acceptodds
Under review as a conference paper at ICLR 2027

Generation-Free Verification in Diffusion Language Models via Anchor Chain Scanning

Abstract

As language models solve increasingly complex reasoning problems, the need for reliable verification grows accordingly. Notably, a single error can propagate through a reasoning trajectory and yield a plausible yet incorrect answer. To address this challenge, recent work has explored generative verifiers based on autoregressive language models and diffusion language models (DLMs). However, such generative methods can hallucinate during verification and often waste computation on unnecessary parts of the reasoning trajectory. To overcome this, we propose De-Anchoring: a generation-free DLM-based verifier that focuses on information anchor chains, defined as repeated occurrences of key information throughout a reasoning trajectory. De-Anchoring simultaneously masks all occurrences of each anchor and uses their DLM reconstruction probabilities to measure how strongly the remaining reasoning supports the original information. Across practical verification tasks, including Best-of-N selection, unsupervised error diagnosis, and auxiliary reward signals for reinforcement learning, De-Anchoring substantially outperforms generative verification baselines. Moreover, De-Anchoring remains robust under aggressive cost reduction, achieving comparable performance with only 12.5% of the generative verification cost.

Primary area:
generative models
Keywords:
Diffusion Large Language Models, reasoning, verification

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_S6d4OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes De-Anchoring, a verifier for diffusion language models. It selects key tokens, such as derived numbers, option letters, and code identifiers, and groups their repeated occurrences into anchor chains. It masks one chain at a time, together with the final answer where defined, and scores the original tokens with one forward pass through a frozen model. It averages these log-probability scores to rank responses without generating new text. Experiments cover response selection and correctness scoring on nine benchmarks with two models, plus an RL application on math tasks.

Critical strengths
  1. The method is simple and needs no verifier training or external model. Each anchor chain takes one forward pass. Tables 3 and 4 support the choice of informative anchors and the need to mask all their occurrences. Table 6 also shows that checking just one chain retains much of the selection accuracy. This makes the method a useful way to reduce verification work.

  2. The gains are consistent across models and tasks. In Table 1, De-Anchoring beats BMC in all 18 model–dataset settings. For example, Dream's HumanEval accuracy rises from 46.34% to 51.22%, and LLaDA's MBPP accuracy rises from 46.95% to 51.75%. Table 11 shows gains with another sampling setup, and Table 2 generally supports better separation of correct and incorrect responses. The evidence goes beyond a single benchmark or sampling choice.

Critical weaknesses
  1. The BMC comparison needs more detail. Section 5.2 says the comparison is budget-matched, but the settings behind Tables 1 and 2 are not clear: the mask ratio, number of reconstructions, and steps per reconstruction. Table 6 also needs an explanation of how K and N_BMC determine the total forward passes. These details matter because better accuracy at lower verification cost is a main claim. The current description makes that claim hard to check.

  2. The RL comparison changes more than the verifier. In Table 13, BMC gives correct responses a continuous score-based reward, with a coefficient that increases during training. De-Anchoring gives a fixed reward to the top half of the correct responses and uses a constant coefficient. The gains in Table 7 could therefore come partly from the reward rule or schedule. This weakens the claim that a better verification signal causes the RL improvement.

Questions for authors
  1. For BMC in Tables 1, 2, and 6, what mask ratio, number of independent reconstructions per candidate, and denoising steps per reconstruction did you use? How were these values chosen? Were total forward passes per candidate matched between the methods? Please also explain how K and N_BMC map to this budget in Table 6.

  2. Could you compare against random masking with the same mask count, answer masking, scoring rule, and total forward passes? Tables 3–5 support the design, but this control would help show how much the chain structure adds beyond the amount of context kept visible. Matching mask counts here would test the design, rather than match compute cost.

  3. In the RL experiment, does De-Anchoring still outperform BMC when both use the same rule for turning scores into rewards and the same reward coefficient schedule?

4 comments
AuthorsOpenReview ↗

Response to Reviewer S6d4 [Part 1]

Dear Reviewer S6d4,

We sincerely appreciate your review with thoughtful comments. We have carefully considered each of your questions and provide detailed responses below. Please let us know if you have any further questions or concerns.


W1 & Q1. For BMC in Tables 1, 2, and 6, what mask ratio, number of independent reconstructions per candidate, and denoising steps per reconstruction did you use? How were these values chosen? Were total forward passes per candidate matched between the methods? Please also explain how K and N_BMC map to this budget in Table 6.

We thank the reviewer for this thoughtful observation. To improve the clarity of our paper, we have incorporated the requested details into Section 6.2 (page 9) of the revised manuscript, with the additions highlighted in green.

