Ranked Visual Belief States for Multi-Turn Vision-Language Dialogue

When a vision-language model, meaning a language model that also reads an image, is told something false about a picture partway through a conversation, it tends to accept the claim and then build on it. A user says "the cup is red" when it is blue, and three turns later the model is describing a red cup and a red-themed kitchen. Iterated belief revision, the branch of logic that studies how an agent should absorb a stream of possibly conflicting information, has a precise answer to this. The agent should keep a ranking of how firmly it holds each belief, not just a list of current beliefs. New information should move only the beliefs it actually conflicts with. And when a later message withdraws an earlier one, the agent should be able to fall back to what it believed before. We propose to give vision-language models exactly this kind of state. The model keeps an explicit, ranked record of atomic visual claims about the image, such as "a cup exists" or "the cup is blue." Each claim's firmness comes from the model's own image-only answers, and each user statement's firmness depends on who said it. Each turn's answer is generated from the current most-plausible beliefs.

The core claim is narrow and testable. Two existing fixes help on single conflicts: ignoring dialogue history, and re-injecting visual features. Neither handles conversations where order matters, where users sometimes tell the truth about things the model got wrong, and where earlier claims are later retracted. A ranked state that obeys the Darwiche–Pearl constraints on revision should beat history-free re-anchoring on exactly those trajectories and match it elsewhere. If it does not, the formal machinery is not earning its keep, and we will say so.

What exists and where the gap is

The failure is well documented. MM-Snowball parses images into atomic visual propositions and builds six-turn dialogues in which an adversarial user perturbs the scene at turn 3. Accuracy follows a V-shape, and LLaVA-1.5-7B falls below 2% in turns 3–5. Its fix, Conflict-Aware Visual Rectification (CAVR), re-injects visual features into middle layers and blends next-token predictions with a stream that sees the image but not the history. Visual Flow finds the same collapse among communicating agents: visual attention drops by up to 60% by turn 20. It fixes this by relaying visual tokens and reweighting attention.

Both papers motivate our proposal, and both treat the fix as continuous re-anchoring. Neither keeps any record of which claims were accepted, from whom, or how firmly. So neither can tell a false user claim from a true correction of the model's own hallucination. Neither can restore turn-1 beliefs when a turn-3 claim is later withdrawn.

Volcano runs a single critique–revise–decide pass. Its accept/reject step has no principled criterion, and its latency rises from 2.7 to 5.8 seconds. Studies of model judges (evaluator unreliability, critique hallucination) show that such critiques themselves hallucinate. This constrains us: firmness must not come from free-form self-critique.

On the text side, Belief-R and ReviseQA show that language models handle static logic well but fail at sequential revision. In Belief-R, chain-of-thought adds only about 1%. In ReviseQA, accuracy falls below 50% on the hardest multi-edit tasks, and turn-by-turn feedback does not reliably help. These results argue against expecting prompting to fix the problem. Both benchmarks are text-only and propose no repair.

The method comes from the logic literature. Kern-Isberner's conditional revision uses Spohn's ranking functions, which assign each possible world an integer degree of implausibility, and shows that they satisfy the Darwiche–Pearl postulates. Sezgin and Kern-Isberner show how to implement bounded revision, which means "accept the new input only up to a limit." That is the natural model for a user whose trust is partial. Williams gives a finite, computable version of ranked belief bases. She also states the open problem we target: where the initial firmness values come from is outside logic.

Our answer is that firmness comes from the vision-language model's own image-only judgment. SHARP finds that faithful and hallucinated hidden states are linearly separable. That suggests such signals carry usable confidence, which we test rather than assume.

Mechanism

For each image, the system extracts a set of atomic propositions about entities, attributes, relations, and counts, using MM-Snowball's parsing procedure. It adds exclusivity constraints, for example that an object has one dominant color. Possible worlds are joint truth assignments. These factor by entity, so enumeration stays small; we cap clusters at roughly 15 propositions and fall back to a SAT solver beyond that.

The initial ranking $\kappa$ assigns each proposition $p$ a firmness derived from the model's image-only probability of "yes" to a templated question about $p$, with no dialogue history. Firmness is set to the rounded log-odds $\lfloor |\log(P(p)/(1-P(p)))| / \log b \rfloor$ in the direction the model favors, where $b$ is a base fixed on held-out data. This probability is optionally averaged over the image transformations used by RITUAL. A language-model parser maps each user utterance to a proposition $\alpha$ with a strength $m$ that depends on the source.

Revision then applies Spohn conditionalization:

$$\kappa'(w) = \begin{cases} \kappa(w) - \kappa(\alpha) & w \models \alpha \\ \kappa(w) - \kappa(\neg\alpha) + m & w \models \neg\alpha \end{cases}$$

Here $\kappa(\alpha)$ is the minimum rank over worlds satisfying $\alpha$. The rule makes $\alpha$ believed with firmness $m$. It also shifts the $\alpha$-worlds and the non-$\alpha$-worlds as blocks, so relative plausibility within each group is untouched; this is the Darwiche–Pearl property.

