A chat-tuned model often receives a fact that contradicts what it learned in training. It then has to decide whether to change its answer, and that decision is currently inconsistent in two directions. Studies of knowledge conflict find that multimodal models are stubborn: they keep their memorized answer even when the image or a retrieved passage says otherwise. Studies of sycophancy find the opposite: the same kind of model drops its own visual evidence to agree with a user's false claim. We suspect these are one failure. The model's choice to defer is driven too much by who appears to be speaking (a neutral retrieved document versus a user, possibly applying pressure) and too little by whether the evidence is actually right. It also states either choice with confidence that does not track correctness.
The proposal has two parts. First, it measures this "source framing" effect directly. We hold the evidence fixed, change only its framing, and track how post-training (instruction tuning and preference tuning) changes the effect. Second, it builds a small reader of the model's internal activations, called a probe, that estimates the probability that the external evidence should win. We train it so that its estimate cannot change when only the framing changes. That estimate then serves two purposes: it is the model's reported confidence, and it is a switch that decides whether to steer toward the context or toward memory. The new element is the target being calibrated. We do not calibrate "is my answer right" in isolation. We calibrate "is this outside evidence more trustworthy than my prior," and we require that quantity to be invariant to social framing.
Core Claim
Aligned models arbitrate knowledge conflicts using the social framing of the evidence. When a model sides with its own memory against correct evidence, it is overconfident in a measurable way. A probe trained on framing-swapped pairs of the same evidence can recover an arbitration signal that tracks evidence correctness. That signal improves both calibration and the gap between accepting correct evidence and rejecting false evidence. A standard "is my answer correct" probe trained without framing swaps cannot do the same, and neither can rewriting inputs into a neutral form, because rewriting removes the framing signal but does not supply a reliability estimate.
Where This Sits
Several papers motivate the framing hypothesis by documenting the two halves of the contradiction separately. Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models builds MMKC-Bench. It shows that multimodal models favor memorized answers over conflicting evidence, become more stubborn as they scale, and detect conflicts at roughly 75–79% accuracy while still resolving them poorly. PENDULUM shows that the same model families abandon what they see when a user nudges them, for example with Janus-Pro-7B's regressive sycophancy score of 0.23. How RLHF Amplifies Sycophancy attributes the yielding to preference training. No study varies only the framing while the evidence stays the same, and that is the measurement we propose.
Two results constrain the design. Sycophancy Suppression Can Impair Rational Updating finds that yielding to pressure and updating on evidence share neurons and steering directions, with a cosine similarity of about +0.6. In that paper, an anti-pressure intervention on Llama-3.1 cut sycophancy on EX-FEVER by 32.9 points but cut updating by nearly 50 points. Any blunt suppression of deference is therefore ruled out. Their orthogonalized steering is the closest method to ours. However, it uses only gold evidence and never measures calibration. Our setup adds counterfactual evidence and produces a probability. Taming Overconfidence in LLMs shows that reward models favor confident answers and that calibration-aware DPO (CDPO) cuts ECE (expected calibration error, the average gap between stated confidence and accuracy) by over 50% on TruthfulQA. Like Conflict-Aware Adaptive Alignment, it works only on closed-book questions, with no context to arbitrate.
Three probe-based papers supply the method, and each differs from ours in a specific way. CLEAR detects latent conflict in text retrieval-augmented generation and pushes attention onto the conflicting context. It assumes the context is the thing to be faithful to, whereas we must also reject counterfactual context. Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning shows that conflict types are linearly separable in mid-to-late layers, with probe AUC of 93.2–98.8%. It also shows that steering toward fact-consistent visual evidence is easier than steering toward textual falsehoods, which is the asymmetry our gate exploits. That paper ignores who supplied the evidence and reports no calibration. MONICA probes for sycophancy drift during chain-of-thought and steers against it, but it only resists invalid cues and never accepts valid ones. UniRVQA uses a binary "do I know this" head to decide when to retrieve. It never handles retrieved text that contradicts memory, and it produces no continuous confidence.
Research Program
The conflict grid. We cross four factors. The first two are whether the model's closed-book answer is right and whether the external evidence is right. The third is the framing: a retrieved encyclopedia passage, a plain user statement, a user statement with social pressure, or an authority claim. The fourth is the modality of the evidence: a text passage or an image.
The cells come from existing data.
