Does Thinking Help Fairness? Reasoning Tokens Resolve Some Biases but Create More

Deng Pan; Joe Germino; Yihong Ma; Elizabeth Daly; Nuno Moniz; Ting Hua; Nitesh Chawla

cs.AISep 25, 2026

Does Thinking Help Fairness? Reasoning Tokens Resolve Some Biases but Create More

Authors: Deng Pan, Joe Germino, Yihong Ma, Elizabeth Daly, Nuno Moniz, Ting Hua, Nitesh Chawla

Organizations: University of Notre Dame · IBM Research, Dublin

Abstract

Thinking in reasoning language models (RLMs) has been subject to debate on whether it resolves or amplifies bias. Prior works have shown competing conclusions in both directions. Using a within-model thinking-vs.-non-thinking ablation across QwQ-32B, DeepSeek-R1-Distill-Qwen-32B, and Qwen3-32B on three high-stakes decision tasks (Adult, COMPAS, Credit), we show that thinking has an asymmetric dual effect on counterfactual fairness: it both resolves counterfactual flips produced by the non-thinking baseline and creates new flips at near-saturating model confidence. In all nine (model, dataset) combinations, the created flips outnumber the resolved flips by roughly 5 times. To explain the effect, we treat the thinking trace itself as a measurable site of fairness change and study it through two dynamic instruments: 1) We propose Counterfactual Depth Probability Gap (CDPG) to track bias evolution along thinking depth, and observe that bias propagates and amplifies with thinking. 2) We also formulate the Bias Transition Matrix (BTM) to show how predictions of counterfactual pairs change from non-thinking to thinking, and find that the asymmetric dual effect originates in the pair-state joint transition.

Figures & tables

Panel of Figure 1 — Figure 1: Anatomy of a cell- $c$ pair (bias created by thinking). Qwen3-32B on Adult. The two sides agree under the non-thinking arm but diverge under thinking ( $\mathrm{CDPG}_{K}{=}0.82$ ); highlighted spans are the bias-relevant phrasing.

		Think
		no flip	flip
No-think	no flip	$a$ : both agree	$c$ : created
No-think	flip	$b$ : resolved	$d$ : both flip

Table 1: Within-pair contingency cells. Each counterfactual pair is classified by whether a flip (

\hat{y}^{a}\neq\hat{y}^{a^{\prime}}

) occurs under the thinking and non-thinking arms.

			Contingency cells (counts)				McNemar	$\overline{\mathcal{D}_{\text{disp}}}$ in cell $b$		$\overline{\mathcal{D}_{\text{disp}}}$ in cell $c$
Model	Dataset	$N$	$a$	$b$	$c$	$d$	$p$	nothink	think K	nothink	think K
Qwen3-32B	Adult	5000	4531	75	362	32	$3.7{\times}10^{-46}$	0.219	0.048	0.055	0.859
	COMPAS	5000	4511	143	344	2	$3.7{\times}10^{-20}$	0.150	0.037	0.046	0.781
	Credit	4996	4867	23	105	1	$1.0{\times}10^{-13}$	0.075	0.012	0.023	0.755
QwQ-32B	Adult	5000	4396	152	443	9	$6.1{\times}10^{-34}$	0.131	0.001	0.029	0.980
	COMPAS	5000	4634	15	351	0	$2.3{\times}10^{-84}$	0.262	0.008	0.002	0.891

Table 2: Dual-effect contingency table with within-pair gap reshaping in cells

b

and

c

. Cell counts

a,b,c,d

; the McNemar column reports the exact two-sided

p

-value of McNemar’s test on the discordant cells

(b,c)

, rejecting

H_{0}{:}\,b{=}c

in every setting; rightmost columns give the mean per-pair counterfactual disparity

\overline{\mathcal{D}_{\text{disp}}}

in cells

b,c

under the nothink and final-thinking arms (

\mathcal{D}_{\text{disp}}^{(i)}\equiv|\mathrm{CDPG}_{K}^{(i)}|

, Eqs. 2 , 6 ).

N

retains pairs parsing on both sides of both arms.

Figure 2: Per-cell change under thinking of counterfactual disparity (

\Delta\mathcal{D}_{\mathrm{cf}}

, Eq. 1 ) vs. group disparity (

\Delta\mathcal{D}_{\mathrm{group}}

, Eq. 3 ). Absolute Think/No-think values are reported in Table 9 (Appendix F ).

