acceptodds
Under review as a conference paper at ICLR 2027

Does the Denoising Trajectory Reveal the Source of Uncertainty in Diffusion Language Models?

Abstract

For a large language model, the source of uncertainty determines the appropriate response. An ambiguous prompt calls for a clarifying question, whereas missing knowledge calls for retrieval or abstention. In autoregressive generation, the model decides one token at a time. During a single generation we therefore observe only the distribution over the next position, conditioned on a prefix that is already set. Whether the uncertainty stems from many valid answers or from missing knowledge only becomes visible across repeated generations. Diffusion large language models instead generate a distribution over all positions at every denoising step from a single generation. We ask whether this denoising trajectory reveals the source of uncertainty, and make three contributions. First, we construct source-of-uncertainty labels (SOUL), a benchmark of 16,224 prompts from seven corpora and an authored set, labeled according to established definitions of different sources of uncertainty. Second, we show that the denoising trajectory contains source-specific information across four diffusion language models. Statistical signals summarizing the trajectory identify the source of a prompt well above chance, and outperform the same signals taken from tokenizer-matched autoregressive models, by 0.08–0.22 F1 on task underspecification. Third, we demonstrate that a convolutional probe applied to the full position × denoising-time trajectory performs within 0.03 F1 of statistical signals, indicating that the information relevant to the source is largely captured by simple, interpretable properties of the denoising process. Because these distinctions are available from the generation a model already produces, they offer a practical basis for deciding when a system should ask, abstain, or answer.

Primary area:
generative models
Keywords:
Uncertainty quantification, Diffusion language models, Denoising Trajectories

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WKh6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies uncertainty quantification in diffusion LLMs (dLLMs). They create a dataset of questions, each of which falls into one of four different kinds of uncertainty: task underspecification, context underspecification, epistemic, and aleatoric. They find that dLLM diffusion trajectories are distinct across these different kinds of uncertainty, and that simple temporal statistics over these trajectories can predict which kind of uncertainty the model is confronted with, as well or better compared to a CNN-based probe.

Critical strengths
  • I think this is an interesting problem domain, and to my knowledge this work is novel in its dataset and approach
  • The approach is simple but effective, and I find the results of Figure 3 especially compelling
Critical weaknesses
  • Writing and presentation could be improved. My reading is that this paper was partly written/edited with AI, and it shows in a number of places.
  • I'm a little unsure about how well the dataset generalizes to real world tasks, or whether particular categories measure what they claim to. In particular:
    • It seems odd that aleatoric uncertainty equates to questions with multiple valid answers ("Name three sports teams in Boston") instead of questions with one true answer that is impossible to know ("I flipped a coin - did it come up heads or tails?"). I could be convinced that the former is valid, but this feels a bit taken for granted.
    • It seems like task underspecification may be a specific kind of context underspecification
    • What about cases where an input prompt has multiple kinds of uncertainty, or where it's ambiguous from the prompt what kind of uncertainty there is?
  • I would have liked to see this applied to real world tasks as well as the authors' constructed dataset. It would be interesting if this kind of analysis could apply to arbitrary prompts to identify what kind of uncertainty that prompt might be encoding.
  • I also would have liked to see quantitative analysis of temporal dynamics, rather than the qualitative descriptions in section 5.1
Questions for authors
  • What are the "interpretable temporal statistics" that you measured (lines 456-7) - is this just the plots shown in Fig 3?
  • How do you compute the commit step tpt_p? (line 299)
  • Could you provide a more formal definition of what posterior distribution the "posterior cube" represents?
  • What does τ1/2\tau_{1/2} in figure 1 represent?
  • Can your analysis of denoising dynamics help us identify cases with multiple kinds of uncertainty, or where the model is uncertain about what kind of uncertainty there is?
4 comments

Dear WKh6, thank you for your time and the careful review.

W1. We tried to make the AI usage transparent in our AI use statement. All ideas, experiments and texts are our own. Since none of us are native speakers, we then used LLMs to rephrase it for readability. If you could point us to specific passages, or even suggest phrasings we can build on, we are happy to revise them.

W2(a) Interestingly, this question motivated this paper. While reading the literature, we found several competing views on where the uncertainty of an LLM originates. We collected them from a neutral standpoint and asked whether these human distinctions also appear inside a dLLM.

