acceptodds
Under review as a conference paper at ICLR 2027

Self-Regularized Flow Language Models

12 views

Abstract

We introduce SERF (Self-Regularized Flow Language Models), a simple self-supervised framework for improving continuous diffusion language models by exploiting the denoising trajectory itself. While continuous flow-based language models offer a promising alternative to autoregressive generation, their training objectives typically rely on supervision from clean target tokens, leaving the denoising trajectory underexploited as a learning signal. SERF addresses this limitation through noisy-to-clean self-distillation: an exponential moving average (EMA) teacher receives a cleaner version of the student's noisy input along the same interpolation path and provides additional supervision. We investigate two variants: SERF-L, which aligns the student's intermediate representations with deeper teacher representations, and SERF-P, which aligns their output token distributions. Both augment the standard training objective without requiring external pretrained models or modifying the inference procedure. Experiments across three continuous language-model paradigms (FLM, ELF, and S-FLM) demonstrate substantial improvements in training efficiency, generation quality, and few-step generation. In particular, SERF-L achieves up to 7.7× wall-clock training acceleration on OpenWebText, improves MAUVE by 24.1% and GSM8K accuracy by 26.0%, and matches baseline performance with much fewer sampling steps. These results establish noisy-to-clean self-distillation as an effective approach to improving continuous diffusion language models. Code and pretrained models will be released.

Primary area:
generative models
Keywords:
diffusion language models, self-supervised learning

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_j2UROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SERF adds a self-distillation regularizer to continuous flow language models. An EMA teacher sees a cleaner point on the same interpolation path as the student and provides an auxiliary target. Two variants are studied: SERF-L aligns an early student block with a deeper teacher block via a projection head and cosine loss; SERF-P matches the student's output token distribution to the teacher's. The method is applied to FLM (one-hot), S-FLM (hypersphere), and ELF (frozen T5 latents). At a matched step budget, both variants improve MAUVE on OpenWebText; SERF-L trained for 300K steps exceeds the released 1M-step FLM at matched entropy, and SERF-L improves GSM8K accuracy of S-FLM by 4.6 points.

Critical strengths
  1. Presentation: The method and figures are easy to follow, and the design is well motivated for the text.
  2. The regularizer is validated on three representation types (one-hot, spherical, latent), and comparisons are made along quality–entropy fronts rather than at a single temperature.
  3. Experimentally, SERF-L-300K beats FLM-1M at every sampling budget on matched-entropy fronts.
Critical weaknesses
  1. Positioning.
    • SERF-L is essentially the objective of SRA (Jiang et al., ICLR 2026) and Self-Flow (Chefer et al., ICML 2026): EMA teacher at the lower noise level, early student block mapped through an MLP head and aligned to a deeper teacher block with a cosine loss, EMA decay. Self-Flow already evaluates the teacher at min(t, s) for two independently drawn times, which is exactly the "random offset" of Sec. 3.3.
    • SERF-P, in turn, is consistency training (Song et al., 2023) applied to the denoiser output: shared noise, an EMA stop-gradient target at the cleaner point, and a CE anchor. The paper cites SRA and Self-Flow as inspiration but never states this correspondence, and the consistency-model connection is absent.
    • Suggestion. I do not consider the transfer to text disqualifying; the empirical findings are useful. But please add a short paragraph or table that contrasts SERF with SRA, Self-Flow, LayerSync, and consistency training point by point, and state precisely what is new.

Jiang, Dengyang, et al. "Representation alignment for diffusion transformers without external components." The Fourteenth International Conference on Learning Representations. 2026.

Song, Yang, and Prafulla Dhariwal. "Improved techniques for training consistency models." International Conference on Learning Representations. Vol. 2024. 2024.

  1. Under shared noise, Eq. (4) gives xt=((1−t)/(1−s))⋅xs+((t−s)/(1−s))⋅x1x_t = ((1−t)/(1−s))·x_s + ((t−s)/(1−s))·x_1, so for t>st > s the pair (xs,xt)(x_s, x_t) determines x1x_1. The tower property 4E[E[x_1|x_t]|x_s] = E[x_1|x_s]$ therefore does not hold, and the minimizer of L_FLM + λ·L_P at fixed x_s is biased toward sharper predictions. The "soft target that reflects plausible alternatives" narrative ( in p. 6) is therefore incomplete.

  2. The random offset also reweights the main loss in time, and the ablation cannot separate the two effects. Algorithms 1–2 evaluate L_FLM at s = min(µ, κ), i.e. s ~ Beta(1, 2) with mean 1/3 rather than U(0, 1). SERF therefore changes the time distribution of the main loss in addition to adding a distillation term. In Fig. 5 (Left), SERF-P with the fixed offset (uniform s) is indistinguishable from FLM, while the random offset helps; time reweighting explains this as well. The fixed-offset control is itself degenerate (with δ = 0.8, 80% of teacher inputs are fully clean), so it does not isolate the offset choice either.

