acceptodds
Under review as a conference paper at ICLR 2027

TIPS: A Tiny Inference-time Policy for Scaling in Diffusion Language Models

Abstract

Masked diffusion language models expose an inference-time choice: which positions become visible after each denoising pass. Many existing decoding methods use hand-designed rules or require visibility to grow monotonically. We present a **T**iny **I**nference-time **P**olicy for **S**caling (**TIPS**), an approximately K-parameter controller that learns both unmasking and remasking while keeping the underlying language model frozen. TIPS uses visibility-aware confidence features and a shallow transformer to select which tokens should be visible, while a fixed schedule controls the expected visible fraction. We train the controller through supervised warmup followed by reinforcement learning from terminal correctness rewards. On GSM8K and MATH-500 with LLaDA-8B-Instruct, the full-canvas and two-block TIPS variants achieve and micro-averaged accuracy at a budget of decoder evaluations, exceeding the strongest baseline in our main comparison by and percentage points, respectively. Both variants are trained at a budget of evaluations and remain competitive at higher budgets without retraining. These results demonstrate that learning when to reveal and reconsider tokens can improve the accuracy–compute trade-off without updating the decoder backbone.

Primary area:
generative models
Keywords:
diffusion language models, inference-time scaling, masked diffusion language models, diffusion models, masked diffusion, discrete diffusion, language models, decoding

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_rsCMOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TIPS uses a roughly 629K-parameter, two-layer DiT controller to manage token visibility in a frozen LLaDA model. The policy can reveal tokens and remask unreliable ones; a logit shift aligns its expected visible fraction with a prescribed progress schedule. It is trained with supervised warmup followed by GRPO using terminal correctness rewards. On 1,819 GSM8K and MATH-500 questions, TIPSfull reaches 58.46% accuracy at 32 decoder calls, compared with 46.34% for WINO. The main results use three training seeds and three inference draws per seed. The paper also studies remasking, policy architecture, and transfer across models and block sizes.

Critical strengths
  1. The paper separates visibility control from the frozen model’s token prediction and allows already revealed tokens to be withdrawn. The policy is small, and its training procedure and visibility schedule are well specified.
  2. TIPS improves accuracy over several baselines at low and moderate decoding budgets. The main results use three training seeds and include ablations on remasking and supervised warmup.
Critical weaknesses
  1. The appendix states that Fast-dLLM’s threshold is tuned using test-set rollouts to reach a target mean NFE, whereas TIPS’s γ\gamma is selected using validation data. Although no answer labels are used, the baseline uses the test-prompt distribution for budget matching, making the protocols asymmetric. Fast-dLLM’s 4-call result is also imputed from its 8-call result.
  2. Evaluation is concentrated on one LLaDA backbone, mathematical tasks, and 256-token responses. Decoder-call counts do not directly measure wall-clock time; the reported policy time is about 8 ms but excludes the decoder and feature extraction, so the end-to-end inference benefit remains unclear.
  3. The paper does include TIPSblock and block-size transfers on LLaDA, but it does not evaluate a native block-diffusion backbone such as LLaDA2.X or SDAR. Since these models may have different blockwise denoising dynamics, the current results provide only indirect evidence that TIPS generalizes to this increasingly relevant model family.
Questions for authors
  1. Why is Fast-dLLM’s budget threshold tuned on test-set rollouts rather than the validation set already used in the paper? Does the main comparison remain favorable when budget matching uses validation data only? How does imputing the 4-call point from the 8-call result affect the low-budget conclusion?
5 comments

Thank you for identifying the calibration asymmetry and the limitations of decoder-call counts as an efficiency measure. We have corrected the Fast-dLLM protocol and conducted end-to-end timing and additional zero-shot transfer evaluations.

Unless otherwise stated, we use the test set of GSM8K and MATH-500, newly trained policy networks use seed 123 for training and accuracy at inference time is averaged over three samples.

1. Why was Fast-dLLM’s budget threshold tuned on test-set rollouts

While TIPS is parametrized directly in terms of decoding budget (i.e., how many steps are to be used), Fast-dLLM’s parametrization is less direct. Our approach has thus been to calibrate Fast-dLLM’s parameters so that, on average, it uses the desired number of steps. However, this calibration is data-dependent: calibrating on the validation set may result in different mean NFEs on the test set.