Two real-world examples of ALEATORIC uncertainty help in answering your question: A WildChat user asks to "write a complex chord progression for a pop song". A CLAMBER prompt asks for "some ideas on where to travel for good sun and beaches". Experts would accept multiple answers to each. A random progression does not make a pop song, and not every place has sun and beaches. However, no lookup and no follow-up question reduces this to a single answer, and this is what ALEATORIC captures.

PromptSourceWhat settles it
A coin that has not been flipped yetUNRESOLVEDnothing today
A coin whose outcome only the user has seenCONTEXTasking the user
Write a complex chord progression for a pop songALEATORICnothing ever

Your coin example therefore lands in UNRESOLVED or CONTEXT, and we will add it to Section 3.1.

W2(b). You are right that both sources of uncertainty lack information from the user. We follow Kirchhof et al. (2025) in keeping them apart, as the follow-up question differs. For TASK, the prompt does not include an instruction, as in "It's my mom's birthday." For CONTEXT, the task is clear but a detail is missing, as in "When did the first Harry Potter movie come out?", which depends on the country. The dLLMs also treat them differently and the classifiers confuse the two in only 7% to 12% of cases (Table 15).

W2(c) and Q5. We agree that real prompts might mix sources of uncertainty. SOUL labels each source independently to allow for multiple labels. 73 prompts carry two of them. We now trained one classifier head per source. Appendix D.8 of the revised paper reports the results. On ALEATORIC and UNRESOLVED, these mixed prompts score as high as single source prompts, whereas the CONTEXT head drops. As each head outputs its own probability, a prompt with several sources can receive several high scores. A prompt with an unclear source shows up as moderate scores across all heads.

W3. We share this goal, and it motivated the design of SOUL. We combined corpora from uncertainty quantification research with real user prompts. Trained without WildChat, our classifier still recognizes the sources of WildChat prompts (see table in djDn). We made all code available in the anonymous repository to keep the method transparent. It can be adapted as needed for your own prompts. We realize that deploying the dLLMs is a real effort, and we hope you understand that we cannot offer a hosted version with open-ended costs.

W4 and Q1. In its current form, Section 5.1 undersells the quantitative results, which are hidden in the appendix. We will move the key figures to the main text. The interpretable temporal statistics are the named scalars of Table 6, such as the entropy half life, while Figure 3 only illustrates the underlying curves. Section D.1 reports their AUROC, for example 0.79 for the entropy half life on TASK with Nemotron.

Q2. The commit step of position pp is tp=min⁡{t∣h[t,p]=1}t_p = \min\lbrace t \mid h[t,p] = 1\rbrace, the first step after which the position is unmasked and keeps its token. The commit step also highlights what makes dLLMs attractive for our question. Unlike an autoregressive model, a dLLM keeps a full distribution for each position until it commits, so the path to each decision becomes observable. This path is what the posterior cube (see Q3) stores and what our signals measure.

Q3. The cube holds the denoising posterior P[t,p,v]=pθ(yp=v∣x,zt)\mathcal{P}[t,p,v] = p_\theta(y_p = v \mid x, z_t), where xx denotes the prompt and ztz_t the partially denoised answer that enters step tt. It is exactly the distribution from which the sampler draws. The rates of the sampler are affine in P\mathcal{P} (Holderrieth and Erives, 2026), so every signal in Table 6 is a chain of aggregations of the cube. We will add this definition to Section 4.1 and revise the notation in Table 6.

Q4. Thank you for pointing this out. τ1/2=min⁡{τ∣H(τ)≤12(H(0)+H(1))}\tau_{1/2} = \min\lbrace \tau \mid H(\tau) \le \tfrac{1}{2}(H(0) + H(1))\rbrace denotes the entropy half life, the time at which the averaged entropy has fallen halfway. We will add this definition to the caption and mark τ1/2\tau_{1/2} on the axis.

Thank you again for your review. As the discussion period closes, we wanted to follow up on our rebuttal. We’d welcome any further feedback or remaining concerns, and hope you’ll consider our response in your final assessment.

W1: Fair enough. I will say, that as a native english speaker, in general I'd much rather read text written by a non-native speaker than text written/heavily edited by an LLM.