Figure 3: The dual effect plays out consistently as the thinking trace deepens. Mean

|\mathrm{CDPG}_{k}|

as a function of normalised thinking depth

k/K

, computed within each contingency cell, pooled across the nine (model, dataset) settings.

Figure 4: Per-side

P(\mathrm{Yes})

evolution through the thinking trace. Cell

b

(left, resolved by thinking) and cell

c

(right, created by thinking). The leftmost point (

k{=}0

) is the nothink baseline. Pooled across the nine (model, dataset) settings.

Adult

COMPAS

Credit

A

P(Y|Y)

P(Y|N)

\Delta P(Y)

P(Y|Y)

P(Y|N)

\Delta P(Y)

P(Y|Y)

P(Y|N)

\Delta P(Y)

\alpha

0.889

0.076

+0.033

0.998

0.281

+0.174

0.908

0.012

-0.026

\alpha^{\prime}

0.903

0.077

+0.032

0.997

0.272

+0.182

0.912

0.011

-0.025

\boldsymbol{\Delta=T^{\alpha}-T^{\alpha^{\prime}}}

\mathbf{-0.014}

\mathbf{-0.001}

\mathbf{+0.001}

\mathbf{+0.001}

\mathbf{+0.009}

\mathbf{-0.008}

\mathbf{-0.004}

\mathbf{+0.002}

\mathbf{-0.001}

Table 3: Group-level transition matrix

T^{A}

for Qwen3-32B, stratified by protected attribute.

\alpha/\alpha^{\prime}={}

Male/Female (Adult, Credit), White/Non-White (COMPAS);

P(Y|Y)

and

P(Y|N)

denote the probability of transitioning to a Y prediction from a Y or N prediction respectively.

\Delta P(Y)=P(Y)_{\text{th}}-P(Y)_{\text{nt}}

measures the overall shift in the Yes-prediction rate (Y = earns over $50K for Adult, will recidivate for COMPAS, will default for Credit). Other models in Appendix G .

	Adult				COMPAS				Credit
from	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{M}$	$n$	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{M}$	$n$	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{M}$	$n$
$\mathrm{NN}$	$0.894$	$0.035$	$0.071$	$3{,}720$	$0.686$	$0.207$	$0.108$	$3{,}157$	$0.981$	$0.004$	$0.015$	$3{,}170$
$\mathrm{YY}$	$0.050$	$0.866$	$0.084$	$1{,}173$	$0.001$	$0.996$	$0.002$	$1{,}698$	$0.068$	$0.899$	$0.033$	$1{,}802$
$\mathrm{M}$	$0.290$	$0.411$	$0.299$	$107$	$0.000$	$0.986$	$0.014$	$145$	$0.917$	$0.042$	$0.042$	$24$
total	$3{,}415$	$1{,}191$	$\mathbf{394}$	$5{,}000$	$2{,}167$	$2{,}487$	$\mathbf{346}$	$5{,}000$	$3{,}256$	$1{,}634$	$\mathbf{106}$	$4{,}996$

Table 4: Pair-state transition matrix

T^{\mathrm{pair}}

for Qwen3-32B (forward: nothink

\to

think). Within-row conditional probabilities; row counts

n

and column totals shown. Highlighted

\mathrm{M}

column totals are the post-think cell-

c

d

population sizes.

Model	Dataset	$\|c\|$	$c_{\alpha\uparrow}$	$c_{\alpha^{\prime}\uparrow}$	$\text{asym}(c)$	$\delta_{c}$ (%)
Qwen3	Adult	362	186	176	0.03	$+0.20$
	COMPAS	344	170	174	0.01	$-0.08$
	Credit	105	61	44	0.16	$+0.34$
QwQ	Adult	443	231	212	0.04	$+0.38$
	COMPAS	351	82	269	0.53	$-3.74$
	Credit	595	232	363	0.22	$-2.63$

Table 5: Directional decomposition of created flips (cell

c

\delta_{c}

reported in %; definitions in § 4.4 . Shaded cells: the four (model, dataset) settings with

\Delta\mathcal{D}_{\text{group}}{>}0

in Table 9 .