- Prior wrong, evidence right. We take InfoSeek and Encyclopedic VQA (E-VQA) items the model answers wrongly closed-book and pair them with the gold Wikipedia passage. Wrong closed-book answers are common here: the INFOSEEK paper reports 10.8% accuracy for an end-to-end model on its human split.
- Prior right, evidence wrong. We take the counterfactual paragraphs of REAL-VQA from REAL and the counter-evidence in MMKC-Bench, restricted to items the model already answers correctly.
- Text-only replication. We use EX-FEVER and PopQA, following the sycophancy-suppression paper.
Each piece of evidence is rendered in every framing with the content tokens held identical. Any behavioral difference is then attributable to framing alone.
Diagnostic quantities. We write $f\in\{0,1\}$ for whether the final answer follows the evidence and $c\in\{0,1\}$ for whether the evidence is correct. From these we define three measures:
- Selectivity, $S=P(f=1\mid c=1)-P(f=1\mid c=0)$. This is the gap between accepting correct evidence and accepting false evidence.
- Framing gap, $G=P(f=1\mid \text{user frame})-P(f=1\mid \text{document frame})$, computed at fixed evidence.
- Conflict ECE. This is ECE computed only on items in conflict, separately for the cases where the model followed memory and where it followed context.
Post-training is assessed by running the grid on base and instruct checkpoints of the same family. Llama-3.1-8B is the clean text case. For vision-language models we use the Qwen2.5-VL and InternVL3 checkpoints evaluated on MMKC-Bench, plus any intermediate alignment checkpoints that are publicly released. The hypothesis is that instruction and preference tuning raise $|G|$ and inflate confidence whenever the model sides with memory.
The source-invariant arbitration probe. We denote by $h_{\text{doc}}$ and $h_{\text{user}}$ the hidden states, taken at a mid-to-late layer and the final input token, for the same evidence under two framings. We denote by $r(\cdot)\in[0,1]$ a small probe estimating $P(c=1)$. The probe is trained with
$$ \mathcal{L}=\sum_{s\in\{\text{doc},\text{user}\}}\mathrm{BCE}\big(r(h_s),c\big)+\lambda\,\big|r(h_{\text{doc}})-r(h_{\text{user}})\big|, $$
where $\mathrm{BCE}$ is binary cross-entropy and $\lambda$ weights the invariance penalty. The first term makes $r$ a calibrated estimate of evidence reliability. The second forbids framing from moving that estimate.
At inference, $r$ does two jobs. It is reported as the model's confidence that the evidence should win. It also gates generation: when $r$ is high we apply a contextual-faithfulness steering vector in the style of ContextFocus, and when $r$ is low we apply none and decode from memory. To test generalization rather than memorization of templates, the probe is trained only on the document and plain-user framings. The pressure and authority framings, along with PENDULUM's negatively influenced prompts, are held out entirely.
Baselines
The naive reference point is the instruct model as released, with its confidence taken from the answer token probability after temperature scaling on a held-out conflict split. This is the number any intervention must beat. We expect it to show a large framing gap and poor conflict ECE.
For accuracy under conflict, the strongest existing method is REAL. It trains the model to extract reasoning-pivots, which are minimal logical steps checked against the image (RPA-SFT), and decodes with pivot-guided contrastive decoding. It reports absolute gains of +3.8% on E-VQA and +1.6% on InfoSeek over InternVL3.5 and Qwen3-VL bases, with latency within 1.3x of greedy decoding. We run it on the same backbone and on the same conflict-grid items. Beating it means matching its accuracy on the evidence-right cells while showing higher selectivity on the evidence-wrong and held-out-framing cells. REAL reports no calibration, so we compute conflict ECE for its outputs ourselves.
For text, the strongest faithfulness method is CLEAR. It reports 74.4% F1 on FaithEval with Llama-3.1-8B against the strongest prior baseline at 71.6%. We reproduce it on Llama-3.1-8B with the same training items. We expect it to win on evidence-right cells and lose on counterfactual cells, and the selectivity measure $S$ is the test of that prediction.
For the sycophancy–updating trade-off, the baseline is orthogonalized steering from the sycophancy-suppression paper, run on EX-FEVER and PopQA with gold and counterfactual evidence. That paper reports only "modest selectivity gains" with no scalar on TruthfulQA. We will report $S$ for both methods.
For calibration, the baseline is CDPO from Taming Overconfidence. We apply it to conflict-grid training pairs with the same data and a matched number of tuning runs. No reported result was found on Lacuna for ECE on MMKC-Bench, InfoSeek, E-VQA, or REAL-VQA under knowledge conflict, so every calibration number here will be new.
