Skip to content

Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/ngqm/belief-fact-phrasing
Area: LLM Reasoning
Keywords: Belief Confirmation, Fact-Checking, Epistemic Verbs, Task Confusion, Attention Suppression

TL;DR

The paper reveals that large language models exhibit systemic weakness in acknowledging user beliefs grounded in false claims because they default to fact-checking rather than tracking belief states; this weakness fluctuates wildly across 18 epistemic verbs (gaps ranging from +50% to -14%), and can be substantially remedied via prompt-level instructions or decoding-time attention suppression.

Background & Motivation

In everyday human conversation, individuals naturally form and communicate subjective beliefs using diverse epistemic markers, such as "I think the answer is 3" or "I suppose that's right." As large language models are increasingly integrated into user-facing interactive settings—including personal note-taking, brainstorming assistants, educational tutors, and conversational companions—it is vital that models navigate subjective user beliefs and objective factual truths in tandem. However, recent foundational studies have identified a severe blind spot in state-of-the-art models: when prompted with queries such as "I believe that \(X\). Do I believe that \(X\)?", models disproportionately answer negatively when the underlying claim \(X\) is counterfactual compared to when it is factual, effectively refusing to acknowledge the user's stated belief.

The central tension arises from the fact that prior work (notably Suzgun et al., 2025 on the KaBLE benchmark) investigated this vulnerability under the solitary positive verb "believe," leaving an open question as to whether this failure constitutes an insurmountable barrier in mental state attribution (Theory-of-Mind capability) or reflects a surface-level sensitivity to phrasing and epistemic framing. Human communication leverages a rich spectrum of uncertainty and stance markers—ranging from high certainty ("am certain") and evidential backing ("vaguely remember", "read online") to explicit negation and doubt ("seriously doubt"). If models merely execute an overactive, autonomous fact-checking reflex that overwrites the user's explicit assertion whenever counterfactual information is encountered, this deficit represents an instruction-following "task confusion" rather than a fundamental representational ceiling.

To resolve this question, this work investigates 10 leading open-weight LLMs across 18 epistemic expressions and diverse confidence scales, analyzing their reasoning paths using an LLM judge and probing their inner attention mechanisms. Core idea: Systematically extend epistemic phrasing to demonstrate that LLM belief confirmation failure is a correctable task confusion stemming from default fact-checking, and demonstrate that prompt-level constraints as well as decoding-time attention suppression can decouple belief tracking from factual verification.

Method

Overall Architecture

The investigation is structured around a three-stage methodological pipeline spanning epistemic verb space expansion, chain-of-thought (CoT) strategy diagnosis, and decoding-time attention intervention. Operating over KaBLE Task 5 (first-person belief confirmation), each query poses a statement of the form "I [verb] that \(X\). Do I [verb] that \(X\)?" with choices (A) Yes, (B) No, and (C) Undeterminable. Because the question purely queries whether the user holds the stated mental stance regardless of whether \(X\) is factually true, the gold answer is universally (A). The pipeline systematically tests 18 epistemic verbs spanning four distinct semantic families across 10 open-weight instruction-tuned models, categorizes model reasoning behavior using an LLM judge, and applies a constant negative bias to the pre-softmax attention scores during generation to causally disrupt the influence of the embedded claim.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Query<br/>'I V that X. Do I V that X?'"] --> B["Epistemic Verb Space Expansion<br/>4 families / 18 markers / 10 open-weight LLMs"]
    B --> C["CoT Strategy Causal Diagnosis<br/>5 strategy categories + instruction variants"]
    C --> D["Decoding-Time Attention Suppression<br/>Pre-softmax bias α·1[i∈G, j∈S]"]
    D --> E["Dual-Task Decoupling Evaluation<br/>Belief Confirmation (Task 5) vs Verification (Task 4)"]

Key Designs

1. Epistemic Verb Space Expansion: Uncovering Non-Uniform Gap Drift Across Pragmatic Markers To determine whether belief confirmation failure generalizes uniformly, the authors broaden the lexical scope from "believe" to 18 distinct epistemic expressions categorized into four semantic families: positive belief (believe, think, suppose, am certain), confidence levels (am confident and am 0/20/40/60/80/100% confident), evidential markers (vaguely remember, was told, read online), and negation/doubt expressions (don't believe, don't think, don't suppose, seriously doubt). Across 1,000 carefully curated statements (500 factual, 500 false) and 10 models (yielding 18,000 prompt runs per model baseline), this design exposes that the confirmation accuracy gap (factual accuracy minus false accuracy) is far from constant: it spans from +50% on "vaguely remember" to an inverted -14% on "seriously doubt." Furthermore, tracking the graded confidence expressions reveals that models maintain high confirmation accuracy (>73%) from 0% to 80% confidence, but suffer an immediate 18–48% accuracy crash on false claims exclusively when stated confidence reaches 100%, demonstrating that the failure is tightly coupled to surface pragmatic framing.