This validation-to-test drift in NFEs is why our original manuscript calibrates on the test set: so that the number of NFEs becomes comparable across inference methods. Our original reasoning was that this gained comparability was worth allowing Fast-dLLM to “see the test set” (which is something we expected to be purely detrimental to TIPS’ case). However, your question led us to do the actual comparison (table in point 2 below), which made us realize we were wrong on both counts: validation-calibrated test mean NFEs are close to nominal, and calibrating on test was not purely detrimental to TIPS’ case (validation-calibrated widened Fast-dLLM’s lead over TIPS at 64 NFE by 0.88 points; the old configuration helped Fast-dLLM by 4.07 points at 32 NFE).

In summary. Table 1 and the AUC will use validation-calibrated Fast-dLLM. Conclusions are unchanged for 8-32 calls. Fast-dLLM’s lead at 64 NFE grows slightly. TIPS remains the top score at 128 and 256, though too close to be called a winner. The test-calibrated numbers and both NFE columns will be presented in appendix.

2. Validation-only calibration of Fast-dLLM

We reran calibrations for Fast-dLLM using the validation set and the test set, and compare both calibration configurations below.

Nominal targetTest-calibrated test mean NFETest-calibrated test accuracy (%)Validation-calibrated test mean NFEValidation-calibrated test accuracy (%)Accuracy change (pp)
88.00017.488.00017.480.00
1616.41814.7315.80815.23+0.49
3232.47337.9330.73233.86-4.07
6464.20865.5960.94266.47+0.88
128128.22467.07124.41066.96-0.11
256256.00066.96256.00066.960.00

(Note minor discrepancies with the original Table 1 at nominal target 128 and 256 for test-calibrated test accuracy. While producing this table, we noticed that previous calibration saturated the mean NFE at 86.661, which is why both entries reported 67.01 test accuracy: we recalled that earlier in the project, we had decided to use the recommended Fast-dLLM threshold, i.e., 0.9, as going above that threshold had negligible effect on performances. Further note that Fast-dLLM uses argmax when determining its output: past a certain budget, it cannot further improve the answer. In essence, properly calibrating Fast-dLLM in that regime demands to have it do nothing during its final steps: we will discuss this in the appendix, and report the effective mean-NFE cap. In any case, we will replace the Table 1 results by the validation-calibrated test accuracy.)

As seen in the last column, the new calibration has minor impact on accuracy overall. The old approach (test-calibrated) was favorable to Fast-dLLM at 32 NFE, and the new one (validation-calibrated) makes Fast-dLLM slightly better than it was at 64 NFEs. This thus increases the gap between Fast-dLLM and TIPS at 64 NFEs. TIPS is still the winner for {8, 16, 32, 128, 256}.

3. The four-call Fast-dLLM entry

A fixed 256-token response decoded as eight sequential 32-token blocks requires at least one backbone evaluation per block. Its minimum is therefore eight calls. We will replace this imputation in the updated manuscript with N/A, and state the reason explicitly.

How does imputing the 4-call point from the 8-call result affect the low-budget conclusion?

From TIPS’ point of view, imputing the 4-call point does not affect the low-budget conclusion, since TIPS already loses against WINO in that extremely-low NFE regime. The imputation was advantageous to Fast-dLLM, inflating its AUC. Our current plan is to exclude 4-call from the AUC calculations for all models, but could alternatively use 0 for Fast-dLLM’s score 4-call score. We welcome the reviewer’s input on this point.

4. Domains, backbones, and response lengths

The submitted paper includes Dream-v0-Instruct-7B transfer (main-text Table 2; Appendix B.4.1, Table 23), frozen math-to-code transfer on HumanEval and MBPP (Appendix B.4.2, Table 24), and coding-trained controllers (Appendix B.4.3, Table 25). We agree that the main text should make this evidence easier to find and will add explicit pointers. We have additionally conducted the following frozen-policy evaluations.