BMC uses a mask ratio of 0.9, following the default setting in the original paper. The two BMC hyperparameters highlighted by the reviewer, the number of independent reconstructions per candidate and the number of denoising steps per reconstruction are respectively denoted by NBMCN_{\mathrm{BMC}} and TT; then, BMC spends NBMC×TN_{\mathrm{BMC}} \times T forward passes per candidate. For a fair performance comparison, we matched this computational budget to De-Anchoring on a per-candidate basis. De-Anchoring verifies m=min⁡(∣V∣,K)m = \min(|V|, K) anchor chains per candidate, using one forward pass per chain, where KK is the cap on verified chains (K=8K = 8 in Tables 1 and 2). Accordingly, BMC must also be allocated exactly mm forward passes for the same candidate, i.e., NBMC×T=mN_{\mathrm{BMC}} \times T = m. However, when mm is prime (e.g., 5 or 7), setting NBMC>1N_{\mathrm{BMC}} > 1 forces T=1T = 1, which the original BMC paper [1] reports to be a highly unfavorable setting.

To mitigate this fairness issue as much as possible, we adopt two measures in Tables 1 and 2. First, we set NBMC=1N_{\mathrm{BMC}} = 1 and T=mT = m, ensuring that both methods use exactly the same total number of DLM forward passes for every candidate. Second, to reduce the potential disadvantage to BMC arising from variance due to random masking, we report its performance averaged over three masking seeds.

Meanwhile, Table 6 further explores BMC's potential by increasing NBMCN_{\mathrm{BMC}}. Starting from the per-candidate budget mm, for candidates whose budget reaches the column-wise cap KK (i.e., m=Km = K), we reallocate the budget as NBMC×TN_{\mathrm{BMC}} \times T with T=K/NBMCT = K / N_{\mathrm{BMC}}, while all other candidates keep NBMC=1N_{\mathrm{BMC}} = 1 and T=mT = m. This setting allows us to explore BMC's potential more thoroughly while maintaining the same number of DLM forward passes for De-Anchoring and BMC. Nevertheless, increasing NBMCN_{\mathrm{BMC}} yielded only marginal differences and still failed to outperform De-Anchoring.

Importantly, we remark that we exclude the forward passes of BMC's external sentence encoder [2] from the reported computational budget, as they are not directly comparable to DLMs forward passes. Thus, BMC incurs additional FLOPs from the external model in all settings reported in Tables 1, 2, and 6. This makes the computational budget comparison favorable to BMC and conservative for De-Anchoring. Nevertheless, the results show that BMC remains within 65.5–66.1 across all budget allocations, consistently underperforming De-Anchoring at every matched KK. Remarkably, De-Anchoring with K=1K = 1 even outperforms BMC with K=8K = 8 across all reported configurations.

AuthorsOpenReview ↗

Response to Reviewer S6d4 [Part 2]


W2 & Q3. In the RL experiment, does De-Anchoring still outperform BMC when both use the same rule for turning scores into rewards and the same reward coefficient schedule?

We thank the reviewer for suggesting this important ablation. To directly address your concern, we ran a controlled comparison in which De-Anchoring and BMC differ only in how the score is computed. Every other rule is shared, including how scores are turned into rewards and the schedule of the reward coefficient. In addition to matching the rule and schedule, we ensure identical verification costs for BMC and De-Anchoring by allocating exactly mm forward passes to BMC per candidate, where m=min⁡(∣V∣,8)m = \min(|V|, 8) is the number of forward passes used by De-Anchoring for that candidate.

Under this fully controlled setting, De-Anchoring still outperforms BMC on each of the three math tasks: 81.73 vs. 79.30 on GSM8K, 24.61 vs. 16.41 on Countdown, and 33.00 vs. 32.80 on MATH. On both Avg and Pooled Avg, De-Anchoring outperforms every BMC configuration, including settings with a larger verification budget (i.e., fixed 8 forwards). This indicates that De-Anchoring provides a more efficient and robust verification signal than BMC, in line with its strong Best-of-N and AUROC results reported in the paper.

RewardGSM8KCountdownMATHAvgPooled Avg
LLaDA-8B-Instruct (no RL)78.2418.3631.6042.7359.61
diffu-GRPO78.7723.83‾\underline{23.83}32.4045.0060.82
+ BMC (B-rule, 8 forwards)79.6122.2731.8044.5661.02
+ BMC (D-rule, 8 forwards)79.2323.83‾\underline{23.83}33.8045.62‾\underline{45.62}61.45‾\underline{61.45}
+ BMC (D-rule, mm forwards)79.3016.4132.8042.8460.34
+ De-Anchoring (D-rule, mm forwards)81.7324.6133.00‾\underline{33.00}46.4562.94

Avg is the unweighted mean over the three benchmarks; Pooled Avg weights each benchmark by its number of test questions (GSM8K 1319, Countdown 256, MATH 500). B-rule and D-rule denote the BMC rule and the De-Anchoring rule, respectively. Each rule fixes both how scores are turned into rewards and the schedule of the reward coefficient.