I do wish there was a good recipe for making writing not sound LLM-y, work like [1, 2] may be a helpful reference, but if there was good advice for this I would give it to 4/5 papers I was assigned to review this cycle. That said, the writing here is the least LLM-y of those 4 papers, but a good example of the AI accent is the figure 2 caption - e.g. "a detail only the user holds" is an unusual way to say "a fact known only by the user" (also, does the user necessarily know this? anyways...), and "chained aggregations collapse the cube axis by axis" is a complex phrase that probably needs some unpacking for the average reader. Similar comment for "supply questions that need a detail only the user holds" in the main text - some LLMs seem to really like phrases like "holds", "carries", "load-bearing", etc.

W2: Thank you for clarifying these points. I think your responses are reasonable, although I also think the text's discussion of uncertainty should have more nuance to capture these subtle distinctions.

W3: I did not mean to imply that I want to test this with my own prompts (although that would be neat), but I think an interesting application of this method could be understanding uncertainty in unsupervised settings. Can this give us insights about whats going on in a prompt before we collect labels? And sorry, but can you offer a short description of the result and takeaway for the table in djDn?

W4: Thanks for the explanations here. Will you be updating your submission with these changes before the review period ends?

[1] Clark, T. H., Martin, M. B., & Mahowald, K. (2026). Why is AI-generated language so hard to understand? [2] https://claudisms.ai

Dear WKh6, thank you for the concrete pointers and the references.

W1. We share your perspective and appreciate your encouraging words. We already implemented some changes in orange (see paper) to account for LLM-y language and will work on this human-y until the final deadline.

W2. Thank you. Section 3.1 now contains a paragraph that separates the sources of uncertainty (changes in blue), with your coin example and the distinction between TASK and CONTEXT.

W3. Thank you for clarifying. The leave-one-corpus-out evaluation comes close to this setting, as the classifier never sees a prompt of the held-out corpus (or its labels). On the real user prompts of WildChat it still reaches an AUROC of up to 0.82. If we just take a look at our signals, a high entropy at the start of denoising can already support indentifying TASK and UNRESOLVED prompts (Section 5.1). Thus, in a prompt-only setting, such signals could help to assign sources of uncertainty in an unsupervised setting.

The tables in our reply to djDn ask two questions. Does the trajectory carry source information beyond corpus and answer length? And how does it generalize to unseen datasets? We added this ablation in our paper in section D.6 and table 18 (blue). The takeaway is that the trajectory reveals information about the source of uncertainty beyond what we can derive from the prompt origin or length of the answer. The second table shows that this information also transfers to unseen corpora.

W4. Yes, we have uploaded a revised paper with all changes marked in blue (for contents) and orange (for language).

Reviewer_djDnOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether a single diffusion-LM generation reveals the source—not just the magnitude—of a prompt’s uncertainty. SOUL, a 16,224-prompt benchmark from seven public corpora plus 6,221 authored prompts, labels four sources by what would resolve them—TASK (ask the intent), CONTEXT (ask a user-held detail), UNRESOLVED (abstain), ALEATORIC (answer one of several valid replies)—plus CONTROL, via two LLM annotators with 94–97% agreement. A generation is summarized by a position×time posterior cube: 26–27 interpretable signals and a small CNN probe. On four dLLMs the classifiers reach F1 0.42–0.47 on the five-way task (vs. 0.13–0.20 naive, 0.25–0.34 answer-length-only), beating tokenizer-matched AR models by +0.01–0.06 overall and +0.08–0.22 on TASK, the only consistently positive source; removing the time axis within a model drops F1 to 0.29–0.35, and the probe adds ≤0.03 F1. CONTEXT remains indistinguishable from CONTROL. It is framed as a routing floor, not a policy.

