acceptodds
Under review as a conference paper at ICLR 2027

Hard to Unsee: Source Addressability After Multimodal Integration

Abstract

A system that integrates several sources often learns which one to trust only afterwards, when the original evidence can no longer be re-read. We ask what a vision-language model can still do with a late, source-conditioned instruction once only an intermediate representation remains. An image and one or two texts disagree about one attribute; an inserted summary records or omits which source said what; access to the evidence is kept or removed by a verified attention mask; only then is a source queried by name or declared unreliable. After a source-neutral summary, Qwen3-VL-8B and Pixtral-12B do not retrieve what a named non-default source reported above a shortcut ceiling, and a late reliability statement is followed only when it agrees with a persistent, image-favoring default. Carrying arbitrary source handles through integration, without binding any value to them, restores most named-source retrieval; handles introduced after the bottleneck do not; answers follow a handle when its source assignment is permuted; the effect survives eviction of the evidence from the cache; and blocking attention to the carried handle at question time removes it while a matched sham does not. The hierarchy replicates in Pixtral-12B and on controlled natural images, is weaker for visual than for text targets, and reliability-conditioned selection lags retrieval only in Qwen3-VL-8B. In these settings, later source-conditioned access depends strongly on a referent carried through integration, while explicit source-value binding provides additional support for reliability-conditioned selection.

Primary area:
interpretability and explainable AI
Keywords:
vision-language models, multimodal reasoning, source attribution, source addressability, source monitoring, multimodal conflict, provenance, causal interventions, mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_bZjmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies whether vision-language models can retrieve and use information from a specified source after conflicting evidence has been integrated into an intermediate representation and direct access to the original evidence has been removed. The authors vary whether an inserted summary omits source identities, lists source names or arbitrary handles without explicitly pairing them with values, or records explicit source–value associations. They distinguish named-source retrieval from direct source selection and reliability-conditioned selection using controlled tasks involving counts, colors, and chart values. In Qwen3-VL-8B and Pixtral-12B, source-neutral summaries lead to poor retrieval of non-default sources, while carrying source handles through integration substantially improves retrieval.

Critical strengths
  • The paper introduces a setting that uses an intermediate representation and removes access to the original evidence before the source-conditioned question arrives.
  • The paper reports prerequisite failures in Qwen3-VL-8B and Pixtral-12B.
Critical weaknesses
  • I strongly suspect that the paper was generated using AI. The appendix appears to consist entirely of raw LLM output without human review.
  • Why does the paper use only eight pages for the main text but more than 20 pages for the appendix?
  • The benchmarks cover only simple questions about counts, colors, and chart values, and the experiments evaluate only two vision-language models.
  • The experimental setting, protocol, and dataset are not clearly described in the current text.
  • Some parts of the paper and appendix are unreadable.
  • The paper’s contribution is entirely unclear.
Questions for authors
  • Can the authors understand their own appendix?
1 comment

The contribution. When a vision-language model has integrated conflicting sources into a summary and direct access to the evidence is then removed, it keeps a default answer but can no longer retrieve what a named non-default source reported. For Qwen3-VL-8B with a source-neutral summary, retrieval is 0.16, below a 0.50 shortcut ceiling. A failure to follow a later "that source is unreliable" can therefore be an access failure rather than a trust failure. Arbitrary source handles carried in the summary, with no value attached to them, restore most retrieval (0.80). Blocking the queried handle, but not an unqueried one, removes it in Qwen3-VL-8B and Gemma-3-12B-it (specificity +0.55 and +0.64 for retrieval).

What is new relative to prior work. Source-monitoring and binding studies examine attribution and label binding while the evidence remains readable. Other work shows that image information persists after image-attention knockout, without asking whether a named source can still be addressed. Late-correction studies after compression are text-only. Our design combines four elements:

  • conflicting named sources;
  • verified removal of access after integration;
  • an intermediate with or without source referents;
  • a later named-source query.

This combination is what exposes the access failure and the dependence on carried handles. We do not claim a new mechanism.

The protocol, in one example.

  • Turn 1 presents three sources: Source B, a text ("There are 3 green triangles."); Source A, a text reporting 8; and Source C, an image showing 6.
  • Turn 2 is an inserted summary. It is either neutral ("Reported counts of green triangles: 8, 6 and 3.") or carries handles ("Sources consulted: Source B, Source A and Source C. Reported counts of green triangles: 8, 6 and 3."). The handle version names the sources but pairs none with a value.
  • Turn 3 asks: "According to Source B, how many green triangles are there?"
  • Masking. An attention mask prevents later positions from reading turn 1. We verify the mask in every run: once the evidence and the summary are both blocked, swapping in another item's evidence leaves the outputs exactly unchanged.
  • Scoring. Only sources whose value differs from the model's default answer are scored.