    • Suggestion. If time and computation allows, I would suggest: Add (i) FLM with s ~ Beta(1, 2) and λ = 0, (ii) SERF with L_FLM evaluated at an independent uniform s, and (iii) an EMA self-distillation control with t = s. These would make the attribution convincing.
  3. The 128-step entropy-matched Gen. PPL is an extrapolation. At 128 steps, both FLM and SERF-L have sampled entropies above 5.44 at all three temperatures, so "interpolating" to H⋆ = 5.44 extrapolates outside the sweep. The result is sensitive to the fitting function; a quadratic fit in log space would give SERF-L a higher value than FLM.

    • Suggestion. Compare within the observed entropy overlap (or extend the τ sweep until the fronts bracket 5.44); show Gen. PPL–entropy Pareto fronts. And it would be better to also show FLM results at 1024 steps as presented in their paper.
Questions for authors
  1. Does the EMA-teacher alignment remain effective when the token embedding space is learned jointly with the denoiser (e.g. LangFlow/RePlaid)? In SERF both FLM (one-hot) and ELF (frozen T5) fix the input space; with learned embeddings the teacher's embedding table lags the student's, and the SERF-L target may drift.

RePlaid: Yang, Zhihan, et al. "Continuous diffusion scales competitively with discrete diffusion for language." arXiv preprint arXiv:2605.18530 (2026).

Langflow: Chen, Yuxin, et al. "Langflow: Continuous diffusion rivals discrete in language modeling." arXiv preprint arXiv:2604.11748 (2026).

  1. Missing controls: The introduction argues that text lacks a suitable external representation encoder. Why are BERT-style encoders or hidden states of an autoregressive LM ruled out as REPA-style teachers? Additionally, it would be interesting to see controls on the EMA teacher against a frozen pretrained teacher, a stop-gradient copy of the student, or shared versus independent noise. One or two of these at 100K steps would suffice.

  2. Implementation details: Are samples drawn from EMA or raw weights, for SERF and for the baselines? How is the shared-noise pair (x_s, x_t) constructed for S-FLM and for ELF's native schedule, where Eq. (4) does not apply directly?

Additional comments and suggestions are given under Weaknesses.

10 comments

We thank the reviewer for the detailed and constructive review, and for recognizing that the method is clear and well motivated for text, that it is validated across three representation types with comparisons along quality–entropy fronts, and that SERF-L-300K outperforms FLM-1M at every sampling budget. We also appreciate and share the reviewer's view that SERF's empirical findings are useful. Following the reviewer's suggestions, we report new experiments and analyses:

  • a point-by-point comparison with SRA, Self-Flow, LayerSync and consistency training, added in Appendix A.4 (W1),
  • a label-smoothing analysis that bounds the deviation of the SERF-P target from the posterior (W2),
  • comparisons with uniform label smoothing and an independent-noise teacher (W2),
  • all three suggested attribution controls: time reweighting alone, a uniform student time, and t=st = s (W3),
  • and an extended temperature sweep so that the entropy-matched Gen. PPL is interpolated rather than extrapolated, with results robust to the fitting function (W4).

We address each point below, and would kindly ask the reviewer to reconsider their assessment in light of these results.

W1 Positioning/novelty

(This reply is identical to our response to Reviewer K6iV.)

We thank the reviewer for highlighting these connections. The original submission explicitly builds on image-domain self-distillation, citing SRA and Self-Flow in Secs. 1, 2, and 4.4. We have now expanded this discussion with a point-by-point comparison with SRA, Self-Flow, LayerSync, and consistency training in Appendix A.4, and revised the remark in Sec. 3.2.2.

SERF-L and transfer to language. SERF-L follows SRA and Self-Flow’s principle of aligning an early student layer with a deeper EMA-teacher layer evaluated at lower noise. Our random noise-level pairing likewise follows the cleaner-teacher principle of Self-Flow. We do not claim these principles as new.

Our contribution is to establish their effectiveness for continuous diffusion language models, which is not guaranteed by their success in images. For example, a motivation for image representation alignment is to improve semantic features learned through denoising pixels or VAE latents, whereas token prediction already directly supervises discrete linguistic units. Whether intermediate alignment still provides useful additional supervision, therefore, merits investigation. We demonstrate benefits across three substantially different representations—one-hot vectors, jointly learned hyperspherical embeddings, and frozen T5 latents—and two tasks, unconditional generation and mathematical reasoning.

SERF-P and consistency training. We agree that SERF-P shares the noisy-student/cleaner-teacher construction of consistency training (CT). The quantities being learned, however, have different mathematical roles. Consistency training learns an endpoint mapping that is invariant along a probability-flow ODE trajectory, enabling direct few-step generation. SERF-P instead distills categorical token distributions to regularize the standard denoiser, retaining clean-token cross-entropy and the original sampler.