Critical strengths
  1. A genuinely new question, answered positively and cleanly isolated. Within a fixed dLLM, deleting the time axis collapses F1 from 0.42–0.47 to 0.29–0.35 (D.5; same architecture and scale); gains over tokenizer-matched AR concentrate where a temporal signature is expected (TASK +0.08–0.22 F1; commit NLL up to 0.84 stratified AUROC); and CNN probe gaps of ≤0.03 F1 show simple statistics carry the information. The negative result is equally informative: sources that leave the resolution process unchanged (CONTEXT) are invisible, confirmed across all analyses.
  2. Careful, reusable measurement infrastructure. SOUL unifies seven corpora under one action-based taxonomy with documented agreement, and the authored set is balanced across every source and style to break the corpus–label confound. Evaluation includes corpus×length-stratified AUROCs, length-only baselines, a time-axis ablation, and an honest post-mortem on the failed rarity proxy; code, data, and out-of-fold predictions are released.
Critical weaknesses
  1. The headline multivariate numbers are confounded by prompt provenance and length, with no multivariate control. Table 3 pools across corpora and lengths, yet labeled prompts are answered at several times CONTROL length (~half hit the 512-token cap), corpus identity strongly predicts labels, and a length-only classifier already attains 0.25–0.34 F1. Without a corpus+length baseline or authored-set-only results, it is unclear how much of the 0.42–0.49 F1 is trajectory-specific; the attributable deltas are modest and TASK-centered (+0.01–0.06 over AR), so the “practical routing” framing may overread.
  2. No multi-sample/auxiliary-model comparison, and limited scope checks. SOUL never evaluates clarification ensembling or multi-response interpretation at any budget, leaving the accuracy-vs-cost trade-off implied by the framing unquantified. All measurements use greedy decoding, which is known to alter commit dynamics; labels come from two LLM annotators validated only on subsamples (125-item calibration; 120 human audits) with an undisclosed adjudication procedure; and 38% of prompts are LLM-authored—together bounding label reliability and generalization.
Questions for authors
  1. From the released out-of-fold predictions, can you report corpus-only, length/NFE-only, and corpus+length baselines for the Table 3 task, together with results within the balanced authored set? This would isolate trajectory-specific information from the prompt-provenance and answer-length confounds.
  2. At a feasible budget, such as 3, 5, or 10 generations or one auxiliary-LLM pass, how do clarification ensembling or multi-response source identification compare with single-generation trajectory signals?
2 comments

Dear djDn, thank you for your time and the actionable review. You made some very valid points and we ran all requested analyses to adress them. Besides including them in the final version of the paper we would like to discuss them here with you:

W1 and Q1. We computed every requested baseline with the protocol of Table 3 and report the five class macro F1. Since markdown does not support merged cells, a concept in the first column applies to all rows below it until the next concept. Bold marks the best value within each concept.

ConceptFeaturesDGemmaDreamNemotronLLaDA2.1
single featureCorpus only0.320.320.320.32
Length only0.340.290.250.31
NFE only0.34-0.320.29
Trajectory (Table 3)0.460.420.470.42
combined featuresCorpus and length0.400.390.370.39
Trajectory, corpus and length0.540.490.530.50
authored set onlyLength only0.350.310.270.33
Trajectory0.540.450.540.49

The table shows that the trajectory carries information that corpora and length do not carry. This finding also holds for the balanced authored set.

Our new evaluation that holds out one corpus at a time confirms the transfer to unseen corpora as well. Values are AUROC within length bins, averaged over the four dLLMs:

Held out corpusAmbigQACLAMBERKUQPopQAWildChatAuthoredMean
Length only0.500.510.510.540.510.530.52
Trajectory0.530.630.670.590.680.680.65

W2 and Q2. We now quantify the accuracy-vs-cost trade-off with a multi-sample and auxiliary-model comparison. It covers a stratified subset of 1,000 prompts with all four dLLMs, one, three, five and ten generations, and one or two auxiliary calls (see the table in our reply to Reviewer ra8q). In response to your questions, we want to highlight three findings. First, the "posterior multiple runs" block extends our approach to multiple generations. It shows that the convolutional probe already matches or exceeds ten sampled generations of our signals from a single generation. Second, the "auxiliary LLMs" block shows that clarification ensembling after Hou et al. performs worse. It uses about seven times the compute and two auxiliary calls, yet stays below our single generation performance. Third, the same block shows that multi-response interpretation after Nan et al. does not improve with more responses. An LLM that reads only the prompt matches it (with our labeling conventions as context, see Section A.1). SOUL asks a different question, namely whether the computation of the model itself reveals the source. Among the methods that read this computation, our signals and our probe lead at every budget, at no cost beyond the generation itself.

We also recorded the subset again with temperature sampling at T=0.5T = 0.5. One sampled generation yields almost the same F1 as a greedy one and averaging the draws adds 0.05 to 0.10. The signals therefore do not depend on greedy commit dynamics. We will state this in the paper.