Corrections. Re-checking the submission against the raw records found errors, all corrected in the revision:

  • the model-written-summary results had been cut, and Section 7 wrongly said they were not tested (now Appendix J);
  • a Pixtral value was in the wrong slot (0.99 → 0.78);
  • a late-label control never told the model what the labels referred to (withdrawn and replaced by a grounded test);
  • the sham control was misdescribed;
  • a decision level in the sensor-corruption control was chosen after outputs were seen (the pre-set level is now reported, and the result is "not resolved");
  • one analysis dropped 33 items contrary to its design (restored).

We report unfavourable pre-set results exactly as frozen. Pixtral's ablation prerequisite failed, its layout re-test is indeterminate, we make no claim under the combined reliability cue, and model-written handle summaries failed their generation gate.

The appendix and AI use. You are right that the submitted appendix contained raw program output and editing damage. LLMs were used extensively for planning, code, analysis and drafting. We have revised the AI-use statement accordingly and take responsibility for the content. In the revised appendix, raw output is replaced with prose and captioned tables, corrupted experiment names are restored, and secondary controls move out of the main text.

Scope and length. The tasks are deliberately simple and single-attribute, with one natural-image check (100 VQAv2 photographs, colour, one model). The failure and the rescue hold in two models; two further models failed the prerequisite even with the evidence visible. The specificity result holds in Qwen and Gemma. The main text now presents one sequence, with secondary controls in the appendix:

  1. the addressability failure;
  2. the handle rescue;
  3. queried-versus-unqueried specificity;
  4. grounded late naming.
Reviewer_nupROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether a multimodal model can retrieve and use source-specific information in response to instructions received after integration. The authors construct conflicting image and text inputs, insert intermediate summaries, and block direct access to the original evidence through attention masking or KV-cache eviction. The experiments compare summaries that omit source identifiers, retain identifiers without explicitly binding them to values, or explicitly preserve source–value mappings. The main finding is that retaining source identifiers across integration substantially restores subsequent source-conditioned retrieval in the studied settings. The authors investigate this effect through label reassignment, cache eviction, and attention ablation at source-identifier positions, with results from two principal models and a small natural-image validation.

Critical strengths
  1. The paper clearly distinguishes source retrieval from reliability-conditioned answer selection. Failure to follow a reliability instruction need not imply a failure of reliability updating itself; the model may be unable to retrieve the specified source’s content. Separately testing source queries, direct source-use instructions, and reliability-conditioned selection helps avoid conflating these different failures.
  2. The central finding is interesting and supported by complementary controls. Retaining source identifiers without explicitly stating source–value mappings substantially restores retrieval, which is the paper’s most compelling result. Arbitrary-label reassignment, physical cache eviction, and attention ablation at the queried identifier provide stronger support than a prompt comparison alone.
Critical weaknesses
  1. The scope of the central finding remains insufficiently characterized beyond templated, single-attribute settings. The controlled design is useful for isolating source addressability, but the main experiments use inserted summaries and conflicts over a single attribute. The natural-image validation changes the visual inputs while retaining templated text sources and summaries, and remains limited to color questions on one model. It therefore leaves open whether the identifier benefit persists when summaries are model-generated or when each source contains multiple facts that must be distinguished. These gaps limit my assessment of the finding’s broader significance.
  2. The source specificity of the identifier-ablation effect is not fully established. Label reassignment and cache eviction provide useful evidence that the retained identifiers support source-conditioned retrieval. The attention-ablation experiment further shows a substantial performance drop when access to the queried identifier is blocked. However, its matched sham blocks adjacent function tokens rather than another source identifier. A comparison against ablating an unqueried identifier would help determine whether the impairment selectively affects access to the queried source or more broadly disrupts the summary’s source structure. The current intervention supports the importance of the queried positions, but leaves this more specific causal interpretation unresolved.
Questions for authors
  1. Does the benefit of source identifiers replicate with model-generated summaries? The main results rely on inserted, templated summaries, leaving unclear whether the observed benefit depends on their specific organization. Could the authors report the identifier benefit with model-generated summaries, together with summary compliance rates? This comparison would help establish whether the central finding extends beyond the templated setting.
  2. Is the effect of identifier ablation specific to the queried source? The current experiments compare ablating the queried source identifier with ablating adjacent ordinary tokens. Could the authors also compare ablation of the queried identifier against an unqueried source identifier, and examine whether the effect changes with the queried source? This would help distinguish disruption of access to a particular source from a more general disruption of the summary’s source structure. Please also distinguish incorrect source selection from unparseable responses, since identifier ablation increases the latter in the reported results.
  3. Please reconcile two reporting inconsistencies. Page 8 states that model-written summaries were not tested, whereas page 14 describes their experimental procedure and page 16 lists 720 model-written summaries. Please reconcile these statements and clarify the status of these experiments. Additionally, page 4 describes the two BN directional accuracies for the Pixtral two-source replication anchor as 0.50 and 0.99, whereas Appendix F.3 explicitly reports BN caption-unreliable accuracy as 0.78. Please verify the conditions and sample populations associated with these values and correct or explain the discrepancy.
  4. Does the identifier benefit persist when each source reports multiple attributes? The current tasks primarily involve a single conflicting attribute. Could the authors add a small controlled extension in which each source reports two attributes, such as object count and color, with conflicting values across sources? The final question would ask for a specific attribute from a named source—for example, “According to Source B, what color is the object?” This would test whether retained identifiers help retrieve the correct source–attribute association when multiple facts must be preserved.
1 comment

