Knowledge Preference Bias

• Knowledge Preference Bias is a phenomenon where systems prefer one knowledge source or representation over others, leading to systematic distortions in decision-making.
• It manifests in diverse scenarios including source-selection failures, confidence miscalibration, and optimization-induced preferences in models, recommender systems, and expert evaluations.
• Practical implications include amplified bias in recommendations, suboptimal decisions in tutoring and policymaking, and the need for specialized mitigation strategies.

As an umbrella expression, knowledge preference bias can denote a family of failure modes in which a model, expert, or policymaker systematically privileges one knowledge source, one preference representation, or one interpretive schema over alternatives under uncertainty, conflict, or incomplete support. In the cited literature, the phenomenon appears in several technically distinct forms: source-selection failures between instruction, context, and parametric memory; confidence-miscalibrated preference for easy but unhelpful knowledge over hard but useful knowledge; amplification of behaviorally observed preferences in recommendation; expert overweighting of familiar conceptual models; modality-dominant shortcuts in multimodal systems; and evaluator favoritism toward outputs produced by particular model families (Zhou et al., 2024, Zhao et al., 14 Nov 2025, Kang et al., 17 Oct 2025, Whitehead et al., 2022, Ye et al., 22 May 2025, Koyama et al., 13 Apr 2026). This suggests that the unifying object is not merely error, but a structured preference over knowledge sources or representations that can distort downstream decisions.

1. Conceptual scope and distinctions

The literature does not treat knowledge preference bias as a single standardized construct. Some papers study it directly as a preference-ordering problem between knowledge sources, whereas others approach closely related phenomena under different names. Heckerman and Jimison, for example, are explicit that their framework is not about “knowledge preference bias” in the modern fairness or cognitive-bias sense; instead, it addresses beliefs and preferences that are only partially supported and should therefore not be forced into brittle point estimates (Heckerman et al., 2013). By contrast, work on emotional support conversation treats strategy preference bias as rooted in the model’s internal knowledge distribution and “knowledge boundaries,” rather than as a demographic or social bias (Zhou et al., 16 Sep 2025). GraphDPO makes a different distinction again: its “bias” is not fairness-oriented, but an optimization-induced preference over knowledge content, in which forgetting triples become dispreferred and alternative completions become preferred (Liu et al., 28 Jul 2025).

A common misconception is therefore that knowledge preference bias is always equivalent to fairness bias, stereotype bias, or simple class imbalance. The cited work shows a broader landscape. In some settings the problem is epistemic: the model prefers parametric memory over retrieved evidence, or a coarse utility assessment over a decomposed and better-supported one. In others it is optimization-induced: the training objective itself creates a ranking over facts, strategies, or candidate outputs. In still others it is institutional: ignorance becomes a justification for intervention, but that justification is filtered through the intervening party’s own preferences. The term is best understood as a cross-domain analytic label for systematic distortions in how knowledge, evidence, or preferences are ranked and used.

2. Meta-level uncertainty and hierarchical source selection

A foundational treatment of weakly supported beliefs and preferences appears in Heckerman and Jimison’s decision-analytic account of partial confidence (Heckerman et al., 2013). Their central claim is that if an assessor is “unable or unwilling to provide a precise value” for a probability or utility, this should be interpreted not as a failure of decision theory but as evidence that the quantity depends on unresolved factors. For beliefs, the assessor identifies influencing events and conditions on them, yielding decompositions such as

$$p(X)=\sum p(X\mid \text{Events})\,p(\text{Events}).$$

If confidence is still lacking, the same procedure is applied recursively. The same logic is stated to apply to preferences held with partial confidence. The key meta-level object is a random variable $\phi(X)$, defined as the value the probability or utility for $X$ would take after a specified assessment scenario is completed. A distribution over $\phi(X)$ represents confidence: broad distributions correspond to low confidence, sharply peaked distributions to high confidence. Refinement is then treated as a meta-decision problem, and the recommended rule is to refine when the value of perfect information on the refinement variables exceeds the cost of refinement. This reframes under-supported utilities as candidates for decomposition rather than authoritative fixed numbers.