The labels come from a fixed questionnaire (Section A.1), and code maps the recorded answers to labels. Two LLM labelers agree on all labels for 94.4% of a calibration batch of 125 prompts with Cohen's κ\kappa between 0.88 and 1.00, and we discarded a third labeler because it followed the prior of each corpus. An adjudicator answers every disagreement and every unsure item again under the consistency rules of the questionnaire instead of a majority vote, as Example 2 illustrates. We will make this more explicit in the paper. Two checks validate the labels independently of the labelers. The labels agree with the intended source for 96.8% of all 6,221 authored prompts, and a human who did not see the LLM labels agrees with them on 97% of 120 random prompts.

We would be grateful if you reconsider the score in light of these findings.

Thank you again for your review. As the discussion period closes, we wanted to follow up on our rebuttal. We’d welcome any further feedback or remaining concerns, and hope you’ll consider our response in your final assessment.

Reviewer_ra8qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether diffusion language models can identify the source of uncertainty from a single denoising trajectory. The authors introduce SOUL, a benchmark of 16,224 prompts labeled with four uncertainty sources: task underspecification, context underspecification, unresolved knowledge, and aleatoric uncertainty.

The approach extracts statistical signals from denoising trajectories and evaluates their ability to distinguish uncertainty sources across four diffusion language models. Comparisons with autoregressive models and a convolutional probe show that diffusion trajectories provide additional information, particularly for task underspecification. However, the overall classification performance remains moderate, and the improvements vary across uncertainty categories.

Critical strengths
  1. The paper moves beyond measuring uncertainty to identifying its underlying source, which is useful for deciding whether a model should answer, clarify, or abstain. The SOUL benchmark combines existing datasets with newly generated prompts and provides a structured evaluation of different uncertainty categories.

  2. The authors evaluate four diffusion models, compare against autoregressive baselines, and conduct within-model ablations to isolate the contribution of denoising trajectories. The finding that simple statistical features perform comparably to a convolutional probe is particularly interesting. The paper also acknowledges cases where trajectory information provides little benefit.

Critical weaknesses
  1. The broader problem of identifying uncertainty sources has already been explored by Hou et al. (2024), Kirchhof et al. (2025), and Nan et al. (2025). Vazhentsev et al. (2026) also studies lightweight uncertainty signals extracted from diffusion trajectories. Although SOUL's single-generation approach is a useful distinction, its methodology mainly builds on existing uncertainty measures and standard classifiers. The absence of direct comparisons with previous source-identification methods under a common evaluation setting makes the practical advantage difficult to assess.

  2. The uncertainty labels are strongly correlated with source datasets and response lengths. Although the authors include controls for these factors, the main classification results are still reported on pooled data, without leave-one-corpus-out evaluation. Additionally, 6,221 prompts are synthetically generated, while independent human validation covers only 120 examples. The benchmark also allows overlapping uncertainty sources, but the main five-class evaluation excludes such cases. These concerns are significant because the reported macro-F1 of 0.42–0.49 remains modest, and improvements over autoregressive models are inconsistent across categories.

Questions for authors
  1. Can the authors compare SOUL with existing uncertainty-source identification methods, particularly Nan et al. (2025), under matched datasets and computational budgets? What advantage does single-generation detection provide in practice?

  2. Can the authors conduct leave-one-corpus-out evaluation to establish whether the model identifies uncertainty sources rather than dataset-specific patterns? How reliable are the labels beyond the 120 human-validated examples?

  3. How would the results change when prompts contain multiple uncertainty sources? Can the authors evaluate stronger trajectory probes and alternative decoding strategies, especially for context underspecification?

2 comments

Dear ra8q, thank you for your time and the constructive review. You raised several valid concerns, and we ran five new experiments to address them. We will include all of them in the final version of the paper and would like to discuss them here with you:

W1 and Q1. Kirchhof et al. (2025) argue in a position paper for sources beyond aleatoric and epistemic uncertainty but propose neither an estimator nor data. SOUL turns their view into a benchmark. Hou et al. (2024) and Nan et al. (2025) infer the source from repeated queries and an auxiliary model. Hou et al. split the entropy over clarified inputs into two terms, so they cannot express TASK or separate UNRESOLVED from ALEATORIC. Vazhentsev et al. (2026) read the denoising trajectory to detect errors and do not distinguish sources. Our work is the first to show that a single dLLM generation carries information about the source of uncertainty.

