) to each agent. D : Model-predicted success rates closely match empirical performance ordering. E : Model reproduces the sampling–inference loss decomposition, capturing that reasoning primarily reduces inference loss while leaving a residual sampling gap. F : Model reproduces invalid-choice profiles via the occlusion-awareness component. Green denotes pooled humans; grey indicates extended-reasoning conditions.
Figure 10: Model parameters explain human and LLM behavior. A : Memory β spans stubborn ( β<0 ) to forgetful ( β>0 ) evidence integration. B : Strategy separates sampling gain m in the sampling temporal kernel ks(t) from final-choice sharpness κf . C : Choice inertia γ captures perseveration on the most recent valid sampling choice. D : Bias spread (difference between max and min ωx bias vector components) summarizes systematic sampling and inference choice preferences ( ωs,ωf ). E : Occlusion awareness θ captures avoidance of invalid or occluded choices.
Figure 11: PCA geometry of layer-1 activations. A: Per-PC variance explained for the conditional-average dataset at layer 1, shown separately for low (purple) and high (green) reasoning. PC1 captures 27.8% of the variance under low reasoning and 19.6% under high reasoning ; the flatter spectrum under high reasoning indicates richer round-by-round dynamics distributed across more components. B: Cumulative variance explained for the same data; the top four PCs together account for ∼64% ( low ) and ∼61% ( high ) of the total variance. C–D: Pairwise 2D projections of the example game’s layer-1 activations (same game as in Fig. 5 ) across the top four PCs, under high reasoning ( C ) and low reasoning ( D ). Per-choice centroids are shown as filled circles colored by choice (A: red, B: blue, C: yellow, D: green); the trajectory’s start (round 2) and end (round 16) are marked by open and filled black circles. Across all PC pairs, the high-reasoning trajectory consistently migrates from the neighborhood of centroid C toward centroid A as evidence accumulates, whereas the low-reasoning trajectory remains anchored near C throughout the game.
Setting
Symbol
Selected value
Choice likelihood weight
λchoice
1.0
Sampling-bias regularizer
λω
0.005
Inertia regularizer
λγ
0.005
Memory bound
β
[−3,3]
Sampling-gain bound
m
[0.05,7]
Inertia bound
γ
[0.05,3]
Appendix
Table 5: Selected gamma-fit objective weights and parameter bounds. Alignment losses are ordered as written in the text.
Figure 12: Posterior evolution by rounds and bias across humans and language models. In each row the first two panels show the true and model-generated posterior evolution for the final choice ( inference ), followed by the true and model-generated sampling-choice posterior ( sampling ). The remaining panels show the true bias for sampling and inference and their corresponding fitted bias weights.
Figure 13: Posterior evolution by rounds and bias across humans and language models. In each row the first two panels show the true and model-generated posterior evolution for the final choice ( inference ), followed by the true and model-generated sampling-choice posterior ( sampling ). The remaining panels show the true bias for sampling and inference and their corresponding fitted bias weights.
Figure 14: Posterior evolution by rounds and bias across humans and language models. In each row the first two panels show the true and model-generated posterior evolution for the final choice ( inference ), followed by the true and model-generated sampling-choice posterior ( sampling ). The remaining panels show the true bias for sampling and inference and their corresponding fitted bias weights.
Figure 15: Posterior evolution by rounds and bias across humans and language models. In each row the first two panels show the true and model-generated posterior evolution for the final choice ( inference ), followed by the true and model-generated sampling-choice posterior ( sampling ). The remaining panels show the true bias for sampling and inference and their corresponding fitted bias weights.
Figure 16: Posterior and bias structure is captured by the mechanistic model. For each agent (rows), we compute phase-specific summaries during sampling (left block) and inference (middle block), and compare empirical estimates ( True ) to quantities generated by the fitted mechanistic model ( Model ). Sampling : heatmaps show the agent’s True Posterior signal (normative posterior mass assigned to the agent’s eventual choice under the game history) alongside the corresponding Model Posterior obtained by unrolling the fitted belief dynamics. Inference : analogous comparison for the final inference phase. Bias : right block contrasts True Bias (empirical action preference beyond belief) with Model Bias recovered from the multiplicative bias parameters. Across agents, the model reproduces both the magnitude and qualitative structure of posterior dynamics and systematic biases, supporting the claim that (β,m,κf,ω,θs,θf,γ) provide a sufficient explanation of the behavioral signatures.
Model
Reasoning
Δ eLPD/game
CI low
CI high
p-value
FI
Optimal agent (DP)
–
8.953
8.707
9.196
< 0.001
99.6%
Humans
–
6.518
6.401
6.636
< 0.001
98.1%
Base models
gpt-5-mini
Base
6.962
6.756
7.160
< 0.001
98.6%
grok-3-mini
Base
7.354
7.159
7.548
< 0.001
98.4%
gemini-2.5-flash
Base
7.716
7.492
7.936
< 0.001
99.5%
Appendix
Table 6: Per held-out game, improvement is δi=logp(yi∣M)−logp(yi∣B) , where M is the fitted model and B is an agent-specific frequency baseline. Reported Δ eLPD/game is δˉ=N1∑i=1Nδi across out-of-sample games (repeated 5-fold CV; repeated predictions per game averaged first). CI low/high : bootstrap confidence interval for δˉ . p-value : sign-flip permutation test for H0:δˉ≤0 vs H1:δˉ>0 (multiple-testing adjusted). FI : fraction of held-out games with δi>0 .