AuthorsOpenReview ↗

Response to Reviewer S6d4 [Part 3]


Q2. Could you compare against random masking with the same mask count, answer masking, scoring rule, and total forward passes? Tables 3–5 support the design, but this control would help show how much the chain structure adds beyond the amount of context kept visible. Matching mask counts here would test the design, rather than match compute cost.

We thank the reviewer for suggesting this valuable ablation study. Random masking that controls for the number of masked tokens in each chain provides an opportunity to directly test the effectiveness of our design, as the reviewer suggested. We conducted this experiment, and the results further demonstrate the clear advantages of our anchor-chain structure.

DatasetMasking MethodBoNAUROC
GSM8KDe-Anchoring (anchor chains)80.160.764
GSM8KSame-count random positions62.96 ± 1.710.544 ± 0.005
MATHDe-Anchoring (anchor chains)58.330.710
MATHSame-count random positions51.07 ± 0.800.516 ± 0.017
CountdownDe-Anchoring (anchor chains)58.020.688
CountdownSame-count random positions31.30 ± 2.250.455 ± 0.020

Anchor-chain masking vs. random masking of the same masking amount. LLaDA-8B-Instruct, problems whose candidate sets contain both correct and incorrect responses. Mean ± std over three seeds.

Following Section 5.1 of our paper, we consider queries whose candidate sets contain both correct and incorrect responses. We evaluated GSM8K, MATH, and Countdown on the same judgment set used in Tables 3–5 (LLaDA-8B-Instruct; random masking results are reported as mean ± standard deviation over three seeds). Specifically, we replace each anchor chain with an equal number of randomly selected token positions, ensuring that the computational cost, number of masked tokens, and scoring rule remain identical, with only the masked positions differing. The results demonstrate that the anchor-chain design structurally establishes a robust and reliable verification setting.


[1] Ruan et al. Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Models. ICML 2026.

[2] Reimers and Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP 2019.


If you have any further questions, please let us know.

Thank you very much, Authors

Thank you for the clarification and additional experiments. The BMC settings and the explanation of the forward-pass budget address my concern about compute matching. The RL comparison with the same reward rule and schedule addresses the confounding issue. The random-masking experiment also provides clearer evidence for the anchor-chain design.

My main concerns have been addressed. I maintain my rating of 3, with greater confidence in the recommendation. Please include these details and additional results in the revised paper.

Reviewer_AXE1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a generation-free verfication approach for DLM, DE-ANCHORING, which masks the task related anchor token spans and let DLM to reconstruct the masked tokens to validate the trajectory consistency. The evaluation on LLaDA-8B and Dream-7B across various domain benchmarks shows the method is valid and effective.

Critical strengths
  • This work proposed a new generation-free verification approach for DLM, DE-ANCHORING. The core idea is to mask the occurance of anchor tokens and reconstruct with consistency.
  • The evaluation runs across 2 models with multiple baselines and diverse benchmarks, showing the approach can improve the verification performance on reasoning, math, knowledge QA.
Critical weaknesses
  • A major concern for this work is, the author propose this work as a generation-free verification, but it lacks enough justification. Even under this DLM setup, the model needs to run generation (even not from scratch). Also, all the tasks evaluated carried with gold reference answer, which makes it easy for a rule-based verifier to run generation-free verification. It would be better to show the potential on open-ended tasks, such as
  • The approach highly depends on hand-crafted tokens for masking to work with the approach, which is hard to generalize to broad usecase.
  • The mechanism underlying the verification capability remains unexplained. For the same DLM used for generation and verification, it is unclear why the model can reliably identify incorrect anchor tokens through masked reconstruction whereas it failed to produce the correct values during generation. A controlled analysis comparing token probabilities during generation and masked reconstruction would help substantiate the claimed verification mechanism.
Questions for authors
  • The approach makes sense when the trajectory contains the inconsistency itself, but I wonder if the model is incorrect in a different way, such the model use wrong knowledge, wrong formular, wrong understanding, etc., can this approach still work?
4 comments
AuthorsOpenReview ↗

Response to Reviewer AXE1 [Part 1]

Dear Reviewer AXE1,

We sincerely appreciate your review with thoughtful comments. We have carefully considered each of your questions and provide detailed responses below. Please let us know if you have any further questions or concerns.


W1. A major concern for this work is, the author propose this work as a generation-free verification, but it lacks enough justification. Even under this DLM setup, the model needs to run generation (even not from scratch). Also, all the tasks evaluated carried with gold reference answer, which makes it easy for a rule-based verifier to run generation-free verification. It would be better to show the potential on open-ended tasks.

We would like to clarify the three potential misunderstandings, and present additional results on open-ended tasks to fully address the reviewer’s concerns.