Zero-shot transfer to LLaDA2.0-mini. We transfer the math-trained controller to LLaDA2.0-mini, without retuning. We use native BL32, applying TIPS within the active block. Here H denotes denoising steps per active block.

MethodMicro accuracy (%)
Frozen TIPS, H=421.13
Native static, H=414.90
Frozen TIPS, H=838.39
Native static, H=839.97
Frozen TIPS, H=1645.52
Native static, H=1644.42

At H=4, TIPS improves accuracy by 6.23 points. Results at higher H are less clear. Nonetheless, this supplies evidence of useful transfer to a native block-diffusion backbone.

Zero-shot transfer to a longer response canvas. We doubled the canvas from 256 to 512 tokens. No length-specific retraining or tuning was used.

Calls TTIPS, 256 tokens (%)Frozen TIPS, 512 tokens (%)Confidence, 512 tokens (%)Fast-dLLM BL32, 512 tokens (%)
1646.8843.2810.509.35
3258.3151.9026.0012.37
6462.5857.1728.5358.55
12865.8659.8115.8971.25

Frozen TIPS demonstrates effective zero-shot transfer to twice its training canvas length, substantially outperforming confidence-based decoding across all budgets and Fast-dLLM at lower budgets.

Together, the backbone-transfer and longer-canvas evaluations provide evidence of transfer beyond the original backbone and response length without controller retraining or retuning, with the clearest gains at low per-block budgets.

5. End-to-end inference efficiency

We agree that the previously reported controller-only timing omits material costs, and measure end-to-end generation on a single H100 80GB, from prompt processing to output. Latency uses a length-stratified fixed subset of 256 prompts.

The top table contains single-request results while the bottom table contains batch-size 8 requests (so the bottom table’s numbers should be divided by 8 when comparing to the top table). All methods use decoder budget 32, and Fast-dLLM thresholds are calibrated on validation set.

MethodAccuracy (%)End-to-end latency (s)
Confidence37.990.637
WINO, full canvas42.171.544
TIPS_full58.310.798
TIPS_block, BL12859.250.799
Fast-dLLM Cache Off33.860.626
Fast-dLLM Prefix KV Cache32.710.779
Fast-dLLM DualCache25.510.930
MethodAccuracy (%)End-to-end latency per batch (s)
Confidence37.930.820
WINO, full canvas40.301.640
TIPS_full58.620.843
TIPS_block, BL12859.430.844
Fast-dLLM Cache Off33.641.492
Fast-dLLM Prefix KV Cache33.320.733
Fast-dLLM DualCache25.780.558

At single-request, TIPS_full improves accuracy by 20.32 percentage points over confidence decoding with 25.3% more end-to-end time. We suspect at single-requests, caching’s execution overhead outweighs the saved computation. For batched requests, with an additional 111ms overhead we can outperform Fast-dLLM with prefix caching by up to 26.11%, while Fast-dLLM with DualCache saves even more time, its accuracy deteriorates significantly.

Additionally, we believe it is important to note that latency depends on both the decoding policy and its execution backend. While TIPS offers faster inference by reducing the decoder evaluations needed to reach a target accuracy, our implementation uses eager execution and leaves concrete optimization opportunities e.g. compile the controller, fuse operations, etc. These optimizations could improve TIPS’s accuracy–latency trade-off significantly. However, as inference optimization is out of the scope of this paper, we leave this as a promising direction for future work.

Thank you again for prompting these protocol and evaluation improvements. We hope the corrected calibration, removal of the unsupported four-call point, direct native-backbone evaluation, complete timings, and experiments already included in the submission address your main concerns. We would appreciate your reconsideration of the rating in light of this evidence, and welcome clarification on any remaining concerns so that we can address them during the discussion.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer,

Thank you again for your detailed and constructive review.

We wanted to kindly follow up as the discussion period approaches its conclusion. In response to your concerns, we have corrected the Fast-dLLM calibration protocol, added end-to-end latency measurements, and evaluated zero-shot transfer to LLaDA2.0-mini with native BL32 decoding. These results and the corresponding clarifications are now incorporated into the revised manuscript.

