Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models

Ayush Rajesh Jhaveri; Anthony GX-Chen; Ilia Sucholutsky; Eunsol Choi

cs.CLApr 2, 2026

Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models

Authors: Ayush Rajesh Jhaveri, Anthony GX-Chen, Ilia Sucholutsky, Eunsol Choi

Organizations: New York University

Abstract

Confirmation bias, the tendency to seek evidence that supports rather than challenges one's belief, hinders one's reasoning ability. We examine whether large language models (LLMs) exhibit confirmation bias by adapting the rule-discovery study from human psychology: given a sequence of three numbers (a "triple"), an agent engages in an interactive feedback loop where it (1) proposes a new triple, (2) receives feedback on whether it satisfies the hidden rule, and (3) guesses the rule. Across eleven LLMs of multiple families and scales, we find that LLMs exhibit confirmation bias, often proposing triples to confirm their hypothesis rather than trying to falsify it. This leads to slower and less frequent discovery of the hidden rule. We further explore intervention strategies (e.g., encouraging the agent to consider counter examples) developed for humans. We find prompting LLMs with such instruction consistently decreases confirmation bias in LLMs, improving rule discovery rates from 42% to 56% on average. Lastly, we mitigate confirmation bias by distilling intervention-induced behavior into LLMs, showing promising generalization to a new task, the Blicket test. Our work shows that confirmation bias is a limitation of LLMs in hypothesis exploration, and that it can be mitigated via injecting interventions designed for humans.

Figures & tables

Figure 1: Confirmation bias leads to narrow exploration. We show two trajectories for rule discovery task, where an agent aims to infer a hidden numerical rule over multiple turns. Starting from an initial triple, the agent guess a hypothesis and test a new triple, receiving binary feedback on whether the proposed triple satisfies the hidden rule. A compatible test is consistent with the agent’s current hypothesis, whereas an incompatible test contradicts it. Trajectory 2 proposes compatible tests in both turns, showing confirmation bias. In contrast, Trajectory 1 introduces incompatible test in the second turn, allowing elimination of incorrect hypothesis and lead to correct hidden rule discovery.

Rule group	Initial triple samples
All odd; $b$ is (strict) max; mixed signs; at least one multiple of 3	$(-9,\,55,\,-71)$ , $(-77,\,75,\,-47)$
All prime numbers; non-increasing; all positive; contains a prime	$(79,\,43,\,31)$ , $(89,\,67,\,2)$

Table 1: Example rule sets (each set has 4 rules) and sampled triples. Each rule is defined over integer triples

(a,b,c)

. All triples are sampled from the feasible intersection of the rules.

	Non-Thinking				Thinking
	Task Suc.	Avg. # Turns	Avg. # Tkns	I:C	Task Suc.	Avg. # Turns	Avg. # Tkns	I:C
	% $\uparrow$	until Suc. $\downarrow$	per Turn $\downarrow$	$\uparrow$	% $\uparrow$	until Suc. $\downarrow$	per Turn $\downarrow$	$\uparrow$
	Qwen3-8B				Qwen3-8B
Baseline	0.06	5.00	8.4	0.35	0.21	10.88	1088.2	0.15
TiO	0.06	15.25	8.4	0.30	0.36	9.01	1716.1	0.35
Dual-Goal	0.08	0.00	13.4	0.57	0.40	7.37	2142.7	0.55

Table 2: LLMs task performance and the extent of confirmation bias (I:C ratio) on Wason’s rule discovery game. We observe varying success rates for LLMs. Overall, larger models with thinking mode show higher task success rates. TiO represents the Think-in-Opposites intervention.

Panel of Figure 2 — Figure 2: Confirmation bias correlates with task success. Higher I:C (more disconfirmatory testing) is associated with higher task success. Each point represents a (model, variant) averaged over 80 episodes, and shaded regions show 95% confidence intervals.

Panel of Figure 3 — Figure 3: Interventions shift exploration in thinking but not non-thinking models. Red denotes Baseline; blue denotes intervention runs (Dual-Goal / Think-in-Opposites). Ellipses show 95% covariance regions. Each point represents a (model, variant) averaged over 80 episodes.

Student (+ Teacher)	Prompt	Task	# Turns	I:C
		Suc.	until Suc.
Qwen3-8B	-	0.21	10.88	0.15
Qwen3-8B	TiO	0.36	7.37	0.35
Qwen3-32B	-	0.41	10.83	0.14
Qwen3-32B	TiO	0.61	6.38	0.54
Intervention Distillation

Table 3: Model performance after intervention distillation. We mark distillation performance gap that are statistically significant (from its base model) with superscript ∗ for task success and I:C ratios.