First, De-Anchoring does not "generate" tokens. Our method does not unmask or commit new tokens; it only scores the consistency of an already generated trajectory using probabilities, as formally defined in Eqs. (2)–(5) of Section 4.2. In contrast, Self-Evaluation [1] and BMC [2] are generative verifiers: the former explicitly generates feedback text, while the latter regenerates and commits the entire trajectory. Moreover, Table 6 empirically shows that De-Anchoring reduces verification cost to as little as 0.78% of the trajectory generation cost, yet achieves better micro-averaged performance across 2 models and 9 datasets than generative verifiers that use 8× more verification compute.

Next, we have already presented evaluations on open-ended code generation tasks, HumanEval and MBPP. For example, Dream's HumanEval accuracy improves from 46.34% to 51.22%, while LLaDA's MBPP accuracy improves from 46.95% to 51.75%. Unlike tasks with a single fixed gold reference answer, these benchmarks evaluate correctness through execution-based unit tests, allowing open-ended valid generations.

Lastly, to fully address the reviewer’s concerns, we further evaluate the potential of De-Anchoring on additional open-ended tasks. Arena-Hard [4] uses conversational prompts involving coding, explanation, and creative writing, while WildBench [3] includes real-world user requests for code debugging, mathematical problem-solving, and data analysis. These tasks are fully open-ended, with no fixed answers, unit tests, or answer keys. The results show that De-Anchoring consistently outperforms both BMC and Semi-AR under the same compute budget.

BenchmarkModelDe-Anchoring vs Semi-ARDe-Anchoring vs BMC
WildBenchLLaDA57.062.2
WildBenchDream58.856.1
Arena-HardLLaDA59.259.3
Arena-HardDream63.855.7

Bold indicates pairwise wins achieved by De-Anchoring (win rate > 50%). Pairwise win rates (%) judged by GPT-4o (temperature 0), with wins scored as 1 and ties as 0.5 (50% = parity). Each method selects from 6 candidates per prompt (5 sampled at temperature 0.6 and 1 Semi-AR greedy response). BMC and De-Anchoring use equal forward-pass budgets per candidate.

AuthorsOpenReview ↗

Response to Reviewer AXE1 [Part 2]


W2. The approach highly depends on hand-crafted tokens for masking to work with the approach, which is hard to generalize to broad usecase.

We acknowledge that the defined token categories are based on hand-crafted rules. However, these rules are simple yet generalizable across a broad range of tasks. Moreover, identifying better alternative rules is a non-trivial challenge, and our additional experiments demonstrate that the current rules outperform more complex alternative parsing strategies.

First, our parsing rules have the advantage of covering a broad range of tasks. De-Anchoring uses only three anchor categories, yet covers nine diverse reasoning benchmarks spanning mathematics, numerical puzzles, scientific reasoning, truthfulness and hallucination resistance, general knowledge, multi-step logical reasoning, and open-ended code generation. Furthermore, we demonstrated that our parsing approach can also generalize to detecting inconsistencies in agentic workflows. For example, we applied the same parsing used for open code generation to tool calls of agentic traces by treating the argument values as anchors. Using at most 2.2 forward passes per call, De-Anchoring distinguished incorrect tool calls from correct ones better than BMC across three ARMs (AUROC 0.72–0.77 vs. 0.61–0.69). For further details on this application, please refer to our responses to W1 and Q1 of Reviewer mcit.

Second, our parsing rules are substantially simpler yet more effective than alternative parsing strategies designed to capture consistency. For example, Reviewer mcit suggested dependency-based and execution-trace-based parsing as alternatives. In our new experiments with them, we confirm that our current approach consistently outperformed both in Best-of-N selection and AUROC, thanks to its structure that prevents error leakage during verification. For example, De-Anchoring improves BoN over dependency-based and execution-trace chains by up to +13.6%p on GSM8K, +11.6%p on MATH, and +5.8%p on Countdown. For a detailed explanation of this experiment, please refer to our response to Q2 of Reviewer mcit.

AuthorsOpenReview ↗

Response to Reviewer AXE1 [Part 3]


W3. The mechanism underlying the verification capability remains unexplained. For the same DLM used for generation and verification, it is unclear why the model can reliably identify incorrect anchor tokens through masked reconstruction whereas it failed to produce the correct values during generation. A controlled analysis comparing token probabilities during generation and masked reconstruction would help substantiate the claimed verification mechanism.

We thank the reviewer for this insightful comment. The underlying mechanism of verification lies in the difference between the contexts available during generation and verification. We would like to first explain the mechanism and then present the results of controlled experiments that provide supporting evidence.

First, during generation, an anchor is decoded while the rest of the chain and the final answer remain masked, so even an incorrect intermediate value may receive high confidence if it appears locally plausible. In contrast, during verification, the model has access to the full preceding and subsequent context except for the masked anchor, allowing it to detect inconsistencies that were not apparent during generation.

Second, we conducted the suggested controlled analysis. The results reveal that the unmasking probabilities of the same tokens can differ substantially between the generation and verification stages, and that this discrepancy is more pronounced for incorrect trajectories than for correct ones. De-Anchoring exploits this asymmetry to distinguish incorrect trajectories from correct ones.