AspectSERF-PStandard CT
Matched outputToken probability distributionPredicted clean sample
ObjectiveClean-token CE + λ × teacher CEConsistency loss, with an endpoint boundary condition
Student–teacher gapRandom, potentially largeAdjacent noise levels, with a progressively refined grid
SamplingOriginal ODE samplerDirect one- or few-step generation
PurposeImprove denoiser trainingLearn a few-step generative mapping

Our focus is on the effect of this loss as an auxiliary regularizer on training efficiency and generation quality, while retaining the original denoising objective and sampler. We do not claim that SERF-P learns a consistency function. We address the bias introduced by its teacher targets separately in our response on the training target.

What is new. Our contribution is the systematic investigation of noisy-to-clean self-distillation for continuous diffusion language models, establishing its benefits for training efficiency and generation quality across three representations and two tasks. SERF-P additionally studies distillation of categorical token distributions, requiring neither a projection head nor the selection of student and teacher layers. The revised discussion distinguishes these contributions from the self-distillation principles inherited from prior work.

W2. Tower property and the soft-target motivation

We thank the reviewer for this analysis. The reviewer is correct that, under shared noise, the pair (xs,xt)(x_s, x_t) determines x1x_1, so the teacher's prediction does not satisfy the tower property. Below we show that SERF-P is a form of label smoothing, which supports our soft-target claim: every per-sample target places probability on plausible alternative tokens rather than only on x1x_1. We further show that the resulting deviation from the posterior is bounded by λ, and that the shared-noise teacher is nonetheless the better design.

Label-smoothing view. Up to the constant factor (1+λ)(1+\lambda), LFLM+λLPL_\text{FLM} + \lambda L_P is cross-entropy against the target (1−α) x1+α q(1-\alpha)\,x_1 + \alpha\,q, with α=λ/(1+λ)\alpha = \lambda/(1+\lambda), i.e., label smoothing with the teacher's prediction qq in place of the uniform distribution. Each per-sample target is softer than the one-hot label and assigns probability to alternative tokens that the teacher considers plausible given a less corrupted context. Unlike the uniform distribution, however, qq is correlated with x1x_1 through the shared noise, which is the source of the bias.

The bias is bounded. At a fixed xsx_s, the minimizer is p⋆=(1−α) E[x1∣xs]+α E[q∣xs]p^\star = (1-\alpha)\,\mathbb{E}[x_1|x_s] + \alpha\,\mathbb{E}[q|x_s], so its deviation from the posterior is at most α≈0.09\alpha \approx 0.09 at λ=0.1\lambda = 0.1.

The teacher's target is what helps. Two ablations show that the gains come from the shared-noise teacher rather than from smoothing alone. First, replacing qq with the uniform distribution, i.e., standard label smoothing with the same α\alpha, performs worse than SERF-P, and even below FLM. Second, building the teacher's input from independent noise, xt=t x1+(1−t) e′x_t = t\,x_1 + (1-t)\,e', under which the pair (xs,xt)(x_s, x_t) no longer determines x1x_1, also performs worse. This supports our choice of a teacher on the same interpolation path.

MethodSmoothing distributionMAUVE (↑)
FLM–0.237 ± 0.009
+ Label smoothingUniform0.212 ± 0.014
+ SERF-P, independent noiseTeacher qq, independent noise0.291 ± 0.017
+ SERF-PTeacher qq, shared noise0.325 ± 0.016

OWT, 50K training steps, 1024 sampling steps, same α\alpha for all variants. Best MAUVE over τ∈{0.95,1.0,1.05}\tau \in \{0.95, 1.0, 1.05\}, ± std over 5 MAUVE seeds.

We have added these controls to Appendix B.

W3. Time reweighting vs. distillation

We thank the reviewer for this observation, which is correct: with s=min⁡(μ,κ)s = \min(\mu, \kappa), SERF also shifts the time distribution of LFLML_\text{FLM} to s∼Beta(1,2)s \sim \text{Beta}(1, 2). We ran the suggested controls (SERF-L, OWT, 50K steps). For (ii), instead of evaluating LFLML_\text{FLM} in a separate forward pass at an independent s′s', we sample the student time uniformly, s∼U(0,1)s \sim U(0, 1), and the teacher time as t∼U(s,1)t \sim U(s, 1). This removes the Beta(1, 2) reweighting from both losses while keeping the same conditional teacher time given ss as the random offset, so it isolates exactly the effect in question without requiring an extra forward pass.

RunLFLML_\text{FLM} time ssTeacher time ttBest MAUVE (↑)
FLMU(0,1)U(0, 1)–0.237 ± 0.009
(i) FLM, λ = 0Beta(1, 2)–0.238 ± 0.022
(ii) SERF-L, uniform student timeU(0,1)U(0, 1)U(s,1)U(s, 1)0.268 ± 0.016
(iii) EMA self-distillationBeta(1, 2)t=st = s0.246 ± 0.010
SERF-LBeta(1, 2)max⁡(μ,κ)\max(\mu, \kappa)0.270 ± 0.017

OWT, 50K training steps, 1024 sampling steps. Best MAUVE over τ∈{0.95,1.0,1.05}\tau \in \{0.95, 1.0, 1.05\}, ± std over 5 MAUVE seeds.