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: CDPG measurement pipeline. Counterfactual paired inputs $(Q^{\alpha},Q^{\alpha^{\prime}})$ are passed through the same reasoning model; each thinking trace is split into $K{=}10$ equal-token segments. At every prefix depth $k$ we read the answer-token softmax $P^{A}_{k}{=}P(\hat{y}{=}1\mid Q^{A},S^{A}_{1:k})$ and report $\mathrm{CDPG}_{k}{=}P^{\alpha}_{k}{-}P^{\alpha^{\prime}}_{k}$ .

		Cell- $c$ propagation slope
Dataset	$n_{c}$	$K{=}5$	$K{=}10$	$K{=}20$
Adult	54	0.78	0.65	0.59
COMPAS	56	0.77	0.62	0.54
Credit	16	0.61	0.42	0.32

Appendix

Table 6: CDPG segmentation sensitivity for Qwen3-32B. Fitted linear slope of the cell-

c

mean

|\mathrm{CDPG}_{k}|

trajectory against normalised depth

k/K

, recomputed at

K\in\{5,10,20\}

;

n_{c}

is the number of cell-

c

pairs in the rerun sample.

Figure 6: CDPG trajectories are robust to the segmentation choice. Mean

|\mathrm{CDPG}_{k}|

per contingency cell as a function of normalised depth

k/K

for

K\in\{5,10,20\}

(Qwen3-32B, rerun subsample pooled across the three datasets; per-cell pair counts in the legend).

			Contingency cells (counts)					McNemar	$\overline{\mathcal{D}_{\text{disp}}}$ (think K )
Model (family)	Dataset	$N$	$a$	$b$	$c$	$d$	$c:b$	$p$	cell $c$	cell $b$
gemma-2-27B (Google)	Adult	997	928	11	54	4	$4.9\times$	$6.0{\times}10^{-8}$	0.886	0.006
	COMPAS †	704	702	0	2	0	—	$0.5$	0.446	—
	Credit	1000	942	0	58	0	—	$6.9{\times}10^{-18}$	0.940	—
Llama-3.1-8B (Meta)	Adult	889	756	20	109	4	$5.5\times$	$5.1{\times}10^{-16}$	0.684	0.095
	COMPAS	1000	842	18	140	0	$7.8\times$	$1.3{\times}10^{-24}$	0.293	0.106

Appendix

Table 7: The asymmetric dual effect replicates on two non-Qwen model families. Same protocol and metrics as Table 2 , at

N{=}1{,}000

sampled records per setting (

N

retains pairs parsing on both sides of both arms); the thinking trace is elicited by prompt since neither model exposes a native <think> channel.

\overline{\mathcal{D}_{\text{disp}}}

(think K ) is the mean per-pair counterfactual disparity (Eq. 2 ) at full depth within cells

c

and

b

. † gemma-2-27B declines the COMPAS recidivism question on a large share of records (296 of

1{,}000

pairs unparsed), leaving too few discordant pairs for a powered test; the row is shown for completeness and included in the pooled counts.

Condition	Example prompt (Adult)
Thinking mode	{description} Question: Does this person earn more than $50K per year? Answer with a single word: Yes or No.
Non-thinking mode	[Same user message as above. The chat template is rendered locally and the assistant turn is pre-filled with a closed empty <think>\n\n</think>\n\n block, so generation resumes after </think> and the model skips the reasoning phase.]

Appendix

Table 8: Prompt formats for the thinking and non-thinking arms on the Adult dataset. {description} is the natural-language profile from Appendix A .

			Think		No-think		$\boldsymbol{\Delta}$
Model	Dataset	$N$	$\mathcal{D}_{\text{cf}}$	$\mathcal{D}_{\text{group}}$	$\mathcal{D}_{\text{cf}}$	$\mathcal{D}_{\text{group}}$	$\Delta\mathcal{D}_{\text{cf}}$	$\Delta\mathcal{D}_{\text{group}}$
Qwen3	Adult	5000	0.079	0.006	0.021	0.012	$+0.058$	$-0.006$
	COMPAS	5000	0.069	0.001	0.029	0.010	$+0.040$	$-0.009$
	Credit	4996	0.021	0.004	0.005	0.004	$+0.016$	$\phantom{+}0.000$
QwQ	Adult	5000	0.090	0.004	0.032	0.021	$+0.058$	$-0.017$
	COMPAS	5000	0.070	0.037	0.003	0.003	$+0.067$	$+0.034$

Appendix

Table 9: Counterfactual disparity

\mathcal{D}_{\text{cf}}

(Eq. 1 ) and demographic-parity gap

\mathcal{D}_{\text{group}}

(Eq. 3 ) under the Think and No-think arms, with within-cell change

\Delta=\mathcal{D}^{\text{Think}}-\mathcal{D}^{\text{No-think}}

. Positive

\Delta

amplifies the bias, negative

\Delta

mitigates (palette from Table 1 ).