Thank you for the precise questions. The central one, whether the ablation effect is specific to the queried source, now has a direct answer under frozen protocols in two model families.

1. Is the effect specific to the queried source? (W2, Q2) We ran the comparison you suggested under a protocol fixed before any output: the 90 stable-handle scenes, evidence blocked, 180 non-default pairs per cue. Separate conditions block the queried handle, each unqueried handle, or a sham. Specificity S is the accuracy drop from blocking the queried handle minus the drop from blocking an unqueried one.

ModelS, retrievalS, use
Qwen3-VL-8B+0.55 [+0.48, +0.62]+0.62 [+0.56, +0.69]
Gemma-3-12B-it+0.64 [+0.56, +0.71]+0.53 [+0.46, +0.61]
  • Qwen: blocking the queried handle lowers retrieval by 0.57, to 0.22. Blocking an unqueried one lowers it by 0.025 [−0.006, +0.058].
  • Gemma: it was added after Pixtral failed its prerequisite, by criteria fixed before any Gemma output, and it passed its own prerequisite (0.81).
  • Pixtral: it shows the same pattern (S = +0.44 [+0.36, +0.52] for retrieval, +0.39 [+0.29, +0.49] for use), but it fails the prerequisite (0.68 < 0.70), so we report it only descriptively.

Does the effect change with the queried source? By list position (exploratory, Qwen), the drop appears at every position. By modality, the ablation cannot tell: 172 of the 180 targets are text sources, so specificity for image-source targets is untested. Without ablation, handles also rescue retrieval of image sources (+0.40 [+0.29, +0.51]), less than of text sources (difference +0.175 [+0.050, +0.300], descriptive; now including 33 items the submitted code had dropped).

Under the queried-handle block, Qwen mostly answers with the default source's value (0.48); none is unparseable (the submitted "unparseable 0.21" combined off-list values and "no such source" replies). This is behavioral necessity of the queried handle, not a mechanism, and we make no claim under the combined reliability cue.

2. Model-generated summaries and the 720 summaries (W1, Q1, Q3a). The page-8 sentence was wrong: Qwen3-VL-8B wrote the 720 summaries in C.4 (an explicit and a neutral one per two-source item). Compliance was 227/360 (0.63) neutral and 95/360 (0.26) explicit. With the image declared unreliable:

  • Neutral: 0.25 [0.20, 0.31], against 0.11 templated on the same items. The image value still dominates.
  • Explicit: 0.28 [0.20, 0.38], against 0.99 templated. This is confounded, because 93 of the 95 had already called the caption incorrect, compliance selected the items, and no text-only reader was run.
  • Paired comparison: the pre-specified paired comparison is +0.04 [−0.10, +0.19], which is inconclusive.

These results are restored in Appendix J. Because these summaries name "the image" and "the caption", they cannot test identifiers; when we asked the model to write handle summaries for the 90 three-source scenes, it produced none in the required form (0/90), so the identifier benefit with model-written summaries is untested. The revised abstract and Limitations restrict it to inserted summaries.

3. The Pixtral values (Q3b), and a preprocessing discrepancy. Re-derived from the raw records (same run, same 90 items):

ValueConditionCount
0.50Evidence blocked, neutral summary, image declared unreliable45/90
0.78Same condition, caption declared unreliable70/90
0.99Evidence accessible, explicit summary, image declared unreliable89/90

The text should have read "0.50 and 0.78" and is corrected; the interaction (+0.36 [+0.24, +0.47]) is unaffected. The default-agreement account therefore holds fully for Qwen and only partly for Pixtral, as the abstract now says.

While checking, we found that our Pixtral runs fed images without the row-break and end-of-image tokens its own processor inserts. A frozen native-layout re-test gives a handle rescue of +0.43 [+0.32, +0.53] vs +0.46 [+0.38, +0.54] originally; the difference, −0.04 [−0.11, +0.04], excludes a large effect but misses our ±0.10 equivalence band (indeterminate; disclosed in §5.10).

