acceptodds
Under review as a conference paper at ICLR 2027

Uncommitted Drafts Track Words, Not Knowledge: Evasion and Representation Monitoring in Diffusion LMs

Abstract

A masked diffusion language model holds, at every denoising step, a distribution over its entire uncommitted answer—a whole-canvas draft that recent work reads as an intrinsic monitorability affordance with no autoregressive counterpart. We ask what a monitor of that draft can see, and whether it survives training against it. We build the first secret-keeping, sandbagging, and belief model organisms in diffusion LMs, on Qwen3-0.6B-diffusion and LLaDA-8B. The draft tracks the text a model is writing—including words it will suppress—not the knowledge that shapes that text. Where the secret is the text's topic, a reader of masked-position mass names a word the model emits in under 12% of runs at 0.72–0.92 (chance 0.05), and names it by denoising step 2, while a black-box reader of the committed text needs 8 to 24 steps. Where text and knowledge diverge, the draft fails with no adversary at all: a sandbagger's draft holds the wrong answer it will emit (0.22–0.52) and the correct one only at chance, and a model that acts on an unstated belief in 70% of its answers, never writing the word, trips a word-level draft monitor in just 14% of them at a 5% false-alarm rate. Against a decoy arm given identical training, a word penalty closes the secret's draft signal by 0.52–0.71 while the emitted hints stay informative, and halves the belief organism's word monitor without changing its behaviour. The knowledge is not lost: it stays legible in late-layer representations—where penalty and decoy are indistinguishable at every layer—and deleting a late-layer direction removes the behaviour (0.97 → 0.12, against 0.92 for a random direction). And the concept lives upstream of the token: in a companion organism that states the belief, blocking every token of the word at decoding—identical weights, identical seeds—leaves the dog-specific ideas untouched (0.667 against 0.684, difference -0.02 [-0.14, +0.11]) against floors of 0.009 and 0.000. Draft monitors watch words; the knowledge is in the representation. Monitor concepts, not words.

Primary area:
interpretability and explainable AI
Keywords:
diffusion language models, AI safety, monitorability, model organisms, interpretability, chain-of-thought monitoring, sandbagging, obfuscation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_aESWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies monitoring of discrete diffusion language models using models trained to hide secrets, underperform and hold unstated beliefs. The main focus is on how whole-answer drafts provided by discrete diffusion models at each decoding step might facilitate monitoring. The authors show that while the drafts facilitate early detection, they primarily track intended words and not withheld knowledge, since targeted training to penalize key words weaken detection without fully removing informative hints or changing model behavior. Further hidden state measurements and causal interventions show that the underlying concepts still affect model behavior even when related words are suppressed. This motivates monitoring internal representations rather than token probabilities. I found the text very hard to follow, so I apologize if I have misrepresented the paper in this summary somehow.

Critical strengths
  • Monitoring of discrete diffusion models is a relevant and timely topic that has not received much attention in the literature.
  • The reported findings that the whole-answer drafts provided by discrete diffusion models enable earlier detection but do not seem to provide more faithful or robust monitoring is significant and could motivate further research.
Critical weaknesses
  • The paper’s presentation could be substantially clearer. Several paragraphs combine motivation, experimental details, results, and discussion, making it difficult to follow the progression of the argument. I encourage the authors to separate these elements and establish a clearer narrative, beginning with the motivation and experimental setup before presenting the results and their interpretation. The abstract suffers from a similar problem; it is dense and full of numbers out of context, which prevents the main message from shining through.
  • I am not fully convinced about the novelty of the paper.
    • The paper seems to apply the same experimental setup proposed in [1] to discrete diffusion language models. The methodological novelty seems limited to the reading of the diffusion-specific draft readout.
    • The main finding that internal representations are better monitoring targets than token probabilities is not particularly surprising. Prior work has already established that activation-based readouts can recover information not reflected on the model’s output [2, 3].

References

[1] Cywiński, Bartosz, et al. "Towards eliciting latent knowledge from LLMs with mechanistic interpretability." arXiv e-prints (2025): arXiv-2505.