Using LLaDA-8B-Instruct, we analyze the problems in GSM8K, Countdown, and MATH whose five candidates disagree in correctness. For each candidate trajectory, we identified the anchor with the lowest probability during verification and binned candidates according to that probability.

Anchor prob. @verifyCorrect: @gen → @verifyIncorrect: @gen → @verifyCorrect shareIncorrect share
[0.8, 1.0]0.93 → 0.990.90 → 0.9782%42%
[0.6, 0.8)0.78 → 0.690.75 → 0.693%7%
[0.4, 0.6)0.70 → 0.490.73 → 0.492%10%
[0.2, 0.4)0.72 → 0.300.68 → 0.292%11%
[0.0, 0.2)0.53 → 0.040.59 → 0.0511%30%

LLaDA-8B-Instruct, GSM8K + Countdown + MATH pooled (1657 correct / 1424 incorrect candidates). Each cell shows the mean probability the same token had at the decoding step where it was produced (@gen) → its probability at verification (@verify); "share" is the fraction of candidates of each class in the bin.

As shown in the table, anchors that receive low probability at verification were not low-confidence when they were generated. In the last row, for instance, incorrect candidates whose weakest anchor falls below 0.2 at verification (mean 0.05) had decoded that same anchor with a mean probability of 0.59. The model thus committed to these tokens with ordinary confidence during generation, yet rejected them only when verification exposed the full context bidirectionally.

Within each bin, correct and incorrect candidates behave alike; what differs is how often such anchors occur. Only 11% of correct candidates contain an anchor that collapses below 0.2, compared with 30% of incorrect candidates, while 82% of correct candidates keep every anchor above 0.8 versus 42% of incorrect ones. This asymmetry gives incorrect candidates systematically lower De-Anchoring scores, making the verification signal reliable.

AuthorsOpenReview ↗

Response to Reviewer AXE1 [Part 4]


Q1. The approach makes sense when the trajectory contains the inconsistency itself, but I wonder if the model is incorrect in a different way, such the model use wrong knowledge, wrong formular, wrong understanding, etc., can this approach still work?

We thank the reviewer for this insightful and thoughtful observation. We empirically confirm that De-Anchoring remains effective even when the trajectory does not contain the inconsistency itself, as long as the model can recognize them within its knowledge boundary.

The table below categorizes the 13,144 incorrect trajectories of LLaDA and Dream on GSM8K, MATH, and Countdown by error type and reports the AUROC of De-Anchoring for each category. Error types are assigned by GPT-4o with a fixed rubric and cross-checked by a deterministic checker that recomputes the explicit equations and common verbal computations in each trajectory. 40.7% of incorrect trajectories are labeled as self-contradictory (i.e., they contain an inconsistency within the trace itself), and De-Anchoring detects them with an AUROC of 0.808, consistent with the findings throughout our work. De-Anchoring also flags malformed trajectories (5.5%) whose reasoning is broken (corrupted, truncated, or missing), with an AUROC of 0.843.

Notably, the remaining errors contain no such inconsistency within trajectory itself, and De-Anchoring still detects each of the error types the reviewer mentioned: a. wrong knowledge (AUROC 0.784), b. wrong formula (AUROC 0.744), and c. wrong understanding (AUROC 0.697). All three are substantially above the random-chance baseline of 0.5.

Error typeIncorrect (share)AUROC
Self-contradictory5,344 (40.7%)0.808
Malformed trajectory722 (5.5%)0.843
Consistent: a. Wrong knowledge171 (1.3%)0.784
Consistent: b. Wrong formula4,403 (33.5%)0.744
Consistent: c. Wrong understanding2,504 (19.0%)0.697

Self-contradictory errors leave a trace that disagrees with itself: a computation is wrong for its own operands, a value derived earlier is later used with a different value, or the reported answer differs from the trace's final result. Malformed trajectories have broken reasoning that is corrupted, truncated, or missing. Consistent errors contain no such inconsistency: a false fact or formula not given in the problem (a. Wrong knowledge), a wrong relation or method for combining correctly read quantities (b. Wrong formula), or a misread number, condition, or target quantity (c. Wrong understanding).

This is possible because the model can detect anomalies within its knowledge boundaries by jointly considering the input prompt and the solution trajectory. As shown in Equations (4)–(5) of Section 4.2, the input condition for De-Anchoring is the context in which the anchor and prediction span are masked, allowing the model to consider both the input prompt and the reasoning process simultaneously. In other words, even when a trajectory is consistently incorrect, it ultimately contradicts either the input prompt or the model's own knowledge. As a result, the model cannot confidently reconstruct its original prediction, leading to a lower score. In contrast, as illustrated in the bottom panel of Figure 3 in our paper, correct trajectories do not exhibit such a score decrease, thereby enabling relative discrimination between correct and incorrect trajectories.