4. Multiple attributes and scope (Q4, W1). We did not test multiple attributes. Every task involves one conflicting attribute, and we make no claim beyond that; the Limitations say so. The natural-image check changes only the images (100 VQAv2 photographs, colour, one model). Handle order and identity were varied within one sentence template: with a scrambled list, retrieval is 0.84 vs 0.80; with nonce names, 0.78 vs 0.80.

Comments on the paper

1 comment

Thank you for the careful reading and for each of your three questions, we have corrected the paper accordingly.

1. Late handles (W1, Q1): we withdraw the submitted interpretation. You are right: no turn told the model what "Source P/Q/R" referred to, so the submitted late-label condition could not test integration. Rerun with the evidence accessible (Qwen3-VL-8B; 83 items, 167 non-default pairs), these labels give 0.36 [0.29, 0.44], the level of a text-only reader of the same summary (0.35), and blocking the evidence changes nothing (−0.01 [−0.07, +0.05]). The submitted 0.37 therefore reflects unresolvable names, and we withdraw that contrast for both models.

We replaced it with a grounded test, run under a protocol fixed before any output. The final question begins with an explicit mapping ("Source P refers to Source B, …"), identical across two conditions that differ only in the summary. The summary lists either P/Q/R, so the mapped-to handle is not carried, or Source A/B/C, so it is carried.

  • With the evidence readable, both conditions work (0.92 and 0.97), so the mapping is understood.
  • With the evidence blocked, retrieval is 0.40 [0.33, 0.48] when the target handle is not carried and 0.71 [0.64, 0.78] when it is: a gap of +0.31 [+0.23, +0.39].

A late name restores retrieval after the bottleneck only when it points to a handle carried through integration. The test covers Qwen, retrieval and one mapping wording; under the combined reliability cue the difference is not detectable (+0.04 [−0.04, +0.12]). See revised §5.8.

2. Ablating an unqueried handle (W2, Q2). We ran this comparison directly (protocol frozen before any output; the 90 stable-handle scenes; 180 non-default pairs per cue). The submitted sham was not that control: it blocked the two tokens before the queried handle, which in two-thirds of trials included another handle's letter; the paper now describes it correctly.

Blocking the queried handle lowers Qwen3-VL-8B retrieval by 0.57 [0.49, 0.64], to 0.22; blocking an unqueried handle lowers it by 0.025 [−0.006, +0.058]. The difference, our specificity measure S, is +0.55 [+0.48, +0.62] for retrieval and +0.62 [+0.56, +0.69] for use.

  • Gemma-3-12B-it, added in the discussion period by criteria fixed before any of its outputs, shows the same pattern: S = +0.64 [+0.56, +0.71] and +0.53 [+0.46, +0.61].
  • Pixtral-12B also shows it (S = +0.44 [+0.36, +0.52]), but it fails our capability prerequisite (0.68 < 0.70), so we report it only descriptively.

Two limits apply. 172 of the 180 scored targets are text sources, so specificity for image-source targets is untested. We also make no claim under the combined reliability cue. The result shows that access to the queried handle is behaviorally necessary; it does not identify a mechanism. The paper reports S separately from the submitted handle-minus-sham effect (+0.56), which is a different quantity (§5.6–5.7, Table 2).

3. Model-written summaries (W3, Q3): they were run, and the Section 7 sentence was wrong. In the 360-item two-source experiment, Qwen3-VL-8B wrote two summaries for every item: an explicit one (what the image and the caption each report) and a neutral one (both values, no attribution). Compliance, judged from the text before any answer was generated, was 227/360 (0.63) for neutral summaries and 95/360 (0.26) for explicit ones.

With the evidence blocked and the image declared unreliable:

  • Neutral: 0.25 [0.20, 0.31] (56/227), against 0.11 templated on the same items and 0.48 for a text-only reader; the image value still dominates.
  • Explicit: 0.28 [0.20, 0.38] (27/95), against 0.99 templated on the same items. This is confounded: 93 of the 95 had already called the caption incorrect, compliance selected the items, and no text-only reader was run.
  • The pre-specified paired comparison (n = 69) is +0.04 [−0.10, +0.19]: inconclusive.

These results are restored in Appendix J, and Section 7 is corrected. Because these summaries name "the image" and "the caption", they are not handle summaries. When we asked the model to write handle summaries for the 90 three-source scenes, it produced none in the required form (0/90; in 84/90 it attached values to sources), so the pre-set generation gate failed and the identifier benefit with model-written summaries remains untested, as the revised Limitations state.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.