	Non-Thinking			Thinking
	Task Suc.	# Turns	I:C	Task Suc.	# Turns	I:C
	% $\uparrow$	until Suc. $\downarrow$	$\uparrow$	% $\uparrow$	until Suc. $\downarrow$	$\uparrow$
	Qwen3-8B			Qwen3-8B
Baseline	0.11	10.24	0.36	0.32	10.26	0.44
TiO	0.19	12.14	0.25	0.40	10.73	0.57
	Qwen3-14B			Qwen3-14B

Table 4: LLM Task performance in the Blicket Task. Each metric is aggregated across 192 evaluation episodes.

Panel of Figure 4 — Figure 4: Confirmation bias correlates with task success. Higher I:C is associated with higher task success. Shaded regions show 95% confidence intervals. Each point represents a (model, variant) averaged over 192 episodes.

Models Distilled with Wason Task
Model	Prompt	Task Suc.	# Turns	I:C
		(until Suc.)
Qwen3-8B	Base	0.32	10.26	0.44
Qwen3-8B	TiO	0.40	10.73	0.57
Qwen3-32B	Base	0.57	12.18	0.54
Qwen3-32B	TiO	0.86	12.98	1.31

Table 5: Performance of distilled models trained on the Wason task are evaluated on the Blicket Test baseline. We mark distillation performance gap that are statistically significant (from its base model) with superscript ∗ for task success and I:C ratios.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Rule	Predicate (over a,b,c)
All even	a%2==0 and b%2==0 and c%2==0
Each divides next	a!=0 and (b%a==0) and b!=0 and (c%b==0)
Exactly two equal	len({a,b,c})==2
At least one even	(a%2==0) or (b%2==0) or (c%2==0)
All end with 6	abs(a)%10==6 and abs(b)%10==6 and abs(c)%10==6
Increasing differences	0 < (b-a) < (c-b)

Appendix

Table 6: Rules (Train).

Rule	Predicate (over a,b,c)
All end with 9	abs(a)%10==9 and abs(b)%10==9 and abs(c)%10==9
Non-decreasing differences	(b-a) <= (c-b)
c is max	c>=a and c>=b
Arithmetic progression (AP)	(b-a)==(c-b)
All divisible by 7	a%7==0 and b%7==0 and c%7==0
Exactly two even	((a%2==0)+(b%2==0)+(c%2==0))==2

Appendix

Table 7: Rules (Validation).

Rule	Predicate (over a,b,c)
All end with 1	abs(a)%10==1 and abs(b)%10==1 and abs(c)%10==1
c is min	c<=a and c<=b
All negative	a<0 and b<0 and c<0
Descending	a>b>c
All odd	a%2==1 and b%2==1 and c%2==1
b is (strict) max	b>a and b>c

Appendix

Table 8: Rules (Test).

ID	Split	Rule 1 (Human)	Rule 2	Rule 3	Rule 4	Feasible $\|S_{g}\|$
1	Train	All even	Each divides next	Exactly two equal	At least one even	1,629
2	Train	All end with 6	Increasing differences	a is min	All distinct	5,394
3	Train	All divisible by 5	a is max	Non-monotone (middle between ends)	At least two multiples of 5	128,080
4	Train	All divisible by 3	Alternating parity (ends same)	Ascending	Non-decreasing	194
5	Validation	All end with 9	Non-decreasing differences	c is max	Arithmetic progression (AP)	2,550
6	Validation	All divisible by 7	Exactly two even	At least one multiple of 4	At least two distinct	72

Appendix

Table 9: Rule groups and their feasible intersections. The rightmost column reports the number of triples

|S_{g}|

[-99,100]^{3}

that satisfy all four rules in the group.

Split	Group ID	Triple
Test	7	$(-1,\,-81,\,-91)$
Test	7	$(-61,\,-71,\,-91)$
Test	7	$(-1,\,-21,\,-81)$
Test	7	$(-21,\,-41,\,-91)$
Test	7	$(-11,\,-31,\,-41)$
Test	8	$(-77,\,75,\,-47)$

Appendix

Table 10: Sampled Test instances (groups 7–10). Each triple is used as a starting point for four episodes, one per rule in the corresponding group.

