acceptodds
Under review as a conference paper at ICLR 2027

Read Less, Recall More in Document VLMs

Abstract

Document vision-language models (VLMs) fine-tuned on forms see sensitive fields both as pixels and as supervised text, yet whether a field's legibility decides what they memorise is untested: resolution studies change the whole page and its visual-token budget. We introduce Visual Canaries, synthetic identifier–secret documents with matched no-canary controls, and run two interventions on 3 released VLMs: a rendering ladder that changes only the secret's glyphs at a fixed target, exposure and token count, and a fixed-weight test that changes only the inference context. On Qwen2-VL-2B-Instruct, lowering the control's transcription accuracy from 1.00 to 0.11 raises the paired memorisation gap from +0.24 to +3.63 [+3.00, +4.26] bits on a 96-distractor rank scale (+13.25 bits on a fitted tail); a corrupted field with the largest glyphs tracks legibility rather than glyph size, the ordering holds on all 3 backbones, and duplication amplifies it by +3.69 bits per doubling. With weights fixed, restoring the page paired with a secret during training recovers +3.18 [+1.90, +4.46] bits although that page holds none of the secret's pixels; a caller with the identifier, the paired page, the caption format and assistant-prefill capability then samples the secret for 96.0% of Qwen and 90.0% of SmolVLM canaries at one query, and for none without the page. Memorisation risk in document fine-tuning thus depends on what the model can read and on the context an auditor supplies: audits must probe with training-like pages, and redaction must cover both pixels and targets.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
memorisation, document vision-language models, privacy auditing, canaries, fine-tuning, training data extraction, legibility

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zqS4OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how field legibility and inference context affect memorization in document VLMs. The authors introduce Visual Canaries, synthetic documents containing identifier secret pairs, together with CMMG, a paired metric that compares fine-tuned models against matched no canary controls. Controlled experiments on three VLM backbones show that reducing a field’s legibility increases detectable memorization when the secret is part of the supervised training target. With model weights held fixed, showing the model the page a secret was paired with during training can substantially improve its retrieval, even when that page does not contain the secret itself. These results expose the limits of text only memorization audits and underscore the role of inference context.

Critical strengths
  1. The rendering intervention changes only the secret field while holding the training target, exposure, page geometry, and visual token count fixed. Combined with the matched controls and the corruption, duplication, and cross backbone experiments, this makes a convincing case that field level visual degradation affects detectable memorization.

  2. The fixed weight experiments were the most interesting part. An association that text only probing completely misses becomes recoverable once the model sees the page it was paired with during training. The matched controls, alternative contexts, and generation based extraction make this more convincing than likelihood scores alone.

Critical weaknesses
  1. The experiments do not cleanly separate memorization from retrieval accessibility. The second experiment shows that retrieval varies sharply with inference context under fixed weights, so the higher CMMG under reduced legibility may reflect easier access to a stored association rather than more storage. The remaining gap between the corruption and font size conditions at similar transcription accuracy also suggests that legibility alone does not explain the effect.

  2. The practical relevance of the context dependent extraction is limited. The strongest results depend on an artificial pairing with another canary's document, knowledge of the exact page and caption format, and assistant prefill access, while ordinary user turn prompts work much less well. Since removing the donor page's own secret eliminates retrieval, the model may have learned a narrow identifier keyed association rather than general context dependent recall, and the few training seeds further limit the precision of the reported rates.

Questions for authors
  1. How can the authors distinguish increased information storage from increased accessibility under the chosen readout? Would the legibility dependent ordering persist across a broader set of inference contexts and prompt formats, rather than primarily the training style text prefix?

  2. Since removing the donor page's own secret eliminates retrieval, could the authors clarify whether this high entropy field acts as the effective retrieval key? Would the same context dependent recovery occur when the paired document contains no other unique secret?

  3. How well does the page conditioned retrieval effect generalize to more realistic document fine-tuning, where the paired page is not deliberately mismatched and an auditor may not know the exact training page association?