On the other hand, when an error is beyond the model's ability to recognize (i.e., when a request requires knowledge far beyond the model's knowledge boundaries), it is inherently difficult for the model to recognize such an error without additional training. This limitation is not unique to our method but is shared by other works on self-verification methods [1,2,5,6]. Therefore, the practical effectiveness of a self-verifying method should be evaluated based on its ability to detect errors within the boundaries where the model can recognize anomalies. The strong Best-of-N and AUROC results consistently demonstrated throughout our work highlight the effectiveness of our method in this regard.


[1] Madaan et al. Self-Refine: Iterative Refinement with Self-Feedback. NeurIPS 2023.

[2] Ruan et al. Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Models. ICML 2026.

[3] Lin et al. WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild. ICLR 2025.

[4] Li et al. From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline. ICML 2025.

[5] Lee et al. Test-time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts. ICLR 2026.

[6] Kadavath et al. Language Models (Mostly) Know What They Know. Anthropic, 2022.


If you have any further questions, please let us know.

Thank you very much, Authors

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

This paper proposes De-Anchoring, a generation-free verifier for diffusion language models. Given a reasoning trace, the method identifies repeated task-relevant information anchors, masks all occurrences of an anchor together with the answer span, and measures the reconstruction probability of the original anchor tokens using a single frozen DLM forward pass. The resulting anchor-chain scores are aggregated into a trace-level verification score.

Critical strengths
  • A well-motivated generation-free verification mechanism. Verification does not require generating a new trajectory. The de-anchoring operation removes repeated copies of a key fact and tests whether the remaining bidirectional context can reconstruct it. This directly addresses the cost and hallucination risks of generative verification for DLMs.

  • Clear presentation and comprehensive evaluation. The method is easy to follow, and the experiments cover two DLM backbones, nine benchmarks, Best-of-N selection, error diagnosis, cost-reduced verification, and auxiliary RL rewards. The comparisons with BMC and other verification baselines make the empirical contribution convincing within the evaluated setting.

Critical weaknesses
  • Coverage beyond static reasoning traces is unclear. The evaluation is broad across mathematics, multiple-choice reasoning, code generation, and planning, but it does not test agentic trajectories involving tool calls, external state, branching execution, or multi-step interaction. The current anchor definition and deterministic task-specific parser may work well for repeated values, option letters, and code identifiers, but it is unclear whether the same score remains meaningful when information is distributed across heterogeneous agentic traces. This is an important open question for the paper’s broader verification framing.

  • Anchor structure and RL portability are not fully isolated. The paper includes useful ablations for anchor classes, full-chain masking, extra random masking, chain subsampling, and the BMC comparison. However, the experiments do not compare the proposed rule-based anchor-chain structure against alternative structures or learned anchor selection, so it is difficult to determine why this structure is preferable beyond the reported task families. The RL study also uses only diffu-GRPO. Comparisons with other DLM RL frameworks, such as ESPO or Coupled-GRPO, would clarify whether the auxiliary reward is generally useful or mainly benefits one training objective.

Questions for authors
  1. How does De-Anchoring extend to agentic traces with tool calls, branching plans, external observations, or state changes? What would count as an information anchor in those settings?
  2. Why is the current rule-based anchor-chain structure preferable to alternative structures, such as learned span selection, dependency-based chains, or anchors defined from execution traces? Can the method be evaluated with at least one alternative structure?
  3. Does the auxiliary reward improve other DLM reinforcement-learning frameworks, including ESPO or Coupled-GRPO, under matched training budgets and policy architectures?
  4. How does De-Anchoring compare with broader generative-verification methods beyond BMC, including the JEV approachs?
  5. To what extent does the score depend on the task-specific parser, the maximum number of anchors, and the number of verified chains (K)?
5 comments
AuthorsOpenReview ↗

Response to Reviewer mcit [Part 1]

Dear Reviewer mcit,

We sincerely appreciate your review with thoughtful comments. We have carefully considered each of your questions and provide detailed responses below. Please let us know if you have any further questions or concerns.


W1 & Q1. How does De-Anchoring extend to agentic traces?

We thank the reviewer for suggesting this timely topic. One of the major threats to the reliability of heterogeneous agentic traces is silent errors in tool calling, which are particularly critical because they are difficult to detect due to the absence of explicit error signals and can ultimately cause the entire agentic workflow to fail. We experimentally demonstrate that De-Anchoring can effectively detect those error signals by verifying tool-call arguments as anchors, thereby improving the reliability of agentic workflows.