Figure 17: A , Fit quality and parameter dispersion across repeated fits. For each agent, the model was fit repeatedly on held-out data splits with multiple random initializations, retaining the best fit within each split by held-out log-likelihood. Each dot is one of the 15 repeated fits per agent (3 cross-validation folds × 5 initializations). B , Parameter recovery for the best model across agents. For each agent, synthetic datasets were generated from the selected fitted parameters. The model was then refit using the same fitting procedure; each dot is a recovered parameter from one of the 15 refits (3 folds × 5 initializations).
Figure 18: Pairwise likelihood landscapes gpt-5-mini Extended : All parameters were fixed at their best fitted values except for one pair, which was varied over a two-dimensional grid while evaluating the negative log-likelihood of the empirical data. τs and τf are not model parameters: they rescale the fitted sampling and final-choice bias vectors, ωs(τ)∝ωs1/τs and ωf(τ)∝ωf1/τf (renormalized), so τ=1 is the fitted bias, τ<1 sharpens it and τ>1 flattens it toward uniform (log-spaced grid, τ∈[0.35,3] ); for each repeated fit, τ is set by the inverse slope of its centered log-bias against the fitted one. Each dot is one of the 15 repeated fits (3 folds × 5 initializations); the star marks the fold-averaged fit.
Figure 19: Pairwise likelihood landscapes Humans : All parameters were fixed at their best fitted values except for one pair, which was varied over a two-dimensional grid while evaluating the negative log-likelihood of the empirical data. τs and τf are not model parameters: they rescale the fitted sampling and final-choice bias vectors, ωs(τ)∝ωs1/τs and ωf(τ)∝ωf1/τf (renormalized), so τ=1 is the fitted bias, τ<1 sharpens it and τ>1 flattens it toward uniform (log-spaced grid, τ∈[0.35,3] ); for each repeated fit, τ is set by the inverse slope of its centered log-bias against the fitted one. Each dot is one of the 15 repeated fits (3 folds × 5 initializations); the star marks the fold-averaged fit.
Figure 20: Pairwise likelihood landscapes Optimal Agent (DP) : All parameters were fixed at their best fitted values except for one pair, which was varied over a two-dimensional grid while evaluating the negative log-likelihood of the empirical data. τs and τf are not model parameters: they rescale the fitted sampling and final-choice bias vectors, ωs(τ)∝ωs1/τs and ωf(τ)∝ωf1/τf (renormalized), so τ=1 is the fitted bias, τ<1 sharpens it and τ>1 flattens it toward uniform (log-spaced grid, τ∈[0.35,3] ); for each repeated fit, τ is set by the inverse slope of its centered log-bias against the fitted one. Each dot is one of the 15 repeated fits (3 folds × 5 initializations); the star marks the fold-averaged fit.
Model
Reasoning
ωs
ωf
β
m
γ
κf
θs
Optimal agent (DP)
–
[0.252, 0.250, 0.246, 0.252]
[0.238, 0.220, 0.251, 0.292]
-0.004
3.763
0.646
10.513
55.004
Humans
–
[0.279, 0.268, 0.232, 0.221]
[0.257, 0.261, 0.239, 0.243]
0.081
1.742
0.059
1.707
29.322
gpt-5-mini
Base
[0.594, 0.285, 0.098, 0.023]
[0.467, 0.317, 0.160, 0.056]
-0.101
0.602
0.415
1.230
5.676
gpt-5-mini
Extended
[0.359, 0.281, 0.232, 0.129]
[0.364, 0.294, 0.208, 0.135]
-0.019
0.701
0.257
1.992
8.590
grok-3-mini
Base
[0.289, 0.329, 0.207, 0.176]
[0.392, 0.273, 0.192, 0.143]
-0.012
0.665
0.050
1.709
12.040
grok-3-mini
Extended
[0.313, 0.309, 0.216, 0.163]
[0.379, 0.304, 0.185, 0.132]
0.013
0.675
0.050
1.941
9.119
Appendix
Table 7: Fitted parameters of the gamma memory model. Values are best-by-fold averages from the selected fit.
Model
Reasoning
Train NLL
Test NLL
NLL
Optimal agent (DP)
–
0.303
0.303
0.303
Humans
–
0.657
0.657
0.657
gpt-5-mini
Base
0.613
0.615
0.615
gpt-5-mini
Extended
0.653
0.654
0.654
grok-3-mini
Base
0.664
0.666
0.666
grok-3-mini
Extended
0.710
0.712
0.712
Appendix
Table 8: Negative log-likelihoods of the mechanistic model repeated fold fits. Values are averages over the 15 repeated fold fits per model-condition.
Departament de Filologia Catalana, Universitat Autònoma de Barcelona, Barcelona 08193, Spain. · Institut für Psychologie, Humboldt-Universitat zu Berlin, Berlin 10099, Germany. · Institució Catalana de Recerca i Estudis Avançats (ICREA), Barcelona 08010, Spain