We would greatly appreciate your feedback on whether these additions adequately address your main concerns, or whether any substantive issues remain unresolved.

Thank you for your time and consideration.

Reviewer_YN5oOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a method to learn the optimal decoding order of DLLMs. It introduces a 600k parameter controller that takes confidence, maksing and scheduling (timestep) information from the outputs of the DLLM forward pass to make a binary decison at each token positon of whether it should be visible in the next timestep or not. This formulation is unique as it allows it to model remasking decisons as well, which makes it distinct from prior work that only learned to unmask mask tokens. Additonally only the controller is trained, keeping the DLLM frozen, which makes it easier to adopt and transfer. They evaluate on Math and GSM8K and show that they get improvements over previous baselines, especially at less number of forward pass calls of the model.

Critical strengths
  1. The formulation is unique, novel and beautifully simple, allowing the modelling of both unmasking and remasking with a single policy through a binary output at each position.

  2. The numbers on Math and GSM8K are strong, especially at low number of model forward pass calls (NFEs). This is promising evidence of the efficacy of this approach.

Critical weaknesses
  1. The paper however seems to focus only on math benchmarks which is a very narrow setting and since the policy being trained is quite small, it is unclear whether it could be powerful enough for general usage or more difficult settings like code or chat, and could be overfitting to this simple narrow setting. Requirement of training a new policy for each domain, while feasible due to it's small size, would hurt the adoption of this method. It would be nice to see at the very least if this approach can genralize to more domains beyond Math

  2. The evaluated configuration are only at high block sizes, which isn't practical for today's DLLMs as they mostly run inference at much smaller blocksizes. Seeing how the method performs at more practical block sizes is important and is unfortunately missing. Also, my understanding is that the Jazbec et al, the paper's main baseline was trained at BL=32, and in the paper it is evaluated at BL=256, while TIPS is trained at the block sizes it evaluates it on. Thus the comparison in Table 1 may not be a fair representation of the baseline's performance due to the training-inference mismatch.

Questions for authors

Please do address the weaknesses above. And also, the point on Jazbec et al's training and infernece settings as it is unclear in the paper. Overall I believe the paper has a lot of promise and I really like the novel formulation that allows the joint modelling of unmaksing and remasking, however I believe some more expeirments are needed to experimentally validate the method. Hence I am initially giving the paper a weak reject and am open to raising my score after the author's response.

After Rebuttal

The author's response has addressed my concerns regarding the missing experiments, which are already present in the appendix, and regarding the comparison with Jazbec et al at smaller block sizes, for which they have provided additonal results and clarified the inference and training settings in the paper. Therefore, I'm raising my score to a 3.

3 comments

Thank you for the encouraging assessment of the joint unmasking/remasking formulation and for specifying the additional evidence that would strengthen the paper. We address both generalization and the BL32 comparison below.

Unless otherwise stated, we use the test set of GSM8K and MATH-500, newly trained policy networks use seed 123 for training and accuracy at inference time is averaged over three samples.

1. Generalization beyond the original mathematical setting

Appendix B.4.2, Table 24 evaluates the frozen math-trained policy on HumanEval and MBPP; Appendix B.4.3, Table 25 evaluates coding-trained controllers. The coding-trained full policy improves the low-budget results on both benchmarks.

The submission also transfers the LLaDA-trained controller to Dream-v0-Instruct-7B without controller retraining (main-text Table 2; Appendix B.4.1, Table 23). Existing block-size transfer is also reported in Appendix B.4.4, Table 27. We will add explicit main-text pointers to these experiments in our updated version of the manuscript.

Zero-shot transfer to LLaDA2.0-mini. We also transferred the same math-trained controller to LLaDA2.0-mini, without retraining. We use native BL32, applying TIPS within the active block. Here H denotes denoising steps per active block.

MethodMicro accuracy (%)
Frozen TIPS, H=421.13
Native static, H=414.90
Frozen TIPS, H=838.39
Native static, H=839.97
Frozen TIPS, H=1645.52
Native static, H=1644.42

At H=4, TIPS improves accuracy by 6.23 points. Results at higher H are less clear. Nonetheless, this supplies evidence of useful transfer to a native block-diffusion backbone.