Model	Variant	Task	First	# Turns until	I:C ${}_{\text{sol}}$	I:C ${}_{\text{uns}}$	I:C ${}_{\text{all}}$	# Tkns per Turn
		Suc. $\uparrow$	Guess	Suc. $\downarrow$	$\uparrow$	$\uparrow$	$\uparrow$	Avg ${}_{\text{sol}}$	Avg ${}_{\text{uns}}$	Avg ${}_{\text{all}}$
Qwen3-8B	Baseline	0.212	0.087	10.875	0.101	0.150	0.147	958.2	1123.2	1088.2
	Dual-Goal	0.400	0.075	7.365	0.813	0.525	0.546	1763.2	2394.1	2141.7
	Think-in-Opposites	0.363	0.000	9.008	0.826	0.314	0.353	1547.9	1811.9	1716.1
Qwen3-14B	Baseline	0.338	0.125	7.393	0.130	0.192	0.186	991.5	1274.1	1178.6
	Dual-Goal	0.575	0.138	7.661	0.609	0.369	0.404	1029.1	1411.8	1191.7

Appendix

Table 11: Results across all rules and episodes for Thinking Models.

Model	Variant	Task	First	# Turns until	I:C ${}_{\text{sol}}$	I:C ${}_{\text{uns}}$	I:C ${}_{\text{all}}$	# Tkns per Turn
		Suc. $\uparrow$	Guess	Suc. $\downarrow$	$\uparrow$	$\uparrow$	$\uparrow$	Avg ${}_{\text{sol}}$	Avg ${}_{\text{uns}}$	Avg ${}_{\text{all}}$
Qwen3-8B	Baseline	0.062	0.013	5.000	0.174	0.354	0.353	8.5	8.3	8.4
	Dual-Goal	0.075	0.075	0.000	–	0.574	0.574	21.0	12.7	13.4
	Think-in-Opposites	0.062	0.000	15.250	0.171	0.298	0.295	7.3	8.5	8.4
Qwen3-14B	Baseline	0.225	0.175	3.000	0.600	0.724	0.723	3.8	8.7	7.6
	Dual-Goal	0.125	0.113	0.500	0.000	0.411	0.411	13.7	15.2	15.0

Appendix

Table 12: Results across all rules and episodes for Non-thinking Models.

Student $\leftarrow$ Teacher	LR sweep range	Best LR	Best step (epoch)
Qwen3-8B $\leftarrow$ Qwen3-8B (Think-in-Opposites)	$[1{\times}10^{-5},\,5{\times}10^{-4}]$ (7 values)	$7.5{\times}10^{-5}$	3400 (3.02)
Qwen3-8B $\leftarrow$ Qwen3-32B (Baseline)	$[1{\times}10^{-4},\,2{\times}10^{-3}]$ (6 values)	$2{\times}10^{-4}$	3200 (2.84)
Qwen3-8B $\leftarrow$ Qwen3-32B (Think-in-Opposites)	$[1{\times}10^{-5},\,5{\times}10^{-4}]$ (6 values)	$2{\times}10^{-4}$	4400 (3.91)
Qwen3-32B $\leftarrow$ Qwen3-32B (Think-in-Opposites)	$[2{\times}10^{-6},\,5{\times}10^{-5}]$ (6 values)	$5{\times}10^{-5}$	2200 (1.96)

Appendix

Table 13: Learning-rate sweeps and selected checkpoints for distillation.

Objects ( $N$ )	Blickets ( $K$ )	Rule	# Episodes
4	2	AND	16
4	2	OR	16
4	2	XOR	16
4	3	AND	16
4	3	OR	16
4	3	XOR	16

Appendix

Table 14: Blicket evaluation dataset.

Thinking models				Non-thinking models
Metric	Intervention	Mean $\Delta$	$p$	Metric	Intervention	Mean $\Delta$	$p$
Task	Dual-Goal	+0.148	$<2\times 10^{-5}$	Task	Dual-Goal	$-0.060$	0.0065
Success	Think-in-Opposites	+0.123	$<2\times 10^{-5}$	Success	Think-in-Opposites	+0.063	0.0033
I:C	Dual-Goal	+0.375	$<2\times 10^{-5}$	I:C	Dual-Goal	+0.048	0.2386
	Think-in-Opposites	+0.168	$<2\times 10^{-5}$		Think-in-Opposites	+0.020	0.5995

Appendix

Table 15: Effect of interventions on the Wason Task.

\Delta

denotes the change relative to the baseline prompt. Statistical significance is assessed using paired permutation tests over episode pairs (

50{,}000

permutations).