Silent errors occur when incorrect tool calls propagate errors to subsequent operations without explicit error signals, as tools may execute successfully despite receiving incorrect arguments. As shown in the experiments below, silent errors occur frequently (44–64%) when incorrect arguments are passed to tools and substantially reduce the success rate of subsequent turns. In our experiments for this new experiment, we used 7–8B-scale autoregressive LLMs (ARMs) as the backbones for agentic workflows to evaluate realistic agentic traces, as recent studies [6, 7] have reported that DLMs are not yet reliable backbones. For example, [6] reported a 0% success rate on multi-turn tool-calling tasks for both LLaDA-8B and Dream-7B.

GeneratorTurns w/ argument mis-transferSilent error rateNext-turn successAll remaining turns success
Llama-3.1-8B8452%0.61 → 0.110.43 → 0.02
Mistral-7B6564%0.47 → 0.060.32 → 0.01
Qwen2.5-7B2744%0.66 → 0.050.50 → 0.00

Silent error in BFCL [1]. We ran each model on BFCL multi-turn (734 turns) and collected the turns where the model passed a wrong argument taken from an earlier context. Silent error rate: share of these wrong calls that the tool executed without any error. The next two columns compare success after a correct turn vs. after an argument mis-transfer (correct → mis-transfer): the very next turn, and all remaining turns.

Since silent errors originate from tool calls with incorrectly transferred arguments, De-Anchoring targets these arguments as anchors to verify their consistency. Specifically, when verifying a tool call at a given turn, we used the conversation history up to the preceding turn as context. We collected two sets of turns: those in which incorrect arguments were passed to tools and those in which arguments of the same type were correctly provided. We then scored these turns using De-Anchoring with LLaDA-8B-Instruct and measured the AUROC for distinguishing between the two sets.

GeneratorForward StepsAUROC (De-Anchoring / BMC)Pair win rate (De-Anchoring / BMC)
Llama-3.1-8B2.20.73 / 0.630.78 / 0.69
Mistral-7B1.90.77 / 0.610.80 / 0.60
Qwen2.5-7B1.60.72 / 0.690.90 / 0.70

AUROC pools all correct and incorrect candidates across turns. The pair win rate is computed only over turns that contain both a correct and an incorrect candidate generated from the identical context; for each such turn we check whether the correct candidate receives the higher score (ties = 0.5), and average over turns. LLaDA-8B-Instruct is used as the verifier.

Overall, De-Anchoring consistently outperformed BMC in identifying incorrect tool calls under the same forward-pass budget (AUROC: 0.72–0.77 vs. 0.61–0.69). Notably, among turns where both correct and incorrect tool calls were generated from the same context, De-Anchoring assigned a higher score to the correct trajectory in 78–90% of cases. Considering that the computational cost corresponds to at most 2.2 forward steps of a DLM on average, these results highlight the potential of De-Anchoring as a lightweight and robust verifier for detecting silent errors in agentic workflows.

For pair win rate, we focus on turns where the generator produced both a correct and an incorrect tool call from the same conversation history. Each turn contributes one pair, and we measure how often the correct call receives a higher score than the incorrect one. Ties count as 0.5. Notably, De-Anchoring correctly ranks 78–90% of these pairs, compared to 60–70% for BMC, demonstrating its stronger ability to distinguish correct from incorrect tool calls under identical contexts.

AuthorsOpenReview ↗

Response to Reviewer mcit [Part 2]


W2 & Q3. RL portability over ESPO and Coupled-GRPO.

We thank the reviewer for suggesting this valuable experiment. To directly address the reviewer's concern, we conducted additional experiments and verified that the auxiliary reward introduced by De-Anchoring yields additional performance improvements even on top of the suggested ESPO and Coupled-GRPO frameworks, respectively, under matched training budgets and policy architectures.

MethodESPO, GSM8KESPO, MATHCoupled-GRPO, GSM8KCoupled-GRPO, MATH
No RL78.2431.6078.2431.60
Base RL (ESPO or Coupled-GRPO)79.4532.0080.67‾\underline{80.67}34.00
+ BMC79.61‾\underline{79.61}33.20‾\underline{33.20}80.2934.20‾\underline{34.20}
+ De-Anchoring80.1433.4080.8234.80

Accuracy (%) of ESPO and Coupled-GRPO with and without auxiliary rewards under the same rollout budget and policy on LLaDA-8B-Instruct. Each cell reports the highest accuracy achieved across a sweep of KL coefficients β ∈ {0.04, 0.14, 0.2}.

Given the limited discussion period and computing budget, we train each variant with a reduced budget of 384 rollouts, about 6.25% of the rollout budget used for the diffu-GRPO runs in Table 7 (Section 6.2). All RL variants share the same policy and optimization settings (lr=3e-5, μ=12, ε=0.5). Compared with LLaDA-8B-Instruct, De-Anchoring reaches gains of +1.90%p and +1.80%p with ESPO and +2.58%p and +3.20%p with Coupled-GRPO on GSM8K and MATH, respectively. In every setting, De-Anchoring adds an extra gain on top of the base RL trainer, which consistently exceeds the gain added by BMC.

