Do LLMs Really Lack Difference Awareness? Uncovering Heuristic Substitution in Group-Differentiating Fairness

Jingshen Zhang; Bo Wang; Yanlin Fu; Yuewen Dang; Dongming Zhao; Zifei Yu; Yuexian Hou

cs.SIOct 18, 2026

Do LLMs Really Lack Difference Awareness? Uncovering Heuristic Substitution in Group-Differentiating Fairness

Authors: Jingshen Zhang, Bo Wang, Yanlin Fu, Yuewen Dang, Dongming Zhao, Zifei Yu, Yuexian Hou

Organizations: School of Computer Science and Technology, Tianjin University · QMindTech · Department of Nuclear Science and Technology, Fudan University · Tianjin Huizhixingyuan Information Technology Co., Ltd

Abstract

In this paper, we investigate the issue of difference awareness in LLMs, defined as a model's ability to recognize and acknowledge meaningful differences between social groups in fairness decision-making. Previous studies have consistently shown that models perform poorly on such tasks, a finding commonly attributed to a lack of difference awareness. We propose an alternative account, which we term Fairness Heuristic Substitution: difference awareness is not necessarily absent, but the complex demand of simultaneously identifying relevant group differences and determining their implications for fairness decisions may lead models to substitute the simpler, more accessible equal-treatment heuristic, i.e., formal equality. Based on this hypothesis, we introduce TDD, a Two-stage Distinguish-Decide framework that explicitly decomposes fairness reasoning into first distinguishing contextually relevant group differences and then making the corresponding decision, thereby discouraging reliance on formal-equality shortcuts. (i) Across four benchmarks and ten models ranging from billions to trillions of parameters, TDD significantly outperforms one-stage baselines, even one strengthened through simulated test-set leakage. (ii) Combining TDD with neuron-level contrastive probing provides mechanistic evidence consistent with the Fairness Heuristic Substitution hypothesis by revealing two signatures of the formal-equality: a compact, locally coordinated representational structure and a pronounced activation bias under default reasoning. (iii) TDD's stage-wise decomposition uncovers overlooked failure modes in fairness reasoning, including heuristic displacement and misaligned stage dynamics.

Figures & tables

Figure 1: TDD , a T wo-stage D istinguish- D ecide framework (Right), separates contextual distinction from fairness decision-making, unlike methods that conflate the two (Left). Deliberative decomposition mitigates fairness heuristic substitution ( § 5.1 ) and reveals neuron-level evidence that formal equality acts as the substitute heuristic ( § 5.2 ).

Principle Condition $y^{*}$	Group 1	Group 2	G1=G2
DIFF ( $\neq$ )	A	B	C
EQUAL ( $=$ )	D	E	F

Table 1: Possible model outputs under the two principles. DIFF (

\neq

) warrants group-sensitive treatment, whereas EQUAL (

=

) requires equal treatment regardless of group membership. Shaded cells indicate correct answers.

A+B+C=D+E+F=1000

Figure 2: DiffAware results across ten models and four datasets. L3/L3.1, G2, and Q2.5/Q3 indicate models from the Llama, Gemma, and Qwen families, respectively. Higher DiffAware scores indicate better difference awareness, i.e., the ability to identify situations where differential treatment is normatively justified. TDD leads in 28 of 40 (70%) model–benchmark pairs, even against a strong Few-shot baseline with simulated test-set leakage. More detailed results, including CtxtAware , and a discussion of TDD are provided in Appendix D .

Figure 3: Causal validation via neuron ablation. The y-axes report DiffAware and EqualAware for DIFF and EQUAL scenarios, respectively; higher scores indicate better fairness performance. For each model, targeted ablations of $\mathcal{N}^{\mathrm{DIFF}}$ and $\mathcal{N}^{\mathrm{EQUAL}}$ are compared with the unablated Oracle and Counterfactual conditions and a random-ablation baseline $\mathcal{N}^{\mathrm{RAND}}$ . Hatched regions mark performance drops relative to the Oracle under targeted ablation. We set $K=100$ here; robustness across different $K$ values is evaluated in Appendix B .

Model	Rep. Size (D/E)	Stable Rank (D/E)	Co-act. Rate (D/E)
L3-8B	44,212 / 40,101 $\downarrow$ 9.30%	15.45 / 10.48 $\downarrow$ 32.17%	4.09% / 4.74% $\uparrow$ 16.10%
L3.1-8B	44,503 / 35,960 $\downarrow$ 19.20%	13.20 / 9.29 $\downarrow$ 29.62%	4.36% / 5.21% $\uparrow$ 19.50%
G2-9B	49,601 / 45,300 $\downarrow$ 8.67%	10.49 / 9.00 $\downarrow$ 14.20%	5.06% / 5.40% $\uparrow$ 6.76%
Q2.5-7B	33,950 / 31,286 $\downarrow$ 7.85%	16.66 / 15.48 $\downarrow$ 7.08%	3.11% / 3.36% $\uparrow$ 8.10%
Q2.5-14B	59,128 / 49,117 $\downarrow$ 16.93%	17.55 / 16.94 $\downarrow$ 3.48%	4.37% / 4.95% $\uparrow$ 13.17%
Q3-4B	40,392 / 36,359 $\downarrow$ 9.98%	16.28 / 14.85 $\downarrow$ 8.78%	6.69% / 7.23% $\uparrow$ 8.10%

Table 2: Comparison of the neural representation structures associated with difference awareness (D) and formal equality (E). Lower Rep. Size and Stable Rank indicate more compact representations, whereas a higher Co-act. Rate indicates stronger neuronal coordination.