	Non-thinking		Thinking
Metric	$\Delta$	$p$	$\Delta$	$p$
Task Success	+0.003	0.936	+0.092	$<2\times 10^{-5}$
I:C	+0.064	0.033	+0.231	$<2\times 10^{-5}$

Appendix

Table 16: Effect of Think-in-Opposites (TiO) on the Blicket test.

\Delta

denotes the change relative to the baseline prompt. Statistical significance is assessed using paired permutation tests over episode pairs (

50{,}000

permutations).

Metric	Model	Mean $\Delta$	$p$
Task Success	Qwen3-8B + 8B TiO	+0.138	0.053
	Qwen3-32B + 32B TiO	+0.225	0.00032
	Qwen3-8B + 32B	+0.225	0.00078
	Qwen3-8B + 32B TiO	+0.438	$<2\times 10^{-5}$
I:C	Qwen3-8B + 8B TiO	+0.163	0.0018
	Qwen3-32B + 32B TiO	+0.185	$<2\times 10^{-5}$

Appendix

Table 17: Effect of distillation on the Wason Task.

\Delta

denotes the change relative to the corresponding baseline model. Statistical significance is assessed using paired permutation tests over episode pairs (

50{,}000

permutations).

Metric	Model	Mean $\Delta$	$p$
Task Success	Qwen3-8B + 8B TiO	+0.021	0.674
	Qwen3-32B + 32B TiO	+0.203	$<2\times 10^{-5}$
	Qwen3-8B + 32B Baseline	+0.042	0.347
	Qwen3-8B + 32B TiO	+0.063	0.150
I:C	Qwen3-8B + 8B TiO	+0.039	0.396
	Qwen3-32B + 32B TiO	+0.345	$<2\times 10^{-5}$

Appendix

Table 18: Effect of Wason Distillation on the Blicket Test. Models are distilled on the Wason task and evaluated on the Blicket test.

\Delta

denotes the change relative to the corresponding baseline model on the Blicket test. Statistical significance is assessed using paired permutation tests over episode pairs (

50{,}000

permutations).

Setting	Sample size	Pos./Neg.	Agreement
Wason announcement	200	100 / 100	198 / 200 (99.0%)
Wason compatibility	200	100 / 100	197 / 200 (98.5%)
Blicket announcement	200	100 / 100	199 / 200 (99.5%)
Blicket compatibility	200	100 / 100	197 / 200 (98.5%)

Appendix

Table 19: Summary of the manual audit for judge validation.

Correct announcement	Model announcement	Judge	Human
All end with 1	All numbers end with the digit 1.	CORRECT	CORRECT
Mixed signs	The triple contains at least one negative number and at least one non-negative number.	CORRECT	CORRECT
Descending	The numbers are in decreasing order.	CORRECT	CORRECT
All cube numbers	The numbers are in a geometric progression with a common ratio of -3.	INCORRECT	INCORRECT
All positive	The three numbers form an arithmetic sequence.	INCORRECT	INCORRECT
Contains a prime	The sum of the first two numbers is twice the third.	INCORRECT	INCORRECT

Appendix

Table 20: Representative examples from the manual audit of Wason announcement judgments.

Previous announcement	Test triple	Judge	Human
All numbers are perfect cubes	[8, 27, 64]	TRUE	TRUE
All three numbers are prime	[5, 17, 19]	TRUE	TRUE
The numbers are in strictly decreasing order	[9, 6, 3]	TRUE	TRUE
All numbers are prime	[2, 4, 35]	FALSE	FALSE
Exactly one even and two odd numbers	[4, 6, 8]	FALSE	FALSE
The numbers must form a strictly decreasing sequence	[1, 2, 3]	FALSE	FALSE

Appendix

Table 21: Representative examples from the manual audit of Wason compatibility judgments.

True rule	True ids	Model announcement	Judge	Human
disjunctive	[0,1,2]	relevant=[object 0, object 1, object 2]; rule=contains at least one of object 0, 1, or 2	TRUE	TRUE
conjunctive	[0,3]	relevant=[object 0, object 3]; rule=objects 0 and 3 must be placed on the device together	TRUE	TRUE
xor	[0,7]	relevant=[object 0, object 7]; rule=the device turns on if exactly one of object 0 or object 7 is present	TRUE	TRUE
conjunctive	[0,3]	relevant=[object 0, object 1, object 2, object 3]; rule=the device turns on if all objects are placed on it	FALSE	FALSE
xor	[0,2,3]	relevant=[object 0, object 2, object 3]; rule=the device turns on if any one of them is present	FALSE	FALSE
disjunctive	[0,2]	relevant=[object 0, object 3]; rule=device turns on if object 0 or object 3 is present	FALSE	FALSE

Appendix

Table 22: Representative examples from the manual audit of Blicket announcement judgments.