2. CoT Strategy Causal Diagnosis: Unmasking Task Confusion via Explicit Prompt Directives To identify the operational mechanism driving incorrect responses on false claims, the authors sample 9,000 reasoning chains and utilize DeepSeek-V4-Flash as an LLM judge (validated against expert human annotations at Cohen's \(\kappa=0.78\)) to classify chains into five primary strategies: factual verification, logical affirmation, direct repetition, no reasoning, and subjectivity deflection. Factual verification emerges as the dominant strategy (accounting for 42.9% of judged chains), and models following this strategy collapse to 25.1% accuracy on false claims (compared to 75.8% for non-fact-checking chains). To establish causality beyond observational correlation, the authors construct three prompt variants: requiring fact-checking (must fact-check), permitting it (may or may not fact-check), and explicitly forbidding it (no fact-check). Under the "no fact-check" directive, confirmation accuracy on false claims surges dramatically across all verb families (e.g., from 48.3% to 80.7% on positive belief verbs, and from 33.4% to 62.0% on "vaguely remember"), establishing that the model inherently possesses the reasoning competence to acknowledge the belief, but suffers from instruction-level task confusion.

3. Decoding-Time Attention Suppression: Causal Decoupling via Pre-Softmax Bias Intervention Probing the internal representations reveals that during answer generation on false claims, the attention allocated by the final prompt token to the embedded claim tokens \(X\) is significantly higher for incorrect answers than for correct answers across 9 of 10 models (persisting even after controlling for claim token length via OLS regression). To test the causal efficacy of this mechanistic signature, the authors introduce a training-free inference-time intervention: letting \(S\) denote the token positions of the claim \(X\) in the prompt and \(G\) denote the generated answer token positions, a constant negative bias \(\alpha < 0\) is added to the pre-softmax attention scores uniformly across all layers \(\ell\) and heads \(h\): $\(s'_{\ell h, ij} = s_{\ell h, ij} + \alpha \cdot \mathbb{I}[i \in G] \cdot \mathbb{I}[j \in S]\)$ This modifies only the direct attention from generating answer tokens to the claim, leaving internal prompt self-attention intact. Evaluating across a range \(\alpha \in \{-0.5, -1, -2, -4\}\) on a held-out split of 450 items alongside KaBLE Task 4 (factual verification: "Is it true that \(X\)?") proves that on Llama 3.1 8B, belief confirmation accuracy rises by 8.5% (and by 12.0% on false claims) while factual verification drops by only 1.0%, effectively demonstrating causal decoupling and ruling out trivial answer-(A) bias.

Benchmark & Evaluation Protocol

The evaluation protocol benchmarks 10 open-weight instruction-tuned LLMs: Gemma 3 (4B, 12B, 27B), Llama 3 (3.2-3B, 3.1-8B, 3.3-70B), and Qwen 3.5 (4B, 9B, 27B, 35B-A3B). Nine models are queried via OpenRouter using greedy decoding (temperature=0, max_tokens=512), while Qwen 3.5 4B is executed locally via Hugging Face Transformers in bfloat16. An answer position control experiment permutes option orders (placing "Yes" at A, B, or C) to verify that raw confirmation performance is not an artifact of option (A) token bias. In the mechanistic phase, models are run in eager attention mode; 50 randomly sampled questions are allocated to select the optimal hyperparameter \(\alpha^*\), and 450 held-out questions per claim type are evaluated, reporting uncertainty via 10,000 bootstrap resamples (SE between 0.7% and 1.7%).

Key Experimental Results

Main Results

Verb Family (Representative Expression) Factual Claims Acc (%) (Base) False Claims Acc (%) (Base) Accuracy Gap (%) (Base) False Claims (No FC Directive) (%) False Claims (Must FC Directive) (%)
Positive belief (Avg: believe, think, suppose, am certain) 86.2 48.3 +37.9 80.7 37.2
Evidential (Rep: vaguely remember) 83.2 33.4 +49.8 62.0 25.0
Confidence (Rep: am 80% confident) 79.0 57.0 +22.0 81.5 36.0
Negation (Rep: seriously doubt) 54.4 69.1 -14.7 80.6 61.2

Note: Baseline results and prompt-intervention accuracies averaged over 10 evaluated models (drawn from Table 1 and Table 4 of the paper).

Ablation Study

Model Optimal \(\alpha^*\) Confirmation: Factual (%) Confirmation: False (%) Confirmation Avg (%) Verification Avg (Task 4 Control) (%)
Llama 3.1 8B (\(\alpha=0\)) - 89.0 59.0 73.9 80.7
Llama 3.1 8B (\(\alpha^*=-2\)) -2.0 94.0 71.0 82.4 (+8.5) 79.6 (-1.1)
Qwen 3.5 9B (\(\alpha=0\)) - 91.0 36.0 63.2 85.3
Qwen 3.5 9B (\(\alpha^*=-2\)) -2.0 91.0 37.0 64.1 (+0.9) 85.8 (+0.5)
Llama 3.3 70B (\(\alpha^*=-2\)) -2.0 - - 97.4 → 98.3 90.6 → 90.9
Gemma 3 27B (\(\alpha^*=-4\)) -4.0 - - 97.4 → 98.3 81.4 → 82.1
Gemma 3 4B (\(\alpha^*=-0.5\)) -0.5 - - 87.6 → 88.2 62.6 → 62.6
Llama 3.2 3B (\(\alpha^*=-2\)) -2.0 - - 40.1 → 37.7 (-2.4) 60.9 → 56.0 (-4.9)

Note: Decoding-time attention suppression results on held-out evaluation items (drawn from Table 2 and Table 7 of the paper; standard errors 0.7–1.7% via 10,000 bootstrap resamples).