[2] Burns, Collin, et al. "Discovering latent knowledge in language models without supervision." arXiv preprint arXiv:2212.03827 (2022).

[3] Azaria, Amos, and Tom Mitchell. "The internal state of an LLM knows when it’s lying." Findings of the Association for Computational Linguistics: EMNLP 2023. 2023.

Questions for authors
  1. The experimental setup was not entirely clear from the paper. Could the authors explain it in detail, including the needed innovations or adaptation required for the discrete diffusion setting?
  2. How does this work inform the design of future monitors besides encouraging the usage of the model's latent representations?
Reviewer_zmTXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether the undecided masked canvas at each diffusion LM step can serve as a trustworthy monitoring channel. The authors port the existing Taboo secret-keeping paradigm to DLM: the model describes a secret word but typically does not output it directly. By aggregating candidate-word probability mass across all masked positions, the authors find that this reader can sometimes recover the secret earlier than reading only the currently committed text. Furthermore, directly suppressing the secret word's draft probability mass substantially reduces this reader while the final hints remain largely usable for recovering the secret.

The paper also uses sandbagging and dog-belief organisms to argue that the draft is closer to the text the model is about to write than to the latent knowledge supporting behavior. Finally, the authors read out a dog-related signal with a late-layer linear ruler and argue via direction deletion that it is related to dog-related behavior.

Critical strengths
  • The question is valuable: how to monitor a model's true computation in AR models is a fairly active topic, and studying this question on a new language-modeling architecture, DLM, is worthwhile.

  • The paper conducts experiments across a fairly wide range of aspects to stress-test the feasibility of canvas monitoring in DLM; the experiments are relatively thorough.

Critical weaknesses
  • The core problem is that the paper does not sufficiently establish what is non-trivially new in these phenomena relative to AR. For logits close to the output end, reading "what the model is about to write" rather than more abstract latent knowledge is a very natural expectation in AR; directly suppressing the probability of a certain token while not changing higher-level semantics or behavior is also consistent with existing monitor-evasion / latent-representation work. As it stands, the paper looks more like porting existing Taboo, secret elicitation, and monitor-evasion phenomena to DLM than demonstrating that DLM is fundamentally different from AR in monitorability.

  • The paper's final claim is not yet complete: it shows that a word-level draft-logit monitor is insufficient and preliminarily demonstrates that a linearly readable pet-related signal exists internally; but it does not form a clear alternative. The paper does not give a comparison or an actionable recommendation for how DLM's inner computation should actually be monitored.

  • All organisms are highly synthetic toy settings: one secret word, four-choice sandbagging, dog belief. They are suitable for diagnosing readouts, but they are far from dangerous intent, agentic misbehavior, jailbreak, or covert planning in real safety scenarios. The paper has not yet shown whether the observed draft signal, warning window, and evasion patterns transfer to more realistic, open-ended safety tasks.

Questions for authors
  1. Please clarify the paper's core claim about DLM relative to AR. For example, is it that DLM can obtain information earlier, obtain future-text predictions at lower computational cost, or observe computation that is originally invisible to AR? Can you add a matched comparison with AR, for example comparing a DLM masked-canvas reader with an AR next-token logit reader, and report results after matching computational budget? This is important for judging whether the paper is "rediscovering phenomena already known in AR" or discovering monitorability properties specific to DLM.

  2. If the conclusion is that word-level draft logits mainly monitor text rather than concepts, how do the authors recommend that practical systems monitor DLM's computation?

  3. Can the conclusion be validated in at least one setting closer to practical risk, for example covert unsafe intent, agentic reward hacking, jailbreak planning, or more complex sandbagging?

Reviewer_ZysXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks whether a monitor of a masked diffusion LM's uncommitted draft (the distribution over still-masked positions at each denoising step) can detect hidden knowledge, and whether the signal survives training against it. The authors train LoRA model organisms on Qwen3-0.6B-diffusion (an autoregressive checkpoint converted to masked diffusion) and LLaDA-8B in NF4: Taboo secret-word organisms, a sandbagger, and an organism that acts on the belief that the user owns a dog without writing the word. A reader of masked-position mass names the Taboo secret at 0.72–0.92, earlier than a reader of committed text. A penalty on the secret's masked-position mass lowers this reader by 0.52–0.71 relative to a matched decoy arm. For the sandbagger and the belief organism the draft fails without any pressure, while a late-layer direction fitted on word-free stimuli reads the belief, and deleting it removes the behaviour (0.97 to 0.12).