Previous announcement	Tested object ids	Judge	Human
relevant=[object 0, object 2]; rule=must include both object 0 and object 2	[0, 2]	TRUE	TRUE
relevant=[object 1, object 7]; rule=exactly one of object 1 or object 7 is required to turn the device on, but not both	[5, 7]	TRUE	TRUE
relevant=[object 0]; rule=object 0 alone is sufficient to turn the device on	[0]	TRUE	TRUE
relevant=[object 0, object 1, object 3]; rule=The device turns on when at least one of object 0, object 1, or object 3 is placed on it.	[2]	FALSE	FALSE
relevant=[object 0, object 2]; rule=contains object 0 or object 2	[1, 3]	FALSE	FALSE
relevant=[object 0, object 1, object 2, object 3]; rule=all objects must be placed on the device	[0, 1, 3]	FALSE	FALSE

Appendix

Table 23: Representative examples from the manual audit of Blicket compatibility judgments.

Language Model Bias Evaluation

Language model bias evaluation tests how models’ outputs, decisions, and representations vary across social groups, cultures, languages, and contexts.

See all 492 papers

Confirming Our Biases? Evaluating the Capabilities, Risks, and Societal Impact of Large Language Models
Aug 7, 2026Mudar Adas, Polina Tsvilodub, Michael Franke +1Prompt Sensitivity LLM Sycophancy
Neuro-Cognitive Modeling Group, University of Tübingen, Sand 14, 72076 Tübingen, Germany · Department of Linguistics, University of Tübingen, Keplerstraße 2, 72074 Tübingen
CogBias: Measuring and Mitigating Cognitive Bias in Large Language Models
Apr 1, 2026Fan Huang, Songheng Zhang, Haewoon Kwak +1Cognitive Bias in Language Models Language Model Bias Evaluation
Indiana University Bloomington · Singapore Management University
FALSIFYBENCH: Evaluating Inductive Reasoning in LLMs with Rule Discovery Games
Jun 3, 2026Leonardo Bertolazzi, Katya Tentori, Raffaella BernardiLLM Evaluation AI Agent Benchmarks
University of Trento · Free University of Bozen-Bolzano

Aug 7, 2026J/K move · Enter open · S save

Confirming Our Biases? Evaluating the Capabilities, Risks, and Societal Impact of Large Language Models

Mudar Adas, Polina Tsvilodub, Michael Franke, Martin V. Butz

Neuro-Cognitive Modeling Group, University of Tübingen, Sand 14, 72076 Tübingen, Germany · Department of Linguistics, University of Tübingen, Keplerstraße 2, 72074 Tübingen

Prompt Sensitivity LLM Sycophancy LLM Prompting Language Model Bias Evaluation

Open paper AskPDF arXiv

It is well established that large language models (LLMs) are sensitive to prompt framing, reflecting patterns in their training data or prior prompts. In this study, we investigate the extent to which LLMs reinforce users biases expressed in the prompts and examine the boundary between implicit framing effects and explicit prompt manipulation. Specifically, we evaluate how susceptible LLMs are to direct and suggestive prompts that encourage models to support or challenge particular positions. We evaluate six LLMs using 160 distinct prompts spanning ten topics across opinion-based and factual domains. The prompts systematically vary in prompting strategy, support versus challenge instructions, prompt polarity, users' expressed beliefs, and topic domain, spanning both opinion-based and factual questions. Our results show that LLMs systematically adapt their responses to align with prompt framing, even in factual contexts. This suggests that prompt framing can outweigh factual consistency in model responses. Overall, our findings delineate the extent and boundaries of LLM manipulability. Furthermore, the results imply that LLMs can reinforce subtle user biases and are susceptible to explicit prompt manipulation even in domains where responses should remain factually stable.

Abstract

Language Model Bias Evaluation

Explore similar work

Confirming Our Biases? Evaluating the Capabilities, Risks, and Societal Impact of Large Language Models

CogBias: Measuring and Mitigating Cognitive Bias in Large Language Models

FALSIFYBENCH: Evaluating Inductive Reasoning in LLMs with Rule Discovery Games