Zero-shot transfer to a longer response canvas. We doubled the canvas from 256 to 512 tokens. No length-specific retraining or tuning was used.

Calls TTIPS, 256 tokens (%)Frozen TIPS, 512 tokens (%)Confidence, 512 tokens (%)Fast-dLLM BL32, 512 tokens (%)
1646.8843.2810.509.35
3258.3151.9026.0012.37
6462.5857.1728.5358.55
12865.8659.8115.8971.25

Frozen TIPS demonstrates effective zero-shot transfer to twice its training canvas length, substantially outperforming confidence-based decoding across all budgets and Fast-dLLM at lower budgets.

Together, the backbone-transfer and longer-canvas evaluations provide evidence of transfer beyond the original backbone and response length without controller retraining or retuning.

2. Practical block sizes and the Jazbec et al. comparison

We must first clarify that the Jazbec et al results of our paper’s Table 1 are trained at BL=256 with expert steering, and are evaluated in the same context. Table 4 (appendix) also has results without expert steering.

However, the reviewer’s intuition that something was amiss at BL=32 was funded. Following the reviewer’s comment, we have trained a fresh TIPS_block policy for eight 32-token blocks, and retrained the Jazbec et al. baseline for the same block size:

Target callsTIPS BL32 accuracy (%)Jazbec BL32 accuracy (%)Best TIPS (Table 1)
817.4917.4829.12
1625.3824.6348.05
3243.3948.0358.46
6457.5463.6464.49
12864.7263.5167.12
25667.7868.1368.18

Jazbec BL=32 gets 48.03 at 32 NFEs, notably higher than the 12.39 reported in Table 1 for Jazbec BL=256+expert steering. We believe that this configuration is worth a separate entry in the revised Table 1 (and a separate curve in Figure 1). This will also be reflected in the abstract: among other things, it now beats WINO at 32 NFE.

TIPS_block at BL32 outperforms Jazbec at 128 calls and achieving comparable accuracy at 8, 16, and 256 calls. (At 8 calls total, neither model make a decision as the whole block is generated in one shot. At 16, they take decisions only once, which may explain why it is comparable. Conversely, at 256, both basically saturate the base denoiser.) Jazbec BL=32 wins over TIPS_block BL=32 at 32 and 64 NFEs. The results for TIPS at BL32 were obtained without tuning, leaving room for further improvements.

We also note that matching block sizes is not strictly necessary for a fair comparison: block size is itself an inference-time design choice, and different methods may perform best under different configurations. In this spirit, we report in the last column the corresponding best TIPS score from Table 1, which beats Jazbec BL32 at all NFEs (though the 256 NFEs are within noise).

Thank you again for your willingness to reconsider your assessment in light of additional evidence. We hope the new backbone-transfer and longer-canvas evaluations, together with the direct BL32 comparison, address the main questions, and would appreciate your reconsideration of the paper with these results.

Thank you for the clarifications and additonal results, especially the analysis of TIPS at a more practical block size along with the corresponding comparison with Jazbec et al. I apologize for missing the results in the appendix on the zero-shot transfer to coding tasks, and I'm glad to see it already available in the paper. I am happy to raise my score, based on the response, and I believe that this work is quite interesting and relevant to the community.

Dear Reviewer,

Thank you for your thoughtful follow-up and for reconsidering your assessment. We sincerely appreciate your constructive feedback and willingness to raise your score in light of our additional experiments and clarifications.

Reviewer_E5r5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes to train a small controller network to jointly predict unmasking tokens and remasking tokens. The controller is trained via supervised warmup + RLVR, and the decoding is implicitly controlled by a pre-defined expected visibility schedule. Results show that the proposed method can improve accuracy in medium/low decoding budget.

Critical strengths

The performance gain at medium/low budget is noticeable and experiments are well-down to support the key design choices.

Critical weaknesses

Efficiency is only measured by decoder-call count and no end-to-end performance accuracy–latency numbers are provided. The actual implication of the work seems to be limited as it mainly shows gains on certain math tasks and particular budget ranges.