Reviewer_uyGNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a fundamental question in document AI: when a model is trained on form-like documents, why does it memorize private field values rather than merely reading them from the page at inference time? To investigate, the authors construct synthetic documents containing fabricated identifiers and secret codes, and for every model they train, they additionally train a twin model that is never exposed to these documents, enabling controlled comparison of the two. Their findings reveal a consistent pattern: the more difficult a field is to read from the page, the more strongly the model memorizes it; since the model cannot reliably extract the value at answer time, it internalizes it during training instead. Moreover, presenting the model with the page that immediately followed the secret during training induces it to reproduce the secret, even though that page itself does not contain the secret value — evidence of contextual (non-target) memorization. Conversely, fields that appear only in the document image and never in the text the model is required to produce exhibit no memorization whatsoever.

Critical strengths
  1. The experiment is clean and fair. The setups change only one thing at a time: they shrink or blur the secret field while keeping the page, the answer text, and the number of exposures exactly the same. That makes the comparison fair, which earlier studies on image resolution could not achieve.

  2. The idea is interesting. Each trained model is compared with a twin that never saw the secrets, so the gap is not just "this model got better at reading forms". The authors also show that using only text to test for memory can fail, which justifies this careful setup.

Critical weaknesses
  1. A blur-only field at 0.39 readability is memorised far more than a small-font field at the same readability. It seems that legibility is not the true driver.

  2. The evidence base is narrow. All fine-tuning uses 256M–3B models, mostly LoRA, on 10 synthetic single-field canaries with one 10-digit secret format. The conclusions talk about document fine-tuning in general, so at least one larger backbone and one ladder trained on real page layouts would make the claims credible.

  3. The second result appears only in an odd setup. The "page brings the secret back" effect is shown when a page is paired with another document's secret. In the normal case, where the page and the secret belong together, the page brings nothing back — so the advice may not apply to real training data.

  4. The explanation is not proven, and the comparison is not fully matched. The story "models store what they cannot read" is supported only by correlation, not by a direct test. Also, the twin model never sees the fake pages at all, so part of the gap may come simply from seeing those pages.

Questions for authors

See the weakness.

Reviewer_z4zyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Main problem dealt by the paper: When a VLM is fine-tuned on documents containing sensitive information, what determines whether that information gets memorized vs simply read from the document? The paper investigate what gets memorized during training + retrievable during inference. They run controlled auditing experiments towards this multimodal memorization.

The main findings from my read are:

  1. Less legible fields lead to more detectable memorization
  2. Memorized information can be hidden until the right context appears

I think these work is pretty inline with my understanding of long-tail memorization phenomena and generally sound. The main implications are that currently memorization might be underestimated and that half measures like redaction might invite more memorization than deter it.

Critical strengths

The paper shows that making sensitive fields harder to read can actually increase their memorization when they are also present in the supervised target, which is an interesting and counterintuitive finding especially for document VLMs where blurring might otherwise seem like a reasonable privacy measure.

The paper show that even when a secret cannot be retrieved through text-only prompting, it can become recoverable when the model is given the document associated with it during training, even when the document itself does not contain the secret, which has important implications for how multimodal memorization should be audited.

I like that the authors use matched controls and several ablations around legibility, corruption, training objectives, and model backbones, which helps establish that memorization depends on more than just how many times the model sees a particular piece of information.

Critical weaknesses

The proposed reading vs memorization mechanism is not sufficiently isolated. The authors argue that models tend to memorize more when they cannot read the information from the input, but reducing legibility also changes the initial target loss, gradients, and the visual input distribution, so it's not entirely clear what is actually driving this effect. Also, Table 12 shows that different corruption methods can lead to quite different memorization even at similar OCR accuracy. I think one useful experiment would be to keep the corrupted image and training setup fixed, but additionally provide either the correct OCR transcription or a shuffled transcription during training. If memorization decreases specifically when the correct transcription is available, this would give stronger evidence for the authors' proposed explanation rather than just the effect of corruption.

Limited generalizability of the experimental setup. Most experiments are done on randomly generated 10-digit secrets, repeated 24 times per canary, and relatively small VLMs (256M–3B) with mostly LoRA fine-tuning. it's not clear how much of these findings would generalize to more realistic document understanding settings where sensitive information has some semantic structure, correlations with other fields, and different exposure frequencies. I think a useful extension would be to have something like a TOFU-style [1] synthetic dataset with realistic identities, profiles, and associated documents, and vary exposure frequencies (say 1, 5, 24 repetitions). This would help establish whether the legibility vs memorization relationship actually holds beyond these fairly artificial random secrets.

[1] TOFU: A Task of Fictitious Unlearning for LLMs, Maini et al.

Questions for authors

See weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.