Model	N1	N2	N3	N4	Dist. Imb. $[0,50]\downarrow$
L3-8B	100.00	98.90	100.00	98.90	49.45
L3.1-8B	99.40	99.55	100.00	81.10	45.01
G2-9B	90.65	75.00	100.00	68.25	33.48
Q2.5-7B	97.65	98.50	97.60	70.25	41.00
Q2.5-14B	98.90	99.70	100.00	86.30	46.23
Q3-4B	99.65	98.65	100.00	77.45	43.94

Table 3: Stage-1 DIFF prediction rates (%) across benchmarks. The balanced reference is 50%; Dist. Imb. denotes the mean absolute deviation from 50, with higher values indicating greater imbalance.

Figure 5: Interventional results obtained by manipulating Stage 1 distinction labels in the TDD framework. Original refers to the unintervened TDD using model-predicted distinction labels. Oracle injects ground-truth distinction labels, whereas Counterfactual injects the corresponding opposite labels. Results are reported in terms of DiffAware and CtxtAware across four datasets.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Causal effects of neuron deactivation in DIFF scenarios. DiffAware is reported as a function of the number of deactivated neurons

K

across seven models. Deact-DIFF deactivates the identified difference-aware neurons, whereas Deact-RAND deactivates a size-matched random neuron set. Oracle and Counterfactual represent no-ablation reference results obtained using the correct and opposite distinction labels, respectively.

Figure 7: Causal effects of neuron deactivation in EQUAL scenarios. EqualAware is reported as a function of the number of deactivated neurons

K

across seven models. Deact-EQUAL deactivates the identified formal-equality neurons, whereas Deact-RAND deactivates a size-matched random neuron set. Oracle and Counterfactual represent the no-ablation reference results obtained using the correct and opposite distinction labels, respectively.

Bench	Method	L3-8B	L3.1-8B	G2-9B	Q2.5-7B	Q2.5-14B	Q3-4B	Q3-8B	GLM-5.1	Kimi-K2.6	GPT-5.6	Avg. Rank
N1: BBQ	Base	15.70 (4)	45.50 (3)	43.44 (4)	15.40 (5)	8.10 (4)	52.20 (2)	52.00 (4)	38.20 (5)	42.20 (3)	17.20 (4)	3.80
	SD	9.20 (5)	39.33 (5)	20.40 (5)	17.00 (4)	4.07 (5)	21.33 (5)	26.10 (5)	41.30 (4)	20.57 (5)	0.83 (5)	4.80
	SD1	18.30	70.80	50.46	37.10	7.60	53.10	52.00	57.40	50.70	2.50	–
	SD2	0.10	2.00	7.32	11.20	0.30	8.20	9.40	39.40	6.50	0.00	–
	SD3	9.20	45.20	3.41	2.70	4.30	2.70	16.90	27.10	4.50	0.00	–
	PD-Only	18.90 (3)	45.40 (4)	54.11 (2)	28.10 (3)	14.20 (3)	43.50 (4)	53.60 (2)	53.30 (3)	39.18 (4)	30.00 (3)	3.10

Appendix

Table 4: Detailed DiffAware results (%) across ten models, five prompting methods, and four benchmarks. Scores for Base, SD, PD-Only, Few-Shot, and TDD are followed by their ranks, with higher scores ranked better; Avg. Rank averages these ranks across the ten models. SD reports the mean of the three Self-Debiasing variants, shown separately in gray as SD1–SD3. Bold values indicate rank 1 among the five methods; SD1–SD3 and label-intervention results are excluded from ranking.