Questions for authors

Why the performance degrades with more budgets? Is this aligned with your expectation / theory? Why do you think such a small network can do the controller job well and do you think some of the empirical negative results are limited by the controller size and the way it is trained?

3 comments

Thank you for the thoughtful comments and for recognizing the gains at low and moderate decoding budgets. We have conducted end-to-end timing and larger-controller experiments to address your questions. We also clarify the scope of the existing coding results and the behavior at higher budgets.

Unless otherwise stated, we use the test set of GSM8K and MATH-500, newly trained policy networks use seed 123 for training and accuracy at inference time is averaged over three samples.

1. End-to-end accuracy and latency

We now measure end-to-end generation on a single H100 80GB, from prompt processing to output. Latency uses a length-stratified fixed subset of 256 prompts.

The top table contains single-request results while the bottom table contains batch-size 8 requests (so the bottom table’s numbers should be divided by 8 when comparing to the top table). All methods use decoder budget 32, and Fast-dLLM thresholds are calibrated on validation set (see discussions with reviewer rsCM for calibration details).

MethodAccuracy (%)End-to-end latency (s)
Confidence37.990.637
WINO, full canvas42.171.544
TIPS_full58.310.798
TIPS_block, BL12859.250.799
Fast-dLLM Cache Off33.860.626
Fast-dLLM Prefix KV Cache32.710.779
Fast-dLLM DualCache25.510.930
MethodAccuracy (%)End-to-end latency per batch (s)
Confidence37.930.820
WINO, full canvas40.301.640
TIPS_full58.620.843
TIPS_block, BL12859.430.844
Fast-dLLM Cache Off33.641.492
Fast-dLLM Prefix KV Cache33.320.733
Fast-dLLM DualCache25.780.558

At single-request, TIPS_full improves accuracy by 20.32 percentage points over confidence decoding with 25.3% more end-to-end time. We suspect at single-requests, caching’s execution overhead outweighs the saved computation. For batched requests, with an additional 111ms overhead we can outperform Fast-dLLM with prefix caching by up to 26.11%, while Fast-dLLM with DualCache saves even more time, its accuracy deteriorates significantly.

Additionally, we believe it is important to note that latency depends on both the decoding policy and its execution backend. Our implementation uses eager execution and leaves concrete optimization opportunities throughout the pipeline, from feature generation and calibration to sampling. Examples may include compiling the controller, fusing operations, etc. These optimizations could improve TIPS’s accuracy–latency trade-off significantly. However, as inference optimization is out of the scope of this paper, we leave this as a promising direction for future work.

2. On domains beyond mathematics

The submitted manuscript also evaluates TIPS on coding tasks HumanEval and MBPP, including frozen math-to-code transfer (Appendix B.4.2, Table 24) and coding-trained controllers (Appendix B.4.3, Table 25). We recognize that these results should have been more visible, and will add explicit pointers in the main text.

3. Budget range and the apparent performance degradation

We’re unsure what the reviewer meant here: on the main benchmark (Table 1), absolute accuracy rises monotonically with budget.

If the reviewer refers to Tables 24 (TIPS_full) and/or 25 (TIPS_block) at T=64, then we hypothesize that the HumanEval drop is due to a train-test mismatch in domain (Table 24) or decoding budget (both tables). Whether or not this was what the reviewer had in mind, we will discuss these points in the manuscript.

If the reviewer is referring to the relative performance of TIPS compared to the baselines as the number of NFEs increases, then we agree that TIPS’ margin is largest at low and moderate budgets (T∈{8,16,32}T\in\{8,16,32\} in Figure 1): at high budget, it remains competitive, but less dominant. A central motivation why dLLMs are being considered today is to decode in fewer passes than there are tokens: the regime where TIPS has a larger margin is where that value is. As to why the gap closes, we believe that TIPS’ advantage comes from deciding which tokens to decode/remask next. As TT approaches LL, each step reveals roughly one token, so there is less to decide. The (frozen) decoder’s own ceiling is what caps performances.

Otherwise, if the reviewer meant something else, then we would kindly ask them to clarify their comment regarding “performance degrades with more budgets”.