Time reweighting alone (i) does not improve MAUVE over FLM. Conversely, removing the reweighting (ii) retains nearly the full gain of SERF-L, while removing the noise gap (iii) loses most of it. The gains of SERF-L therefore come from distillation with a cleaner teacher, not from the change in the time distribution of LFLML_\text{FLM}.

We agree that the fixed offset with δ = 0.8 is a degenerate control for SERF-P, since most teacher inputs are fully clean. For SERF-L, however, the fixed offset uses a uniform student time and already performs on par with the random offset (Fig. 5, left), consistent with control (ii). We have added these controls to Appendix B, and now state this reweighting explicitly in Sec. 3.3.

W4. Extrapolated Gen. PPL at 128 steps

The reviewer is correct: at 128 steps our sweep τ∈{0.95,1.0,1.05}\tau \in \{0.95, 1.0, 1.05\} gave coldest entropies H=5.47H = 5.47 (FLM) and 5.515.51 (SERF-L), both above H⋆=5.44H^\star = 5.44, so the value was extrapolated. We have run τ=0.9\tau = 0.9, giving H=5.34H = 5.34 and 5.405.40: both sweeps now bracket H⋆H^\star.

To reassure the reviewer that our conclusion does not depend on the choice of interpolation, we read Gen. PPL at H⋆H^\star three ways: piecewise linear between the two straddling temperatures (the paper's scheme), a least-squares line a+bHa + bH over the sweep, and a least-squares quadratic a+bH+cH2a + bH + cH^2.

Stepsτ\tau sweepPiecewise linear (paper's scheme)LinearQuadratic
640.85, 0.9, 0.95, 1.086.5 / 82.388.3 / 80.885.3 / 81.2
1280.9, 0.95, 1.0, 1.0578.8 / 75.480.1 / 74.377.8 / 74.1
2560.95, 1.0, 1.0575.7 / 72.676.3 / 72.575.1 / 72.0
5120.95, 1.0, 1.0574.7 / 71.775.7 / 73.274.4 / 71.7
10240.8 to 1.0574.9 / 72.972.0 / 70.974.5 / 72.5

Gen. PPL (↓) as FLM-1M / SERF-L-300K at H⋆=5.44H^\star = 5.44, same scheme and same temperatures for both models in every cell, 5000 samples per temperature.

Under all three schemes, SERF-L achieves lower Gen. PPL than FLM at every sampling budget.

On the last point, Fig. 6 shows the full MAUVE– and Gen. PPL–entropy fronts. The FLM Gen. PPL values in Fig. 6 at τ=1\tau = 1 match those reported in the FLM paper, and we will add the raw values behind Fig. 6 as a table.

Reviewer_K6iVOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SERF, a self-distillation regularizer for continuous diffusion language models. An EMA teacher receives a cleaner version of the student's input along the same interpolation path (same noise, lower noise level) and supervises the student, either by aligning an early student layer with a deeper teacher layer (SERF-L) or by matching the teacher's output token distribution (SERF-P). The regularizer is applied to FLM and ELF on OpenWebText (OWT) and to S-FLM trained on TinyGSM and evaluated on GSM8K. Empirical results show that SERF-L reaches the MAUVE of FLM trained for 1M steps with far fewer training steps, and improves the GSM8K accuracy of S-FLM from 17.7% to 22.3%. Additionally, ablation studies are conducted on the noise level strategy, the aligned layers and the regularization weight.

Critical strengths
  1. The method is simple, adds no inference cost and requires no external pretrained model.
  2. The gains over the baselines are consistent across the settings tested, and the training overhead (about 30% per step) is reported and accounted for in the wall-clock comparison.
Critical weaknesses
  1. The novelty is limited. SERF-L follows SRA (Jiang et al., 2025), which already aligns an earlier layer at higher noise with a later layer at lower noise, and the layer ablation (2→102\rightarrow10 works best) repeats its finding. SERF-P matches the student's prediction to an EMA teacher evaluated on the same noise at a cleaner level, which is a construction similar to the target in consistency training (Song et al., 2023), but the paper only calls it orthogonal to flow map distillation and does not discuss this connection. The paper is mainly a transfer of image-domain self-distillation to continuous language models.

  2. SERF-P is not well justified as a training target. Since the student and teacher inputs share the same noise, the teacher's input carries information about the clean tokens beyond what the student's input has, so the teacher's prediction is not an unbiased target for the student's posterior. In a simple binary example with Gaussian noise (s=0.2,t=0.5s = 0.2, t = 0.5, student input 0.3), the expected teacher prediction given the student input is 0.64, while the true posterior is 0.55. The regularizer therefore pushes the denoiser away from the posterior that the flow relies on, which the paper partly acknowledges when it notes that a larger λ shifts the effective target away from the ground truth (MAUVE drops from 0.320 to 0.259 when λ goes from 0.1 to 0.5).

