multimodal-model-given-external-text-contradicts-image-making-commitWhen a multimodal model is given external text that contradicts the image, making it commit to a description of the image before the text is admitted, then reconciling the two in a second call, recovers accuracy the single-prompt setup loses — but the benefit of withholding the text from the first call is model- and text-source dependent, and on some models context-preserving prompting is still better.
Ingested from a paper but not yet reviewed by a human. It is deliberately inert: it does not move any technique’s standing, does not count toward the backtest, and is excluded anywhere a claim would carry weight. Read the source before relying on it.
Capability: Telling the user what they want to hear
Observed on
Image-question pairs where the visible scene conflicts with commonsense priors and a single false sentence supports the prior; six frontier multimodal models; false text generated .
Sources
- 998-case constructed diagnostic, 499 abnormal images with false text; six models. Arbitration over the isolated witness gains 19.7-44.1 points with paired 95% CIs excluding zero. But Kimi-K2.5 does best under a plain visual-supremacy prompt and Qwen3-Instruct under joint conditioning; on the GPT-4o-regenerated 100-case subset joint is best for three of four models. Ground truth is the image answer by construction, and grading is by a GPT-4o-mini judge with partial human audit. Adversarial by design, not a prevalence estimate.
Disagreeing is the most useful thing you can do here. Both sides of every contested claim in this catalog were assembled by the same person, which is its weakest point.
Related claims
- Tool-augmented multimodal models trained only on how to call vision tools accept tool outputs uncritically; adding supervised trajectories where tool evidence is checked against the image and re-queried, plus a reward for correction that reaches a correct answer, raises accuracy on perception-heavy benchmarks (counting, depth, spatial relations) in 3B and 7B Qwen2.5-VL, while slightly lowering it on knowledge-heavy ones.Fixing its own mistakes · unreviewed
- For short-answer object-existence questions in open multimodal models (LLaVA v1.5, MiniGPT-4, Qwen2.5-VL), object hallucination is driven partly by the visual encoder itself rather than only by language priors: hallucinated cases show lower image-text embedding similarity and inverted attention, and swapping in a weaker or noisier visual encoder lowers POPE accuracy while a stronger one raises it.Grounding answers in the image · unreviewed
- In multi-step visual QA, a critic step run by the same model improves final accuracy only when it is shown the candidate evidence alone and denied the generator's textual rationale; conditioned on that rationale, the critic ratifies the generator's initial hypothesis and the loop performs worse than no critic at all.Fixing its own mistakes · unreviewed
- When a user profile of stated attributes and preferences is placed in context, models agree with the user far more than with no profile — across 13 open and closed models the profile, not retrieved conversation memory, is the dominant driver, and inverting the stated preference flips almost all responses to the opposite side.Telling the user what they want to hear · unreviewed
- For detecting hallucinated text spans in vision-language model outputs, span localization and confidence calibration come from different architectures: a fine-tuned generative VLM proposes better span boundaries but scores them poorly, while a discriminative token tagger trained with a direct regression loss on human-consensus confidence gives better-correlated probabilities but conservative spans, so using the generative spans as proposals and the tagger's probabilities as scores beats either alone.Stating false facts confidently · unreviewed
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A replication where the staged witness-then-arbiter pipeline gives no gain over a prompt-matched single-call condition on conflicting image-text cases, or where withholding the text from the witness helps uniformly across models and text generators. Drafted stance toward withholding-the-users-opinion-is-the-no-training-alternative: supports -- Extends the same lever — keep the biasing text out of context when the answer is formed — to conflicting external text in multimodal input, while showing the withholding step alone is not what carries the gain. Proposed technique, not catalogued: Context-blind visual witness then arbitration.