4. Why a small controller can work, and the roles of capacity and training

Our interpretation is that TIPS solves quite a narrow problem. The large frozen decoder has already computed context-dependent token probabilities. TIPS only converts those beliefs into binary actions. The two input features are also directly aligned with our action space: the active-token log-odds measures support for that token, while its log-probability ratio against the strongest alternative indicates the decoder favoring a replacement (Section 3, Equations 4-6). Furthermore, the external schedule further separates how much of the canvas should be visible from which positions should be visible.

The architecture ablation (Appendix B.2.7, Table 9) finds that cross-position attention improves on a parameter-matched per-token MLP. The feature-width study (Appendix B.2.4, Table 6) finds no benefit from expanding two input features to ten or fifty.

We also trained a four-layer, 1.2M-parameter controller and compared it to our two-layer controller. Results are below:

Decoder budgetTwo layers (%)Four layers (%)
416.7121.39
829.4539.34
1646.8853.44
3258.3160.22
6462.5863.99
12865.8664.98
25666.9665.29

Although two-layer TIPS already outperforms competing methods at 8-32 budgets, these results suggest that additional controller capacity can further improve performance in this regime, e.g. improving accuracy by 9.89 and 6.56 points at 8 and 16 calls, respectively. However, the gains diminish as budget increases, leading the four-layer TIPS to perform slightly worse at 128 and 256. Therefore, larger capacity is not uniformly beneficial across budgets.

Importantly, the larger 4-layer controller remains compact: its 1.2M parameters still amount to only 0.015% of the frozen 8B decoder. Thus, the improvement strengthens the evidence that a small learned controller is sufficient, while suggesting that the original two-layer architecture does not exhaust the available gains.

However, it is important to note that our original two-layer version already outperforms almost all methods in the 8-32 budget regime. Another thing to note is that the larger controller remains compact: its 1.2M parameters amount to only 0.015% of the frozen 8B decoder. Thus, the improvement strengthens the evidence that a small learned controller is sufficient, while suggesting that the original two-layer architecture does not exhaust the available gains.

Re: the way it is trained, our RL objective uses terminal correctness rewards, leaving denser training feedback and shared controllers trained across domains as promising future work. The present contribution is that a compact controller already produces substantial gains by learning visibility decisions from the frozen decoder's beliefs.

Thank you again for the constructive feedback. We believe the end-to-end measurements and direct controller-capacity comparison address the main concerns raised in your review. We would appreciate your reconsideration of the rating in light of these new results as well as the experiments already included in the submission. If any concerns continue to support your current assessment, please let us know which issues remain unresolved so that we can address them during the discussion.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer,

Thank you again for your thoughtful review.

As the discussion period approaches its conclusion, we wanted to kindly follow up on our response. We have added end-to-end accuracy–latency measurements and a direct controller-capacity comparison, alongside clarifications on generalization beyond mathematics and the behavior at higher decoding budgets. These additions are now incorporated into the revised manuscript.

We would appreciate your feedback on whether the new evidence addresses your original concerns, and would be happy to clarify any remaining questions.

Thank you for your time and consideration.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your constructive feedback and helpful experimental suggestions. We have updated our manuscript to incorporate the clarifications and additional experiments discussed in the reviews and our responses. The main updates include:

  • End-to-end efficiency: accuracy and latency comparisons at batch sizes 1 and 8 (Table 2).
  • Baseline comparisons: validation-calibrated Fast-dLLM results in the main comparison and a matched BL32 comparison with Jazbec et al. (Section 4.1; Appendix B.1.1, Table 6).
  • Generalization: transfer experiments with LLaDA2.0-mini and 512-token responses, alongside expanded discussion of the HumanEval and MBPP experiments (Section 4.3).
  • Controller capacity: a comparison of two- and four-layer controllers across decoding budgets (Section 4.4 and Table 12).

The main text now summarizes these additions, with further results and experimental details in the appendix. Revisions are highlighted in red for ease of review.

We believe the revised manuscript and additional experiments directly address the concerns raised during review. We hope the reviewers will reassess the paper in light of these results and reconsider their original ratings.

We appreciate everyone’s time and constructive feedback, which have helped us strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.