  3. The evaluation does not fully support the main claims:

    i. The sampling acceleration claims compare against the baseline's 1024-step result, which is not its best. FLM at 1M steps has higher MAUVE at 128 steps (0.329) and 256 steps (0.349) than at 1024 steps (0.313) (Table 1), and on GSM8K S-FLM is already close to its 1024-step accuracy at 64 steps (Figure 4, right), so the "64× fewer sampling steps" is mostly a result of the baseline saturating early. The gap between SERF-L at 16 steps (18.7%) and S-FLM at 1024 steps (17.7%) is about one point.

    ii. The training acceleration relies on a single run per model, and the reported ±\pm is the std over MAUVE seeds only, not over training runs. The FLM MAUVE curve in Figure 4 (left) peaks at 40K steps, where it equals SERF-L (about 0.26), and is lower from 60K to 100K while its Gen. PPL keeps improving, which makes the 100K-step comparison look favorable to SERF.

    iii. The absolute performance remains low. MAUVE stays below 0.44 on OWT, and on GSM8K the best result (22.3%) is far below the autoregressive baselines in the paper's own Table 2 (53.9% with sampling, 63.3% greedy) and below MDLM and Duo at low temperature (33–36%, as the paper acknowledges). The gain on ELF is small, and each of the three paradigms is tested in only one setting. The Gen. PPL gain is also small (72.9 vs 74.9), so the quality gain rests on MAUVE, and the claim that both variants match or exceed FLM at 1M steps does not hold for SERF-P at 256 and 512 steps (Table 1).

    iv. No control separates the effect of the cleaner teacher from plain self-distillation, such as a teacher at the same noise level (t = s). On GSM8K, the 1.4× training speedup is measured against a single released S-FLM checkpoint rather than a training curve.

Questions for authors
  1. Can the authors show that SERF-P preserves the posterior that the flow relies on, or quantify how far the learned denoiser deviates from it?
  2. Which FLM and S-FLM checkpoints are used in Table 1 and Figure 4, released or retrained by the authors? Appendix A.4 states that the baseline results are taken from the respective papers, but the FLM paper does not report MAUVE, and Table 2 marks S-FLM as the authors' own evaluation (17.7% vs 18.0% reported).
  3. S-FLM sampling truncates the distribution to the top-1 token at every step (Appendix A.3). What does τ = 1 mean in this case, and how does SERF compare with MDLM and Duo at τ = 0.1?
  4. What is the variance over training runs for the main OWT and GSM8K comparisons?
  5. (Suggestion) How does SERF-P compare with simply lowering the sampling temperature of the baseline, given that the SERF-P target may sharpen the predicted distribution?
7 comments

We thank the reviewer for the detailed and constructive review, and for recognizing that SERF is simple, adds no inference cost, requires no external pretrained model, and yields consistent gains across settings with its training overhead fully accounted for. Following the reviewer's suggestions, we report new experiments:

  • a same-noise-level teacher control, t=st = s (W3 (iv)),
  • comparisons with uniform label smoothing and an independent-noise teacher (W2),
  • additional training seeds on OWT and GSM8K (W3 (ii)),
  • and a retrained S-FLM training curve (W3 (iv)).

We also provide an analysis that bounds the deviation of the SERF-P target from the posterior (W2, Q1). We address each point below, and would kindly ask the reviewer to reconsider their assessment in light of these results.

W1 Positioning/novelty

(This reply is identical to our response to Reviewer j2UR.)

We thank the reviewer for highlighting these connections. The original submission explicitly builds on image-domain self-distillation, citing SRA and Self-Flow in Secs. 1, 2, and 4.4. We have now expanded this discussion with a point-by-point comparison with SRA, Self-Flow, LayerSync, and consistency training in Appendix A.4, and revised the remark in Sec. 3.2.2.

SERF-L and transfer to language. SERF-L follows SRA and Self-Flow’s principle of aligning an early student layer with a deeper EMA-teacher layer evaluated at lower noise. Our random noise-level pairing likewise follows the cleaner-teacher principle of Self-Flow. We do not claim these principles as new.

Our contribution is to establish their effectiveness for continuous diffusion language models, which is not guaranteed by their success in images. For example, a motivation for image representation alignment is to improve semantic features learned through denoising pixels or VAE latents, whereas token prediction already directly supervises discrete linguistic units. Whether intermediate alignment still provides useful additional supervision, therefore, merits investigation. We demonstrate benefits across three substantially different representations—one-hot vectors, jointly learned hyperspherical embeddings, and frozen T5 latents—and two tasks, unconditional generation and mathematical reasoning.

SERF-P and consistency training. We agree that SERF-P shares the noisy-student/cleaner-teacher construction of consistency training (CT). The quantities being learned, however, have different mathematical roles. Consistency training learns an endpoint mapping that is invariant along a probability-flow ODE trajectory, enabling direct few-step generation. SERF-P instead distills categorical token distributions to regularize the standard denoiser, retaining clean-token cross-entropy and the original sampler.