A closely related but more explicitly source-selection-oriented formulation appears in work on LLMs. “Establishing Knowledge Preference in LLMs” defines a three-level hierarchy: Instruction Knowledge $\succ$ Context Knowledge $\succ$ Parametric Knowledge (Zhou et al., 2024). The paper’s premise is that many failures in retrieval-augmented QA, recommendation, counterfactual QA, and model editing reduce to violations of this hierarchy. A model is biased in this sense if it prefers parametric memory over retrieved evidence, or retrieved passages over user-provided assumptions, when conflict should be resolved differently.

KnowPO, which studies controllable knowledge selection in retrieval-augmented LLMs, sharpens this view by formalizing two dual failure modes: Contextual Ignorance, in which the model ignores useful retrieved context and answers from parametric knowledge, and Contextual Overinclusion, in which it overuses irrelevant or noisy context (Zhang et al., 2024). Its target behavior is explicitly conditional: if retrieved context contains an answer, the model should follow retrieval; if no retrieved document can answer the question, it should ignore the context and output its parametric answer. This turns knowledge preference bias into a conditional source-selection problem rather than a blanket rule such as “always trust retrieval.”

Taken together, these works define a common formal lens. Partial confidence in a belief or utility is itself a structured input to higher-level decision making, and conflict among instruction, context, and memory is a structured source-selection problem. In both cases, bias arises when under-modeled uncertainty or the wrong source hierarchy is collapsed into a fixed decision rule.

3. Confidence, calibration, and generation-time preference reversal

In multimodal generation, knowledge preference bias often appears as a misalignment between knowledge correctness and confidence. HinD with Knowledge Encouragement Preference Optimization (KEPO) studies this in knowledge-based visual question answering (Zhao et al., 14 Nov 2025). The paper’s central diagnosis is that MLLMs may be under-confident on helpful but weakly visually anchored “visual-agnostic” facts, while being over-confident on visually concentrated but answer-irrelevant content. Confidence for a generated knowledge snippet $K$ is defined as

$$C(K)=\exp\left(\frac{\sum_1^T \log p_{\pi_\phi}(K_t\mid K_{<t})}{T}\right),$$

and KEPO constructs DPO-style preference pairs by choosing, among helpful facts, the least confident one as the winner and, among unhelpful facts, the most confident one as the loser. On OK-VQA, full HinD-Know with HDFT+KEPO reaches PRR@K $94.3$ and score $68.3$; removing KEPO drops performance to PRR@K $\phi(X)$0 and score $\phi(X)$1; using random hit/miss pairing without confidence gives PRR@K $\phi(X)$2 and score $\phi(X)$3. The confidence-aware preference ordering therefore matters, not just the positive/negative label. The same paper also shows that a zero-shot Qwen2.5-VL-7B model can be over-confident on knowledge with Hits$\phi(X)$4, whereas after KEPO the confidence distribution for high-hit knowledge shifts rightward.

A related calibration-centric account appears in emotional support conversation, where strategy preference bias is tied to a model’s epistemic regime rather than only to class frequencies (Zhou et al., 16 Sep 2025). The paper partitions samples into highly known, weakly known, and unknown regions using repeated response sampling: $\phi(X)$5 with $\phi(X)$6, $\phi(X)$7, and $\phi(X)$8 defining the three regions. It also measures strategy entropy

$\phi(X)$9

The proposed RL reward combines correctness-oriented confidence with a region-specific entropy term: low entropy is rewarded in highly known and weakly known regions, while high entropy is rewarded in the unknown region. On Qwen-2.5-7B-Instruct, the method improves ESCov macro F1 from $X$0 to $X$1, lowers bias $X$2 from $X$3 to $X$4, and raises weighted F1 from $X$5 to $X$6 (Zhou et al., 16 Sep 2025). The paper’s broader claim is that preference bias can emerge because the model collapses onto low-risk default strategies when it is near or beyond its knowledge boundary.