The consequences follow directly. A false claim about a confidently seen cup, with low $m$, fails to dislodge the visual belief. A true correction of a low-confidence hallucination succeeds. A later retraction restores the earlier ordering, so turn-1 beliefs return.

At answer time, the current most-plausible beliefs are passed to the model as a short grounded fact list alongside the image and question. Beliefs with rank margin 0 are marked uncertain.

Research Program

There are three parts. The first is calibration: do image-only probabilities, logit-lens readings, or SHARP-style probes give firmness values whose ordering predicts correctness on POPE and on MM-Snowball propositions?

The second is a benchmark extension, MM-Revise, built from MM-Snowball's propositions with up to seven turns, following ReviseQA's depths. It mixes four kinds of user turns: false claims, true corrections of propositions the base model gets wrong, retractions of earlier user claims, and true information about things not visible in the image. Each trajectory has a ground-truth answer per turn, computed from the annotations under a specified trust policy.

The third is the controlled comparison below, including a parser-oracle condition that separates belief-dynamics errors from utterance-parsing errors.

Baselines

The naive reference is the base model (LLaVA-1.5-7B and 13B, plus Qwen-VL) answering with full history. The file reports LLaVA-1.5-7B falling below 2% in MM-Snowball turns 3–5. Chain-of-thought prompting over the same history is the cheapest reasoning baseline.

The strongest existing method is CAVR from MM-Snowball. It is training-free and was designed for this benchmark. Its Snowballing Resistance Score has no reported result found on Lacuna, so we will reimplement it, check it against the paper's qualitative flattening of the V-curve, and run it on identical trajectories, models, and decoding settings. Its blending weight gets the same held-out tuning budget as our $b$ and $m$.

We also run a pure history-free re-query, which answers every turn from the image alone. This is the extreme version of CAVR's visual stream. Volcano-13B is included for single-revision quality; its reported MMHal-Bench score is 2.64 versus 2.54 for base LLaVA-1.5-13B. On single-turn POPE, the reference is RITUAL's reported 88.87% accuracy on LLaVA-1.5 versus 84.13% base. Our system must not fall below the base model there.

Two ablations isolate the key ingredient. The first, flat fact memory, keeps the same proposition list with last-writer-wins updates and no ranks. The second, memoryless ranking, recomputes ranks from scratch each turn from the image plus the current claim, discarding the revised ordering.

Beating CAVR means three things. Our method must achieve higher mean per-turn accuracy on MM-Revise at matched or lower latency. It must show a statistically significant gain, using a paired bootstrap over trajectories, on the true-correction and retraction subsets. It must be no worse on the adversarial-only subset.

Why not just…?

Why not just ignore the dialogue history and answer from the image? On adversarial-only trajectories like MM-Snowball's, this is a genuine threat and may tie or win. Our claim is that it fails whenever the user is right: it cannot accept a true correction of the model's own hallucination, nor information not visible in the image. MM-Revise's true-correction and out-of-view subsets measure exactly this. If history-free re-query matches us there too, the proposal fails.

Why not just accept a user claim whenever the model's visual confidence is below a threshold? This one-shot rule is bounded revision without memory, and it is the most serious alternative. It coincides with our method on any single conflict. It diverges only on sequences: after a retraction, or when a second claim depends on an earlier accepted one, there is no ordering to fall back to. The memoryless-ranking ablation is this rule. We report the gap separately on trajectories with and without order dependence. If the gap is near zero, the honest conclusion is that thresholding suffices, and iterated structure matters only for retraction-heavy dialogue.

Why not just prompt the model to reason carefully or give it corrective feedback? Belief-R reports about a 1% gain from chain-of-thought. ReviseQA finds that explicit per-turn feedback does not consistently help. We still run chain-of-thought on identical inputs, since the multimodal setting could differ.

Why not just use Volcano-style self-critique? Model critics hallucinate discrepancies (evaluator unreliability), and Volcano roughly doubles latency. Our firmness comes from a single image-only probability per proposition, cached across turns.

Decisive Checks

The first check is a correctness signal on the recovery subset, where a false user claim is introduced and later withdrawn. Per-proposition recovery rate, meaning the fraction of turn-1 beliefs correctly restored, should be clearly higher than for flat memory and CAVR.

The second check is an empirical Darwiche–Pearl audit. After each revision we measure the fraction of propositions unrelated to the input whose answers changed. This rate should be near zero for our method and visibly nonzero for the full-history base model.

The third check is calibration. If image-only firmness does not separate correct from hallucinated propositions on POPE's adversarial split (AUROC well above chance), the whole design rests on sand. We would then substitute probe-based firmness or stop.