AspectSERF-PStandard CT
Matched outputToken probability distributionPredicted clean sample
ObjectiveClean-token CE + λ × teacher CEConsistency loss, with an endpoint boundary condition
Student–teacher gapRandom, potentially largeAdjacent noise levels, with a progressively refined grid
SamplingOriginal ODE samplerDirect one- or few-step generation
PurposeImprove denoiser trainingLearn a few-step generative mapping

Our focus is on the effect of this loss as an auxiliary regularizer on training efficiency and generation quality, while retaining the original denoising objective and sampler. We do not claim that SERF-P learns a consistency function. We address the bias introduced by its teacher targets separately in our response on the training target.

What is new. Our contribution is the systematic investigation of noisy-to-clean self-distillation for continuous diffusion language models, establishing its benefits for training efficiency and generation quality across three representations and two tasks. SERF-P additionally studies distillation of categorical token distributions, requiring neither a projection head nor the selection of student and teacher layers. The revised discussion distinguishes these contributions from the self-distillation principles inherited from prior work.

W2. SERF-P as a training target

We thank the reviewer for this careful analysis. We agree that, with shared noise, the teacher's prediction is not an unbiased estimate of the student's posterior. Below we show that SERF-P is a form of label smoothing whose bias is small and controlled by λ, and that the shared-noise teacher nonetheless provides a useful training signal.

Label-smoothing view. Up to the constant factor (1+λ)(1+\lambda), LFLM+λLPL_\text{FLM} + \lambda L_P is cross-entropy against the target (1−α) x1+α q(1-\alpha)\,x_1 + \alpha\,q, with α=λ/(1+λ)\alpha = \lambda/(1+\lambda), i.e., label smoothing with the teacher's prediction qq in place of the uniform distribution. Each per-sample target is softer than the one-hot label and assigns probability to alternative tokens that the teacher considers plausible given a less corrupted context. Unlike the uniform distribution, however, qq is correlated with x1x_1 through the shared noise, which is the source of the bias.

The bias is bounded. At a fixed xsx_s, the minimizer is p⋆=(1−α) E[x1∣xs]+α E[q∣xs]p^\star = (1-\alpha)\,\mathbb{E}[x_1|x_s] + \alpha\,\mathbb{E}[q|x_s], so its deviation from the posterior is at most α≈0.09\alpha \approx 0.09 at λ=0.1\lambda = 0.1.

The teacher's target is what helps. Two ablations show that the gains come from the shared-noise teacher rather than from smoothing alone. First, replacing qq with the uniform distribution, i.e., standard label smoothing with the same α\alpha, performs worse than SERF-P, and even below FLM. Second, building the teacher's input from independent noise, xt=t x1+(1−t) e′x_t = t\,x_1 + (1-t)\,e', under which the pair (xs,xt)(x_s, x_t) no longer determines x1x_1, also performs worse. This supports our choice of a teacher on the same interpolation path.

MethodSmoothing distributionMAUVE (↑)
FLM–0.237 ± 0.009
+ Label smoothingUniform0.212 ± 0.014
+ SERF-P, independent noiseTeacher qq, independent noise0.291 ± 0.017
+ SERF-PTeacher qq, shared noise0.325 ± 0.016

OWT, 50K training steps, 1024 sampling steps, same α\alpha for all variants. Best MAUVE over τ∈{0.95,1.0,1.05}\tau \in \{0.95, 1.0, 1.05\}, ± std over 5 MAUVE seeds.

We have added these controls to Appendix B.

W3 (i). Sampling acceleration

OWT: 2× sampling acceleration on MAUVE. We kindly point out that we have already taken this into account for MAUVE, which is why we claim a 2× acceleration. FLM-1M reaches its best MAUVE at 256 steps (0.349), and SERF-L-300K exceeds this with 128 steps (0.435), i.e., with half the sampling steps.

GSM8K: 64× sampling acceleration. Both SERF and S-FLM show a similar relationship between sampling steps and accuracy: accuracy rises quickly and then saturates (Fig. 4, right). Nevertheless, it is a factual result that SERF-L recovers S-FLM's 1024-step accuracy, which is also its best across all budgets, with 64× fewer sampling steps.

W3 (ii). Training runs

Training-seed variance. We thank the reviewer for this comment. Training each model with multiple seeds is computationally expensive and is not common practice in the diffusion language model literature, where results are typically reported from a single training run. To address the reviewer's concern, we trained two additional seeds (seeds 2 and 3) of SERF-L on both FLM (OWT) and S-FLM (GSM8K):

OWT, MAUVE (↑)Seed 1Seed 2Seed 3
FLM0.237 ± 0.009––
+ SERF-L0.270 ± 0.0170.265 ± 0.0140.274 ± 0.015

50K training steps, 1024 sampling steps, best MAUVE over τ∈{0.95,1.0,1.05}\tau \in \{0.95, 1.0, 1.05\}, ± std over 5 MAUVE seeds. Seed 1 is the run reported in the time-reweighting ablation (Appendix B).

GSM8K, accuracy % (↑)Seed 1Seed 2Seed 3
S-FLM5.78 ± 0.24––
+ SERF-L7.69 ± 0.388.72 ± 0.448.45 ± 0.83

50K training steps, 32 sampling steps, ± std over 4 evaluation seeds.