AuthorsOpenReview ↗

Response to Reviewer mcit [Part 3]


Q2. Evaluation on alternative anchor structures, such as dependency-based chains, or anchors defined from execution traces.

We thank the reviewer for this concrete suggestion. Following the reviewer's recommendation, we conducted additional comparative experiments. The results confirm that the verification context employed by De-Anchoring is the most effective at detecting inconsistencies, compared against the two suggested alternative structures, Dependency-based chain and Execution-trace chain.

Specifically, De-Anchoring uses derived values as anchors and evaluates the consistency of the chain by simultaneously masking their duplicate occurrences. In contrast, Dependency-based chain evaluates consistency by simultaneously masking values that are mathematically interdependent with the anchors, rather than their duplicates. These include the operands in expressions that produce the anchor values and the positions where the anchors themselves serve as operands. Execution-trace chain evaluates consistency by masking all tokens within executable mathematical expression chunks (e.g., all tokens in \frac{2}{3} \times 60 = 40). All three methods differ only in how the chains are constructed, while employing the same scoring procedure for subsequent verification.

DatasetDe-Anchoring (BoN / AUROC)Dependency-based chain (BoN / AUROC)Execution-trace chains (BoN / AUROC)
GSM8K84.07 / 0.83076.11 / 0.76870.49 / 0.716
MATH69.77 / 0.79159.30 / 0.75058.14 / 0.683
Countdown67.50 / 0.73861.67 / 0.69561.67 / 0.651

LLaDA-8B-Instruct, 5 samples per problem, block 32, T=0.6. Evaluated on candidates for which all structures define at least one chain (GSM8K 90%, MATH 43%, Countdown 77% of candidates). AUROC: these candidates, correct vs. incorrect. BoN (%): among these candidates, in problems where they disagree in correctness.

The results demonstrate that De-Anchoring's anchor structure is the most robust. We illustrate the weaknesses of the two proposed alternatives through a concrete verification example below. This example represents a common error pattern observed in DLMs, where the generated trajectory contains an internal inconsistency. The question is: "Bobby has 5 fewer than 3 times as many video games as Brian. Brian has 20 but lost 5 before the comparison." The ground-truth answer is 40. Here, [M] denotes a contiguous masked span, which may consist of multiple tokens (e.g., "60" consists of two tokens).

Given reasoning trajectory

Now, we know that Bobby has 5 fewer than 3 times as many video games as Brian does. We calculate:
\[ 3 \times 20 - 5 = 60 \]

So, Bobby has 60 video games.

<answer>

\boxed{60}

De-Anchoring

Now, we know that Bobby has 5 fewer than 3 times as many video games as Brian does. We calculate:
\[ 3 \times 20 - 5 = [M] \]

So, Bobby has [M] video games.

<answer>

\boxed{[M]}

Dependency-based chain

Now, we know that Bobby has 5 fewer than 3 times as many video games as Brian does. We calculate:
\[ [M] \times [M] - [M] = [M] \]

So, Bobby has 60 video games.

<answer>

\boxed{[M]}

Execution-trace chain

Now, we know that Bobby has 5 fewer than 3 times as many video games as Brian does. We calculate:
\[ [M] \]

So, Bobby has 60 video games.

<answer>

\boxed{[M]}
De-Anchoring chainDependency-based chainExecution-trace chain
verification probability for erroneous "60"0.02%97.3%97.1%
Candidate score−2.159−0.012−0.020

Only De-Anchoring chain masks every repeated occurrence of erroneous 60; the alternatives leave "So, Bobby has 60 video games." visible, and the model copies the error back from it.

The results show that only De-Anchoring provides a strong signal for error detection. As illustrated in the example above, this is because the propagated erroneous value during generation remains in the future context, allowing the same error to be reinforced at the past masked span during verification. As reported in Section 6.1, visible contradictory anchors can provide false support for the same contradiction at the verified position. Consequently, the two alternatives fail to satisfy the "Verify the Full Anchor Chain" principle already analyzed in Section 6.1. Thus, De-Anchoring's anchor-chain structure is essential for preventing information leakage during verification.

Furthermore, both alternatives require heavy regular expressions to identify mathematical dependencies and executable expressions. Covering the diverse and complex cases where natural language and mathematical expressions are intertwined would require increasingly elaborate regex-based heuristics. Indeed, execution-trace chains can be constructed for only 44% of MATH candidates, compared to 80% for De-Anchoring. These findings demonstrate that De-Anchoring's simple rule achieves both simplicity and robustness.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Gentle Reminder of Our Rebuttal

Dear Reviewers and AC,

We sincerely thank you for your time and constructive feedback on our submission.

As the discussion period closes in three days, we would like to kindly follow up on our responses. We have replied to each review individually, and we would be grateful if you could let us know whether our responses address your concerns. If any questions remain, we would be glad to provide further clarification or additional results in the remaining time.

Thank you again for your effort and consideration.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.