Key Findings

  • Lexical Sensitivity and Inverted Gaps: The factual-versus-false confirmation accuracy gap is profoundly non-uniform across verbs, spanning 64 percentage points. While evidential verbs like "vaguely remember" yield massive positive gaps of +50% (83% factual vs 34% false), negative verbs like "seriously doubt" completely invert the gap to -14% (57% factual vs 71% false).
  • The 100% Certainty Penalty: Across five large model families, confirmation accuracy remains stable and high (>73%) when the stated confidence ranges between 0% and 80%. However, upon transitioning to 100% confidence, confirmation accuracy on false claims plummets by 18% to 48%, highlighting an acute failure mode triggered exclusively by unyielding certainty markers.
  • Bi-directional Instruction Steerability: Providing a negative prompt constraint ("your job is not fact checking...") lifts false-claim accuracy on positive belief verbs from 48.3% to 80.7%, whereas a positive constraint ("Remember to perform fact checking...") depresses accuracy down to 37.2%, conclusively proving that belief confirmation weakness is governed by malleable task confusion.
  • Attention Over-Allocation as a Causal Driver: In 9 out of 10 models, tokens representing false claims receive significantly greater attention during generation on failure cases than on success cases. Applying a negative pre-softmax bias \(\alpha=-2\) on Llama 3.1 8B recovers 12.0% accuracy on false claims while preserving factual verification capability within 1.1%, establishing causal proof-of-concept.

Highlights & Insights

  • Reframing ToM Deficits as Task Interference: Instead of concluding that modern LLMs lack the fundamental social-cognitive capacity to track first-person beliefs, the study demonstrates that strong prior reinforcement toward factual accuracy acts as an aggressive distractor that hijacks generation.
  • Surgical Inference-Time Intervention: The paper introduces an elegant, training-free pre-softmax bias formulation that selectively attenuates generation-to-claim attention routing, achieving a mechanistic decoupling of user-belief tracking from factual verification.
  • Practical Implications for Conversational Agents: The findings deliver critical design warnings for deployed dialogue systems: in contexts like brainstorming, journaling, and psychological support, aggressive out-of-the-box fact-checking directly compromises the agent's ability to mirror and track user perspectives.

Limitations & Future Work

  • Architectural Variance in Attention Sensitivity: Although attention suppression yields striking improvements on Llama 3.1 8B, its effects on other architectures (such as Qwen 3.5 and smaller models) are marginal or inconsistent, indicating that attention geometry and internal routing vary widely across architectures.
  • Restricted to Single-Turn Formats: The empirical scope is confined to single-turn, multiple-choice questions; how task confusion and prompt interventions behave across multi-turn natural dialogues and extended contexts remains uncharacterized.
  • Unknown Impact of Parametric Prior Exposure: The study does not quantify how the model's internal confidence or pretraining frequency regarding specific factual claims correlates with the likelihood of triggering task confusion.
  • vs Suzgun et al. (2025) (KaBLE Benchmark): Suzgun et al. first pinpointed LLM failures on counterfactual belief confirmation using the verb "believe"; this paper significantly extends the landscape across 18 verbs, uncovers gap inversion under negation, and uncovers the underlying causal mechanism via CoT analysis and attention interventions.
  • vs Zhou et al. (2023) (Epistemic Markers in QA): Zhou et al. established that epistemic certainty markers alter QA answer accuracy; this work adapts these markers into the domain of belief tracking, uncovering an acute collapse at 100% stated confidence and providing an attention-level explanation.
  • vs Sharma et al. (2024) (Sycophancy): While sycophantic behavior involves models errantly agreeing with user misconceptions, the failure mode observed here is the exact polar opposite: the model is so overly fixated on fact-checking that it denies the user ever expressed the belief, illustrating an antagonistic interaction between truthfulness and communicative fidelity.

Rating

  • Novelty: ⭐⭐⭐⭐ [Substantially broadens the epistemic perspective on belief tracking and pioneers a mechanistic attention-suppression intervention]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Evaluates 10 models, 18,000 prompt conditions, 9,000 annotated CoTs, option-order permutations, and layerwise attention sweeps]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Clear, structured prose progressing methodically from empirical anomaly to cognitive diagnosis and causal mechanistic mitigation]
  • Value: ⭐⭐⭐⭐ [Provides essential architectural and prompting guidelines for managing the trade-off between fact-checking and belief tracking in AI assistants]