Critical strengths
  1. The decoy arm matches data, compute, optimizer, LoRA rank, seed and λ\lambda, and differs only in the target word. Each concealment claim is paired with an informativeness check (TOST equivalence). The causal tests use a random direction of equal norm and a decode-time word ban with identical weights. Section 8 reports six claims that the authors withdrew after a matched control, which is more self-correction than most papers show.
Critical weaknesses
  1. The writing shows unedited LLM drafting, and it hides the results. The AI-use statement says LLMs were used "to draft and edit prose", and it shows:

    • Abstract. It is three paragraphs and 324 words, packed with about twenty numbers and intervals. The ICLR template limits the abstract to one paragraph. It ends in slogans ("Draft monitors watch words; the knowledge is in the representation. Monitor concepts, not words.").
    • Repetition. The same claims and numbers repeat across the abstract, introduction, Sections 4–7 and the conclusion. "Tracks the text" appears four times, and 0.97 → 0.12 appears four times. The introduction restates the abstract almost line by line.
    • Headings and prose. Section titles are aphorisms ("Take the word away: the ideas stay", "Word pressure closes what the draft sees"), not descriptions of what each section tests.
    • Unedited errors. Section 7.1 says "at or below chance 0.25 (chance 0.25)". Section 7.4 (p. 8) starts a sentence with "Third, ..." with no first or second point before it. Section 6.3 is titled "Where the trained refusal lives", but the organisms conceal a word; they do not refuse.
    • Limitations (Section 9). It is a list of bold run-in labels ("Toy secret, direct pressure.", "Bounded in four directions.", "Scale, substrate, sampler, prompts, and weak instruments.") that compresses a dozen caveats into fragments. Most of them were already stated in Sections 6–8. The section reads as a generated checklist. It does not say which limitation threatens which claim, or how much.
    • A broken instrument stays in. The logit lens "reads 0.000 in every arm including the OPEN positive control"; the paper calls it "a broken instrument, not a null" but still reports it.

    Together these make a short empirical story hard to follow, and they make it hard to tell the established results from the hedged ones.

  2. The scientific claims rest on a narrow setup and on results that hold by construction.

    • Scale. The organisms are a 0.6B converted autoregressive model and an 8B model in NF4. One sampler at 32 steps and 20 hint prompts from one template family are used throughout. The main result has 3 seeds, and each dose point has 1.
    • By construction. The penalty lowers exactly the quantity the reader sums, which the authors admit.
    • Readers and transfer. The Taboo readers pick among 20 known candidates, so they measure oracle identifiability, not detection at a fixed false-positive rate. Transfer of evasion to a held-out monitor fails on the 8B model.
    • Belief results. They rest on a judge that agrees with human labellers at κ=0.27\kappa = 0.27, and on a direction that is pet-level, not dog-level: deleting the cat direction also cuts dog behaviour to 0.16–0.30. There is one sandbagger and one belief organism, both trained in.
    • Novelty. The "first of their kind" claim rests on a footnoted arXiv keyword search.

    The general title claim ("drafts track words, not knowledge") is much broader than this evidence

Questions for authors
  1. Does the timing advantage (draft passes 0.5 at step 2, committed text at steps 8–24) hold for a detection-style reader with no candidate list, at a fixed false-positive rate? If not, the "warning window" has no practical use.
  2. Can you show a dog-specific direction (contrasted against the cat organism) whose deletion removes dog behaviour but not cat behaviour? Without it, "the knowledge is in the representation" is supported only at the level of "pets".
  3. How sensitive are the results to the remasking schedule and the number of denoising steps? Section 8 shows that the commit schedule affects what leaks.
  4. With more seeds on LLaDA-8B, does held-out transfer fail, or is the null underpowered?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.