All three seeds of SERF-L outperform the baseline by a clear margin, confirming that the gains are not due to a favorable seed. Given the limited compute and time available during the rebuttal, we prioritized additional seeds for SERF-L, since the question is whether its improvement is robust, and report the baseline from a single training run. The spread across SERF-L seeds (0.009 MAUVE on OWT, 1.0 point on GSM8K) is well below the smallest gap to the baseline (0.028 MAUVE, 1.9 points), so the improvement is unlikely to be explained by seed variation alone.

Training-acceleration comparison. We respectfully note that our training-acceleration claim does not rely on the comparison with FLM at 100K steps. The reference is FLM trained for 1M steps, and SERF-L at 100K steps already exceeds its MAUVE: 0.336 vs. 0.313 at 1024 steps, and 0.369 vs. 0.349 when each model is compared at its best sampling budget (Table 1). FLM's peak at 40K steps (≈0.26) lies well below both values, so using FLM's best intermediate checkpoint instead of its 100K checkpoint would not change this conclusion. The fluctuation of FLM's MAUVE between 40K and 100K therefore does not affect the comparison. Over the same range, SERF-L's MAUVE increases steadily and continues to improve up to 300K steps (0.389).

W3 (iii). Absolute performance and breadth

Absolute performance. SERF is a training regularizer: its goal is to improve a given continuous diffusion language model under an identical architecture, dataset and training setup. The absolute numbers are therefore largely determined by the base models, which we train with the same architectures and hyperparameters as their original papers. Continuous diffusion language models are currently behind autoregressive and discrete diffusion models, and this gap is precisely what motivates methods that improve their training. Within this setting the gains are substantial. SERF-L-300K reaches a MAUVE of 0.435, compared with 0.349 for FLM-1M at its best budget. On GSM8K, SERF-L raises S-FLM's accuracy from 17.7% to 22.3%, the strongest continuous diffusion result in Table 2. As we already acknowledge in the paper, a gap remains to discrete diffusion models at low temperature.

ELF and breadth of settings. The ELF experiment tests whether SERF transfers to a different paradigm: latent flows over frozen T5 embeddings, where SERF-P is not applicable. Under ELF's own training and sampling protocol, SERF-L improves ELF-B at every guidance scale, shifting the entire quality–diversity front upward. Across the paper, SERF is evaluated on three representations (one-hot, hyperspherical and latent) and two tasks (unconditional generation and math reasoning), using each baseline in its primary setting.

Gen. PPL. The comparison of 72.9 vs. 74.9 is between SERF-L at 300K steps and FLM at 1M steps. More relevant, SERF-L already matches FLM-1M's Gen. PPL at 100K steps (74.7 vs. 74.9), with a tenth of the training steps, and improves on it at 300K steps. At an equal training budget, both SERF variants have lower entropy-matched Gen. PPL than FLM at every training step (Fig. 4, left).

SERF-P vs. FLM-1M at 256 and 512 steps. SERF-P is trained for 100K steps, ten times fewer than FLM's 1M, and performs on par with FLM-1M at every sampling budget. At 256 and 512 steps, its MAUVE is comparable (0.334 vs. 0.349 and 0.340 vs. 0.345). At 64, 128, and 1024 steps, and at its best budget (0.361 vs. 0.349), it exceeds FLM-1M.

Reviewer_8GbEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an additional loss to the non-discrete flow matching models for text, based on similar ideas in imaging (distillation loss). It is a direct generalization of such approach. Comparison on OWT data and GMS8k is shown.

Critical strengths

There is a strong need for better flow matching models for text (i..e, it is known that they require much more tokens on pretraining, so any modification can be useful). I think, the proposed loss is definitely useful

Critical weaknesses
  1. Not all references on continious flow diffusion models for texts are given, i.e. Serviugov, Razzhigaev, Dragunov, Kuznetsov, Oseledets (https://proceedings.iclr.cc/paper_files/paper/2026/file/dfc24bd3ec5d74960e104268bbb52849-Paper-Conference.pdf)
  2. The specific choice of interpolation (2) is quite strange, since they come from different geometries.
  3. The idea itself is a direct generalization of the approaches in other domain. FM in text have rather clear theory, and just adding an additional term without validation is not enough.
  4. The experiments are small, and in terms of quality not SOTA at all.
Questions for authors
  1. Can you provide theoretical justification for the proposed additional loss?
3 comments

We thank the reviewer for the feedback, and for recognizing both the need for better flow matching models for text and that "the proposed loss is definitely useful". We respectfully note that the concerns raised relate to related work, the interpolation inherited from the baseline, novelty and scale, rather than to errors in our method or experimental evaluation. We would therefore kindly ask the reviewer to reconsider the soundness score in light of the responses below.