N

matches Table 2 .

Adult

COMPAS

Credit

A

P(Y|Y)

P(Y|N)

\Delta P(Y)

P(Y|Y)

P(Y|N)

\Delta P(Y)

P(Y|Y)

P(Y|N)

\Delta P(Y)

\alpha

0.980

0.300

+0.249

1.000

0.525

+0.525

0.977

0.387

+0.361

\alpha^{\prime}

0.988

0.302

+0.255

1.000

0.515

+0.513

0.961

0.381

+0.355

\boldsymbol{\Delta=T^{\alpha}-T^{\alpha^{\prime}}}

\mathbf{-0.008}

\mathbf{-0.002}

\mathbf{-0.006}

\mathbf{+0.000}

\mathbf{+0.010}

\mathbf{+0.011}

\mathbf{+0.017}

\mathbf{+0.006}

\mathbf{+0.006}

Appendix

Table 10: Group-level transition matrix

T^{A}

for QwQ-32B, stratified by protected attribute. Same conventions as Table 3 .

Adult

COMPAS

Credit

A

P(Y|Y)

P(Y|N)

\Delta P(Y)

P(Y|Y)

P(Y|N)

\Delta P(Y)

P(Y|Y)

P(Y|N)

\Delta P(Y)

\alpha

0.968

0.196

+0.138

1.000

0.215

+0.203

1.000

0.382

+0.381

\alpha^{\prime}

0.977

0.201

+0.147

1.000

0.203

+0.193

1.000

0.392

+0.391

\boldsymbol{\Delta=T^{\alpha}-T^{\alpha^{\prime}}}

\mathbf{-0.009}

\mathbf{-0.005}

\mathbf{-0.009}

\mathbf{+0.000}

\mathbf{+0.012}

\mathbf{+0.010}

\mathbf{+0.000}

\mathbf{-0.010}

\mathbf{-0.010}

Appendix

Table 11: Group-level transition matrix

T^{A}

for DeepSeek-R1-Distill-Qwen-32B, stratified by protected attribute. Same conventions as Table 3 .

		$\|c\|$ (created)			$\|b\|$ (resolved)			$\|c\|/\|b\|$
Model	Dataset	ind.	actual	ratio	ind.	actual	ratio	ind.	actual
Qwen3-32B	Adult	$744$	$363$	$2.05\times$	$18$	$75$	$0.24\times$	$41$	$4.8$
Qwen3-32B	COMPAS	$1272$	$344$	$3.69\times$	$40$	$143$	$0.28\times$	$32$	$2.4$
Qwen3-32B	Credit	$367$	$107$	$3.43\times$	$2$	$23$	$0.10\times$	$153$	$4.7$
QwQ-32B	Adult	$1768$	$444$	$3.98\times$	$50$	$152$	$0.33\times$	$36$	$2.9$
QwQ-32B	COMPAS	$2487$	$349$	$7.13\times$	$8$	$15$	$0.52\times$	$319$	$23.3$

Appendix

Table 12: Independence-bound check per (model, dataset) cell. Independence predictions use

\mathbb{E}|c|=2p(1{-}p)\,|\mathrm{NN}|_{\mathrm{nt}}+2q(1{-}q)\,|\mathrm{YY}|_{\mathrm{nt}}

and

\mathbb{E}|b|=[p(1{-}q){+}q(1{-}p)]\,|\mathrm{M}|_{\mathrm{nt}}

with

p=P(Y|N)

and

q=P(N|Y)

read from Tables 3 (Qwen3-32B), 10 (QwQ-32B), 11 (R1-Distill). Actual counts are read from Tables 4 , 13 , 14 . Across all nine cells, independence overpredicts

|c|

by 2–7

\times

and underpredicts

|b|

by 2–10

\times

, but the directional inequality

|c|>|b|

holds under both independence and the data. The empirical ratios

|c|/|b|

are uniformly smaller than the independence prediction (e.g., 4.8 vs. 41 on Qwen3-Adult): within-pair flip correlation dampens the base-count asymmetry but cannot reverse it.