These results support a general interpretation: knowledge preference bias is often not a simple preference for incorrect information. It is a confidence-calibrated preference reversal, in which the model favors what is easy to predict over what is useful to know.

4. Preference recovery under biased observations

In recommender systems and tutoring, the problem often begins with the fact that observed data are a biased and incomplete proxy for underlying preference. BPL formalizes this in explicit-feedback recommendation under missing-not-at-random data (Kang et al., 17 Oct 2025). Let $X$7 be the full user–item space, $X$8 the observed subset, and $X$9 the unrated subset. The ideal model would minimize

$\phi(X)$0

but standard training minimizes only empirical risk on $\phi(X)$1. The paper’s solution, Bias-adaptive Preference distillation Learning (BPL), combines supervised learning on rated data, adversarial alignment between rated and unrated distributions, and affinity-adaptive dual distillation on unrated data. Its central distinction is between the factual test environment, which measures prediction of actual platform behavior, and the counterfactual test environment, which measures debiasing under randomized or full exposure. On Yahoo!R3, BPL-Hard achieves MSE $\phi(X)$2 and MAE $\phi(X)$3, outperforming InterD’s $\phi(X)$4 and Stable-MRDR’s $\phi(X)$5. The paper’s broader claim is that observed behavior contains both harmful and useful signal, so debiasing must preserve feedback-aligned knowledge where it is trustworthy while expanding beyond it where it is not.

“Crank up the volume: preference bias amplification in collaborative recommendation” studies a closely related distortion at the group/category level (Lin et al., 2019). For a user group $\phi(X)$6 and item category $\phi(X)$7, it defines the input preference ratio $\phi(X)$8, the output ratio $\phi(X)$9, and the bias disparity

$\succ$0

Positive $\succ$1 indicates amplification; negative $\succ$2 indicates dampening. In MovieLens Action/Romance experiments, UserKNN increases output preference ratio on Action by about 50% and ItemKNN by about 20%, for both men and women, even though women were approximately balanced in the input. BiasedMF and SVD++ often reverse the skew, while WRMF stays relatively close to the input distribution. The paper thus shows that recommendation algorithms do not merely reproduce observed preferences; they can amplify them, dampen them, or transfer majority-group structure onto minority groups.

In intelligent tutoring systems, DisKT treats the students’ historical correct-rate distribution over question groups as a confounder $\succ$3 that contaminates the latent student representation $\succ$4 and the prediction score $\succ$5 (Zhou et al., 4 Mar 2025). It defines

$\succ$6

where $\succ$7 is student $\succ$8’s historical correct rate on concept $\succ$9. DisKT then models separate familiar and unfamiliar abilities and uses a contradiction attention mechanism to shield against guessing and mistaking. Across 11 public datasets and 3 synthesized datasets with different bias strengths, the paper reports that DisKT achieves the best AUC/ACC/RMSE overall and the lowest $\succ$0 on the synthetic bias benchmarks. Here, knowledge preference bias is not a source-selection conflict but a confounded estimate of mastery induced by uneven exposure and success histories across concept groups.

5. Expert judgment, rational learning, and paternalistic intervention

Human experts exhibit analogous phenomena when they interpret sparse or ambiguous evidence through prior knowledge. “Modular interface for managing cognitive bias in experts” argues that experts are indispensable precisely where data are incomplete and theory-laden, but that the same conditions make them prone to overweight familiar conceptual models, accepted theories, and socially reinforced interpretations (Whitehead et al., 2022). The paper’s examples are concrete. In Bond et al. (2007), a synthetic seismic section was shown to 412 experts: only 21% identified the correct tectonic setting, and only 23% identified the three main fault strands. The proposed response is the modular MICE architecture, centered on “stop-and-perform” action modules such as Step-back, Pre-mortem, Devil’s advocate, Explicit knowledge, and Forced anonymity. These modules do not assume that experts self-correct; instead, they proceduralize challenges to anchoring, availability, representation bias, and single-expert over-dominance.