We compared all methods on a stratified subset of 1,000 SOUL prompts with DeepSeek V4 Flash as the auxiliary model. All values are five class macro F1. Due to the formatting limits of markdown, rows with an empty first column belong to the concept above them. Bold marks the best value within each concept.

ConceptMethodGenerationsLLM callsDGemmaDreamNemotronLLaDA2.1
posterior onlyVazhentsev et al.100.460.320.420.36
Ours (signals)100.460.330.460.37
Ours (signals + curves)†100.510.420.500.47
Ours (conv. probe)†100.560.430.530.49
posterior multiple runsVazhentsev et al.1000.480.390.500.40
Ours (signals)1000.520.430.510.44
auxiliary LLMsHou et al.≈7.320.340.340.340.33
Nan et al.110.800.830.780.79
Nan et al.1010.750.810.770.76
Ours (prompt only)010.820.820.820.82

† trained on the 15,108 SOUL prompts outside the subset

The table displays that our signals reveal the source of uncertainty from a single generation. Rows marked † train on all prompts outside the subset, the other rows within it. Trained on the same prompts, the convolutional probe stays close to our signals, as in Table 3 of the paper. Additionally, our signals match or exceed those of Vazhentsev et al. at every budget (generation). To instruct the auxiliary models which labels to assign, we give them our labeling conventions as context. They therefore follow nearly the same process that produced the labels, and the prompt-only LLM performs best and marks an upper reference. Nan et al. perform best with a single generation. Hou et al. stay below our single generation on three of four models despite about seven times the compute. Our method instead asks whether the computation of the model itself reveals the source, and trajectory signals read this computation at no cost beyond the generation.

W2 and Q2; leave-one-corpus-out evaluation. On the unseen corpus our signals identify sources rather than corpora. They reach a mean AUROC of 0.65 and up to 0.82 on real WildChat prompts, whereas answer length reaches only 0.52. Corpus and length together reach a five class F1 of only 0.37 to 0.40, and adding the trajectory raises it to 0.49 to 0.54. Our reply to Reviewer djDn reports these baselines in full. The labels rest on more than the human audit of 120 prompts with 97% agreement. The labels of all 6,221 authored prompts agree with their intended source in 96.8% of cases, two LLM labelers agree on 94.4% of a calibration batch with Cohen's κ\kappa between 0.88 and 1.00, and an adjudicator resolves their disagreements. The gain over the autoregressive models depends on the source, as it is consistent for TASK at 0.08 to 0.22 and absent for CONTEXT, and we will state this explicitly.

Q3. 73 prompts in SOUL carry two sources. We tested one classifier head per source. On these prompts, the ALEATORIC and UNRESOLVED heads score as high as on single source prompts, whereas the CONTEXT head drops clearly. Appendix D.8 of the revised paper reports the results.

We liked your idea of stronger convolutional probes and ran additional experiments with various CNN architectures of up to 300K parameters (50x the inital size). Yet, they change F1 by at most 0.02. Furthermore, temperature sampling at T=0.5T = 0.5 yields an F1 within 0.05 of greedy decoding from one generation, and averaging ten draws adds 0.05 to 0.10. CONTEXT remains the hardest source in every variant, and we will make it the central open question of the limitations.

We would be grateful if you reconsider the score in light of these results.

Thank you again for your review. As the discussion period closes, we wanted to follow up on our rebuttal. We’d welcome any further feedback or remaining concerns, and hope you’ll consider our response in your final assessment.

Comments on the paper

1 comment

Dear reviewers, thank you for your reviews. The revised paper includes all changes announced in our replies, marked in blue.

  • Comparison with prior methods at matched budgets, including Hou et al., Nan et al. and Vazhentsev et al. (Section 5.4, Appendix E)
  • Corpus, length and NFE baselines, the authored set and the leave-one-corpus-out evaluation (Section 5.4, Appendix D.6)
  • Larger probes and temperature sampling (Sections 5.3 and 5.4, Appendix D.7)
  • Prompts with two sources (Appendix D.8)
  • Separation of the sources with the coin example, and the labeling procedure (Sections 3.1 and 3.3)
  • Definitions of the posterior cube, the commit step, the interpretable temporal statistics and τ½ (Sections 4.1 and 4.2, Figure 1)
  • Gain over the autoregressive models by source, and CONTEXT as the central open question (Sections 5.2 and 6)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.