	Adult					COMPAS					Credit
from	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{YN}$	$\mathrm{NY}$	$n$	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{YN}$	$\mathrm{NY}$	$n$	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{YN}$	$\mathrm{NY}$	$n$
$\mathrm{NN}$	$0.659$	$0.236$	$\mathbf{0.055}$	$\mathbf{0.050}$	$4{,}149$	$0.445$	$0.484$	$\mathbf{0.016}$	$\mathbf{0.054}$	$4{,}981$	$0.561$	$0.312$	$\mathbf{0.050}$	$\mathbf{0.077}$	$4{,}616$
$\mathrm{YY}$	$0.004$	$0.984$	$\mathbf{0.006}$	$\mathbf{0.006}$	$690$	$0.000$	$1.000$	$\mathbf{0.000}$	$\mathbf{0.000}$	$4$	$0.008$	$0.943$	$\mathbf{0.012}$	$\mathbf{0.037}$	$244$
$\mathrm{YN}$	$0.045$	$0.888$	$0.030$	$0.037$	$134$	$0.000$	$1.000$	$0.000$	$0.000$	$14$	$0.009$	$0.965$	$0.009$	$0.018$	$114$
$\mathrm{NY}$	$0.000$	$1.000$	$0.000$	$0.000$	$27$	$0.000$	$1.000$	$0.000$	$0.000$	$1$	$0.000$	$1.000$	$0.000$	$0.000$	$10$
total	$2{,}743$	$1{,}805$	$235$	$217$	$5{,}000$	$2{,}218$	$2{,}431$	$82$	$269$	$5{,}000$	$2{,}594$	$1{,}792$	$233$	$365$	$4{,}984$

Appendix

Table 13: Pair-state transition matrix

T^{\mathrm{pair}}

for QwQ-32B, per dataset (forward: nothink

\to

think), with

\mathrm{M}

split into

\mathrm{YN}

and

\mathrm{NY}

. Compared with Qwen3-32B (Table 4 ), QwQ runs hotter on every channel: the en-bloc

\mathrm{NN}\!\to\!\mathrm{YY}

push peaks at

48.4\%

on COMPAS (half the originally-agree-No pairs are pushed wholesale to agree-Yes), and within

\mathrm{M}

the cell-

c

split tilts strongly toward Non-White on COMPAS (

\mathrm{NN}\!\to\!\mathrm{NY}=0.054

\mathrm{NN}\!\to\!\mathrm{YN}=0.016

) and toward Female on Credit (

\mathrm{NN}\!\to\!\mathrm{NY}=0.077

\mathrm{NN}\!\to\!\mathrm{YN}=0.050

); the YY side reinforces the Credit tilt (

\mathrm{YY}\!\to\!\mathrm{NY}=0.037

\mathrm{YY}\!\to\!\mathrm{YN}=0.012

	Adult					COMPAS					Credit
from	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{YN}$	$\mathrm{NY}$	$n$	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{YN}$	$\mathrm{NY}$	$n$	$\mathrm{NN}$	$\mathrm{YY}$	$\mathrm{YN}$	$\mathrm{NY}$	$n$
$\mathrm{NN}$	$0.728$	$0.108$	$\mathbf{0.087}$	$\mathbf{0.078}$	$3{,}715$	$0.726$	$0.142$	$\mathbf{0.046}$	$\mathbf{0.086}$	$1{,}356$	$0.565$	$0.338$	$\mathbf{0.056}$	$\mathbf{0.042}$	$4{,}971$
$\mathrm{YY}$	$0.007$	$0.964$	$\mathbf{0.012}$	$\mathbf{0.018}$	$1{,}186$	$0.000$	$1.000$	$\mathbf{0.000}$	$\mathbf{0.000}$	$70$	$0.000$	$1.000$	$\mathbf{0.000}$	$\mathbf{0.000}$	$14$
$\mathrm{YN}$	$0.059$	$0.735$	$0.103$	$0.103$	$68$	–	–	–	–	$0$	$0.000$	$1.000$	$0.000$	$0.000$	$5$
$\mathrm{NY}$	$0.100$	$0.667$	$0.133$	$0.100$	$30$	$0.000$	$1.000$	$0.000$	$0.000$	$4$	$0.000$	$1.000$	$0.000$	$0.000$	$5$
total	$2{,}718$	$1{,}614$	$348$	$319$	$4{,}999$	$985$	$266$	$63$	$116$	$1{,}430$	$2{,}808$	$1{,}704$	$276$	$207$	$4{,}995$