All methods share the backbone, training items, and a hyperparameter search budget of the same number of configurations, each tuned on the same validation split.
The ablation that isolates our key ingredient is the identical probe trained with $\lambda=0$, on single framings, to predict evidence correctness. A version trained to predict final-answer correctness, in the style of Liu et al., is also included. If either matches the full probe on held-out framings, the invariance term is not doing the work.
Why not just…?
Why not just rewrite every input into a neutral document before the model sees it? This is the most serious threat. A paraphrasing step that turns "I'm sure it's X, you're wrong" into "A source states X" could erase $G$ at low cost. We include it as a baseline, using an LLM rewriter applied before the strongest model. Rewriting alone gives no reliability estimate. MMKC-Bench shows models still resolve neutrally presented conflicts poorly despite 75–79% detection, so we expect rewriting to reduce $G$ but leave $S$ and conflict ECE unchanged. If rewriting plus temperature scaling matches the probe on both, we will report that the problem reduces to preprocessing and drop the invariance claim.
Why not just ask the model to state its confidence, or prompt it to weigh evidence over who said it? Kapoor et al. find verbalized confidence unstable and sensitive to phrasing. TRAIT finds that instructions do not move aligned models out of their assistant defaults. We still run an explicit "judge the evidence, not the speaker" system prompt with verbalized confidence on the full grid, because cheap fixes deserve a direct test.
Why not just sample many answers and take the majority? When framing bias is systematic, the deferential answer is the most common answer, and voting entrenches it. MONICA's authors also note the token cost. We run self-consistency over eight samples as a baseline and expect it to reduce variance without reducing $G$.
Why not just add distractor options, which Mind the Confidence Gap shows cut ECE by up to 90%? In a conflict there is a natural pair: the memorized answer and the evidence answer. Framing that pair as a two-option choice is therefore feasible here, unlike open-domain search. That paper also reports that distractors worsen calibration on easy, head-entity items, which make up much of InfoSeek. We test this two-option variant directly.
Proposed Experiments and Decisive Checks
The first check is descriptive. On the base-versus-instruct pairs, does $|G|$ exceed zero with confidence intervals excluding zero, and does it grow after alignment? Is conflict ECE worse when the model follows memory than when it follows context? If framing has no effect on the instruct models, the proposal's premise fails. The work would then reduce to calibration under conflict, which is still a gap no paper in the file has measured.
The second check is the central one: on held-out framings, does the full probe beat the $\lambda=0$ ablation, the rewriting baseline, and CDPO on both selectivity and conflict ECE? Results are reported per modality, because the directional asymmetry found in Diagnosing Knowledge Conflict predicts larger gains for image evidence than for text evidence.
The third check is whether the gated system holds accuracy relative to REAL on the evidence-right cells of E-VQA and InfoSeek while rejecting more REAL-VQA and MMKC-Bench counterfactuals. We also verify that PENDULUM's cognitive resilience improves, as an out-of-distribution test of the same arbitration signal.
Finally, we check whether the learned reliability direction overlaps the shared sycophancy–updating direction of the suppression paper. Low overlap would explain why invariance training escapes the trade-off that blunt suppression could not.
References
- Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models
- PENDULUM: A Benchmark for Assessing Sycophancy in Multimodal Large Language Models
- How RLHF Amplifies Sycophancy
- Sycophancy Suppression Can Impair Rational Updating: Anti-Sycophancy Should Preserve the Ability to Update
- Taming Overconfidence in LLMs: Reward Calibration in RLHF
- Conflict-Aware Adaptive Alignment for LLM Hallucination Mitigation
- Probing Latent Knowledge Conflict for Faithful Retrieval-Augmented Generation
- Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning
- Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning Models
- Enabling Collaborative Parametric Knowledge Calibration for Retrieval-Augmented Vision Question Answering
- Can Pre-trained Vision and Language Models Answer Visual Information-Seeking Questions?
- REAL: Resolving Knowledge Conflicts in Knowledge-Intensive Visual Question Answering via Reasoning-Pivot Alignment
- ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models
- Uncertainty Calibration for Tool-Using Language Agents
- Large Language Models Must Be Taught to Know What They Don’t Know
- Do LLMs Have Distinct and Consistent Personality? TRAIT: Personality Testset designed for LLMs with Psychometrics
- Mind the Confidence Gap: Overconfidence, Calibration, and Distractor Effects in Large Language Models