“Rationally Biased Learning” shows that some apparently cognitive biases can arise from optimal learning under imperfect information (Lara, 2017). The setup is a repeated choice between a safe action and a risky action whose payoff distribution is unknown. Learning occurs only by taking the risky action; choosing safety prevents future information acquisition. The optimal policy has a one-switch structure: experiment while the posterior remains favorable enough, but once the posterior enters an avoidance region, switch to safety forever. Under a Beta-Bernoulli model, if the true bad outcome is rare but a bad draw triggers stopping, the posterior can overestimate the bad-state probability relative to truth. This generates status quo bias, overattention to bad outcomes, and pessimistic overestimation of rare losses without assuming faulty inference. In this sense, knowledge preference bias can emerge from optimal stopping itself: the agent selectively censors future evidence by choosing the safe option.

A related institutional version appears in “Knowledge and Freedom,” which studies how policymakers use decision-makers’ knowledge as a basis for granting or denying autonomy (Grossmann, 2024). The welfare function is

$\succ$1

where $\succ$2 is the weight the Choice Architect places on her own preference. Across two experiments, full decision-maker knowledge causes more than a 60% reduction in intervention rates. At the same time, interveners are less likely to provide information: in the treatment where they could both inform and intervene, those who intervened were 29.7 percentage points less likely to provide information. Moreover, when ignorance-induced mistakes aligned with the CA’s own preference, information provision fell by 13.4 percentage points. The estimated $\succ$3 is about one-third. The paper therefore shows that ignorance is not used in a purely impartial way; it can become a selective justification for intervention, and the imposed option reflects both the Chooser’s full-information preference and the policymaker’s own.

6. Editing, unlearning, distillation, and acquisition as explicit preference engineering

Several recent methods do not merely diagnose knowledge preference bias; they engineer it explicitly. Knowledge editing via KDPO treats model updating as preference optimization (Rozner et al., 2024). The desired new completion $\succ$4 is the positive sample, while the model’s current completion $\succ$5 is the negative sample. Standard DPO is adapted by generating and evaluating the negative under a teacher-forced new-knowledge context, yielding

$\succ$6

The paper’s argument is that teacher forcing localizes the change by penalizing old tokens only under the new-fact context, which reduces collateral damage. Empirically, KDPO is consistently strong on locality and remains more robust than vanilla DPO under 500 sequential edits, while preserving HellaSwag, Winogrande, and MMLU better than several baselines.

GraphDPO applies the same basic logic to knowledge graph unlearning (Liu et al., 28 Jul 2025). Forgetting triples are reframed as dispreferred outputs, and alternative completions become preferred outputs under DPO. The method adds out-boundary sampling to avoid semantically entangled alternatives and boundary replay/distillation to preserve neighboring retained knowledge. Across eight unlearning datasets derived from FB15K-237, WN18RR, CoDEx-L, and YAGO3-10, GraphDPO outperforms approximate baselines by up to 10.1% in $\succ$7 and 14.0% in $\succ$8. This is a direct example of deliberate preference shaping over knowledge items: some facts are made systematically dispreferred, while nearby valid facts are explicitly protected.

Group Relative Knowledge Distillation (GRKD) makes a related move from absolute probabilities to relative preferences among outputs (Li et al., 29 Apr 2025). With teacher scores $\succ$9 and student probabilities $K$0, it constructs

$K$1

and optimizes a group-relative loss

$K$2

combined with a soft-target loss. The paper’s claim is that much of the teacher’s useful inductive bias lies in pairwise class ordering rather than absolute confidence values. In LLM alignment experiments, it reports over 20% improvement on AlpacaEval 2.0 and an average improvement of 9% across four datasets.