Bench	Method	L3-8B	L3.1-8B	G2-9B	Q2.5-7B	Q2.5-14B	Q3-4B	Q3-8B	GLM-5.1	Kimi-K2.6	GPT-5.6	Avg. Rank
N1: BBQ	Base	7.85 (5)	22.75 (4)	21.20 (4)	8.20 (5)	28.95 (4)	26.10 (3)	26.00 (4)	64.00 (5)	30.35 (4)	57.05 (4)	4.20
	SD	15.52 (3)	25.43 (3)	11.92 (5)	12.32 (4)	32.53 (3)	13.12 (5)	19.22 (5)	66.23 (4)	49.40 (2)	49.23 (5)	3.90
	SD1	4.80	22.60	3.45	1.85	27.25	5.50	20.35	60.90	48.55	48.70	–
	SD2	32.60	18.25	7.30	16.50	42.50	7.30	11.30	68.30	48.25	48.90	–
	SD3	9.15	35.45	25.00	18.60	27.85	26.55	26.00	69.50	51.40	50.10	–
	PD-Only	9.45 (4)	22.70 (5)	25.00 (3)	14.80 (3)	26.45 (5)	21.75 (4)	26.80 (3)	71.45 (3)	29.10 (5)	62.60 (3)	3.80

Appendix

Table 5: Overall Accuracy results (%). Bold and underlined values indicate the best and second-best results, respectively.

Bench	Method	L3-8B	L3.1-8B	G2-9B	Q2.5-7B	Q2.5-14B	Q3-4B	Q3-8B	GLM-5.1	Kimi-K2.6	GPT-5.6	Avg. Rank
N1: BBQ	Base	13.57 (4)	31.27 (3)	43.35 (2)	13.47 (5)	13.89 (4)	34.30 (3)	34.21 (4)	79.09 (5)	34.11 (4)	93.99 (2)	3.60
	SD	8.07 (5)	25.19 (5)	22.39 (5)	14.10 (4)	7.55 (5)	15.20 (5)	20.71 (5)	84.53 (2)	42.88 (3)	30.86 (5)	4.40
	SD1	8.46	31.13	6.18	2.65	7.95	2.86	18.15	84.16	37.82	0.00	–
	SD2	0.29	2.96	13.39	12.54	1.92	8.06	9.77	93.59	39.39	0.00	–
	SD3	15.47	41.48	47.61	27.10	12.77	34.68	34.21	75.83	51.42	92.59	–
	PD-Only	15.90 (3)	31.22 (4)	51.92 (1)	22.20 (3)	18.81 (3)	30.31 (4)	34.90 (3)	83.81 (3)	32.58 (5)	91.46 (3)	3.20

Appendix

Table 6: CtxtAware results (%). Bold and underlined values indicate the best and second-best results, respectively.

Algorithmic Fairness

Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.

See all 424 papers

DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training
Apr 18, 2026Ziwen Pan, Zihan Liang, Jad Kabbara +1LLM Auditing Social Bias in Language Models
Emory University · MIT
Explanation Fairness in Large Language Models: An Empirical Analysis of Disparities in How LLMs Justify Decisions Across Demographic Groups
May 9, 2026Gautam VeldandaLLM Auditing Algorithmic Fairness
Independent Researcher
Fair Like Us? Auditing LLM Alignment in Resource Allocation
Aug 31, 2026Qishen Han, Hadi Hosseini, Joshua Kavner +3LLM Alignment LLM Auditing
Rutgers University · Pennsylvania State University · Independent Researcher +1

Apr 18, 2026J/K move · Enter open · S save

DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training

Ziwen Pan, Zihan Liang, Jad Kabbara, Ali Emami

Emory University · MIT

LLM Auditing Social Bias in Language Models LLM Safety Alignment LLM Safety AI Ethics

Open paper AskPDF arXiv

Large language models (LLMs) tuned for safety often avoid acknowledging demographic differences, even when such acknowledgment is factually correct (e.g., ancestry-based disease incidence) or contextually justified (e.g., religious hiring preferences). This identity-blindness yields incorrect responses, unnecessary refusals, or generic "equal-treatment" defaults. We study this via difference-awareness classification: given a question involving demographic groups, the task is not to answer directly, but to classify whether a correct answer requires recognizing group differences (yes) or whether groups should be treated identically (no). Crucially, fine-tuning for accuracy triggers harm drift: model-generated explanations become increasingly harmful as decision accuracy improves, whether by elaborating harmful content, introducing problematic assumptions, or failing to flag harms the baseline identified. To mitigate this, we introduce DART (Distill--Audit--Repair Training), which distills label-conditioned reasoning from a teacher, audits outputs for harm drift cases relative to baseline, and repairs problematic cases via severity-weighted fine-tuning. On eight benchmarks, DART improves Llama-3-8B-Instruct accuracy from 39.0% to 68.8%, with largest gains on equal-treatment prompts (11.3% -> 72.6%), while reducing harm drift cases by 72.6%. It also transfers to 280 open-ended real-world queries across medical, legal, policy, and educational domains, improving difference-appropriate responses from 39.8% to 77.5% while reducing refusals from 34.3% to 3.0%. Our results demonstrate that accuracy and safety need not conflict when explicit detection and repair mechanisms are in place.

Abstract

Algorithmic Fairness

Explore similar work

DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training

Explanation Fairness in Large Language Models: An Empirical Analysis of Disparities in How LLMs Justify Decisions Across Demographic Groups

Fair Like Us? Auditing LLM Alignment in Resource Allocation