Appendix

Table 14: Pair-state transition matrix

T^{\mathrm{pair}}

for DeepSeek-R1-Distill-Qwen-32B, per dataset (forward: nothink

\to

think), with

\mathrm{M}

split into

\mathrm{YN}

and

\mathrm{NY}

. R1-Distill has the highest aggregated

\mathrm{NN}\!\to\!\mathrm{M}

rates of the three models (Adult:

\mathrm{NN}\!\to\!\mathrm{YN}+\mathrm{NN}\!\to\!\mathrm{NY}=0.165

, the per-model peak); the SFT-distillation regime is the most aggressive at producing pair-level splits rather than en-bloc pushes. The COMPAS row reflects

\sim 72\%

refusal under R1-Distill’s prompt sensitivity (only

1{,}430

parseable pairs); on the parseable subset

\mathrm{NN}

tilts strongly toward Non-White (

\mathrm{NN}\!\to\!\mathrm{NY}=0.086

\mathrm{NN}\!\to\!\mathrm{YN}=0.046

), matching QwQ on COMPAS. On Credit, the direction reverses (

\mathrm{NN}\!\to\!\mathrm{YN}=0.056

\mathrm{NN}\!\to\!\mathrm{NY}=0.042

, Male-leaning), opposite to QwQ.

Social Bias in Language Models

Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.

See all 199 papers

The Long Road to the Same Answer: Cognitive Bias Under Escalating Reasoning Budgets in Large Language Models
Oct 7, 2026Obada KraishanCognitive Bias in Language Models Large Reasoning Models
College of Media and Communication Texas Tech University Lubbock, TX, USA
COFT: Counterfactual-Conformal Decoding for Fair Chain-of-Thought Reasoning in Large Language Models
May 28, 2026Arya Fayyazi, Mehdi Kamal, Massoud PedramSocial Bias in Language Models Counterfactual Fairness
Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, California, USA.
Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models
Jun 1, 2026Simone Caldarella, Davide Talon, Rahaf Aljundi +2LLM Evaluation Overthinking in Language Models
University of Trento · 3Fondazione Bruno Kessler · 2Toyota Motor Europe

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

The Long Road to the Same Answer: Cognitive Bias Under Escalating Reasoning Budgets in Large Language Models

Obada Kraishan

College of Media and Communication Texas Tech University Lubbock, TX, USA

Cognitive Bias in Language Models Large Reasoning Models Language Model Bias Evaluation Inference-Time Scaling

Open paper AskPDF arXiv

Reasoning models allocate extra computation at inference time and present their answers as the product of deliberate thought. If this deliberation works the way dual-process accounts of human cognition suggest, longer thinking should weaken the classic decision biases that fast, intuitive judgment produces. Using 30 vignettes covering six biases (anchoring, framing, loss aversion, escalation of commitment, availability, confirmation) from an established benchmark, we run a dose-response study across four model families, pairing each reasoning model with a matched non-reasoning sibling and requesting thinking ceilings of 0, 1,024, 4,096, and 8,192 tokens, for 12,350 API calls. Because a requested ceiling is not the same as realized deliberation, we use the reasoning tokens each call consumed as the dose. First, reasoning models are not less biased than their siblings; the point estimate leans the other way in every family, but the item-level pooled contrast is not reliable (Delta = +0.031, t(29) = 1.45, p = .157). Second, bias magnitude does not reliably fall as realized deliberation grows: no slope is significantly negative, and where anything moves it is the signed score drifting further from the human direction. Third, anchoring is the only bias in the human direction (d = 1.89). Four of the other five lean the opposite way in all seven models; with five items per bias, that reversal is reliable for framing and directional for escalation of commitment, confirmation, and loss aversion, while availability is absent. A one-line instruction to restate the anchor before answering lowered anchoring on all five anchoring items, which no amount of additional thinking did, although the effect does not reach significance (p = .057). The results argue against treating test-time reasoning as a rationality guarantee and for auditing deployed models bias by bias.

Abstract

Social Bias in Language Models

Explore similar work

The Long Road to the Same Answer: Cognitive Bias Under Escalating Reasoning Budgets in Large Language Models

COFT: Counterfactual-Conformal Decoding for Fair Chain-of-Thought Reasoning in Large Language Models

Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models