Even acquisition functions can encode systematic preference over what knowledge to elicit next. In preferential Bayesian optimization, the exact knowledge gradient for pairwise comparisons is analytically derived under a GP surrogate and a probit preference model (Wu et al., 29 Jan 2026). The paper notes that the acquisition “usually has low values along the diagonal region where $K$3 and $K$4 are close to each other since such queries are less informative.” It therefore biases query selection away from near-tie comparisons and, in a 2D Lévy case study, toward comparisons around the current maximizer rather than directly on it. This illustrates a broader methodological point: preference-based optimization procedures are themselves sources of systematic selection bias over knowledge acquisition.

7. Multimodal, graph, and evaluator bias

In multimodal systems, knowledge preference bias often manifests as dominance of one evidential channel. MedCFVQA studies modality preference bias in medical visual question answering, where predictions are “heavily dominated by one modality while overlooking the other,” usually with the question dominating the answer while the image is overlooked (Ye et al., 22 May 2025). To expose this, the paper constructs SLAKE-CP and RadVQA-CP by changing prior dependencies between questions and answers in train and test. The proposed counterfactual debiasing method estimates the shortcut effect of the dominant modality and subtracts it at inference. The largest gain is on SLAKE-CP, where accuracy improves from 0.337 to 0.430; on RadVQA-CP it improves from 0.735 to 0.755. The broader implication is that a multimodal architecture can still exhibit knowledge preference bias if one modality’s priors dominate the intended multimodal reasoning path.

In knowledge graphs, Fair-KGNN argues that bias is not only entity-wise but also multi-hop relational (Chuang et al., 2022). Its motivating example is gender–occupation bias: because there are more male engineers than female engineers, the two-hop path Occupation $K$5 People $K$6 Gender makes Engineer structurally closer to Male than to Female, with probabilities 0.625 versus 0.375. The method combines Sensitive Bias Mitigation (SBM), which reduces differential inclination scores between sensitive groups and non-sensitive entities, with Entity Relation Preservation (ERP), which keeps sensitive-entity embeddings separated so that graph semantics are not obfuscated. On FB15k-237, Fair-KGNN(CompGCN) reaches MRR 0.2238 while reducing $K$7 and $K$8 substantially relative to CompGCN. This extends knowledge preference bias to graph-structural inclination: the graph itself encodes asymmetric association strength between sensitive groups and non-sensitive concepts.

Finally, evaluator models can themselves exhibit knowledge preference bias. Philautia-Eval studies model-specific preference bias in MLLM-as-a-Judge for image captioning (Koyama et al., 13 Apr 2026). Using 4,500 images, 54,000 generated captions, 12 MLLMs, and 1,296,000 scores, it constructs a generator–evaluator matrix $K$9, standardizes it column-wise and row-wise, and defines the diagonal standardized value as the philautia score, or degree of self-preference. In both reference-based and reference-free settings, all 12 models have positive philautia scores. In the reference-based setting, InternVL2.5-8B reaches 3.02, whereas GPT-4o is comparatively low at 0.55. The paper also finds mutual preference bias within model families, including a Qwen-based $$C(K)=\exp\left(\frac{\sum_1^T \log p_{\pi_\phi}(K_t\mid K_{<t})}{T}\right),$$0 block with 9 of 12 positive off-diagonal entries, and interprets this as potentially driven by reused connectors and overlapping instruction-tuning resources. The ensemble Pomms reduces the philautia score to 0.15 while maintaining competitive Kendall correlation with human judgments. This is a particularly direct demonstration that learned evaluators may reward outputs that are epistemically or stylistically familiar to them, rather than judging quality alone.

Across these multimodal and graph settings, a common pattern emerges. Knowledge preference bias is not exhausted by wrong answers or demographic disparities. It also includes shortcut reliance on one modality, multi-hop structural inclination in graph representations, and evaluator favoritism toward outputs that resemble their own generative priors. This suggests that robust mitigation requires not only better prediction, but explicit control over which knowledge source, pathway, or stylistic regime is being privileged at each stage of the pipeline.