W1. Missing reference

We thank the reviewer for pointing out Logit-KL Flow Matching (Sevriugov et al., ICLR 2026), which defines flows along KL-geodesic (logit-linear) paths and trains with a likelihood objective. We have added it to the related work (Sec. 2). SERF is complementary to this line of work: it only requires a noise-to-data path and a denoiser, and could in principle be applied on top of such models.

W2. Choice of interpolation (Eq. 2)

Eq. (2) is not a design choice of SERF. It is the interpolation of FLM (Lee et al., 2026), the baseline we build on: the standard linear flow matching path from Gaussian noise to one-hot data in Euclidean space. SERF is a training regularizer that is independent of the choice of path. To demonstrate this, we apply it to three different geometries: Euclidean one-hot vectors (FLM), the hypersphere with geodesic (SLERP) interpolation (S-FLM), and a latent T5 embedding space (ELF). SERF improves all three. The same holds for other paths, such as the KL-geodesic interpolation of Logit-KL Flow Matching, since SERF only requires a noise-to-data path and a denoiser.

W3. Generalization from other domains and validation

Our paper explicitly builds on self-distillation methods from image diffusion (Sec. 1, 2 and 4.4), and Appendix A.4 now compares SERF with these methods point by point. Its contribution is to show that this approach transfers to continuous diffusion language models, across three representations and two tasks. It also introduces SERF-P, which distills the categorical output distribution specific to language denoisers. Given the community's growing focus on diffusion language models, we believe these findings are valuable. SERF does not alter the flow matching framework: the sampler and the velocity parameterization are unchanged, and the FLM cross-entropy remains the primary objective. The added term is validated by ablations on the noise-level strategy, the aligned layers and the weight λ (Sec. 4.4). The revised appendix adds controls that isolate its contribution: replacing the teacher's prediction with uniform label smoothing or computing it from independent noise, separating the distillation term from the change in time distribution it induces, and removing the noise gap between student and teacher (Appendix B). In all cases, the gains are attributable to distillation from a cleaner teacher.

W4. Scale and SOTA

SERF is a training regularizer: its goal is to improve a given continuous diffusion language model under an identical architecture, dataset, and training setup, not to set a new state of the art. We therefore train at the same scale and with the same hyperparameters as the original baselines. Within this setting, the gains are substantial. SERF achieves up to 7.7× wall-clock training acceleration on OWT, and SERF-L-300K reaches a MAUVE of 0.435, compared with 0.349 for FLM-1M at its best budget. On GSM8K, SERF-L raises S-FLM's accuracy from 17.7% to 22.3%, the strongest continuous diffusion result in Table 2.

Q1. Theoretical justification

For SERF-P, up to the constant factor (1+λ)(1+\lambda), LFLM+λLPL_\text{FLM} + \lambda L_P is cross-entropy against the target (1−α) x1+α q(1-\alpha)\,x_1 + \alpha\,q, with α=λ/(1+λ)\alpha = \lambda/(1+\lambda). SERF-P is therefore a form of label smoothing in which the uniform distribution is replaced by the teacher's prediction qq (see our added analysis in Appendix A.7). At a fixed noisy input xsx_s, the minimizer is p⋆=(1−α) E[x1∣xs]+α E[q∣xs]p^\star = (1-\alpha)\,\mathbb{E}[x_1|x_s] + \alpha\,\mathbb{E}[q|x_s]. Its deviation from the posterior used by the sampler is therefore at most α≈0.09\alpha \approx 0.09 at λ=0.1\lambda = 0.1, and it vanishes at both extremes of the student–teacher gap (t=st = s and t=1t = 1). The auxiliary term is thus a controlled perturbation of the FLM objective, whose size is set by λ, consistent with our λ ablation. For SERF-L, the loss places no target on the output distribution. It acts on intermediate representations through a projection head, while the output is still trained only by the FLM cross-entropy. It encourages early layers to form the representations that a deeper teacher computes from a cleaner input, the mechanism shown to accelerate training in image diffusion (Jiang et al., 2025; Chefer et al., 2025).

Dear Reviewer 8GbE,

We would greatly appreciate knowing whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining questions before the end of the discussion period.

Thank you again for your time and thoughtful review, which has substantially improved our paper.

The Authors

Dear Reviewer 8GbE,

We submitted our responses to your review five days ago, and as the discussion period ends on November 18, we would greatly appreciate knowing whether they have addressed your concerns. For your convenience, we summarize the key points below, together with where to find them:

  • Missing reference (W1): Logit-KL Flow Matching (Sevriugov et al., 2026) is now cited and discussed in the related work (Sec. 2).
  • Choice of interpolation (W2): Eq. (2) is inherited from the FLM baseline. SERF is independent of the path and improves models with three different geometries: one-hot (FLM), hyperspherical (S-FLM) and latent (ELF).
  • Validation of the added term (W3): new controls isolate its contribution (Appendix B). Uniform label smoothing, an independent-noise teacher, and time reweighting alone all fail to match SERF. A point-by-point comparison with related methods is in Appendix A.4.
  • Scale (W4): SERF uses the same scale and hyperparameters as each baseline, with up to 7.7x faster training on OWT and a GSM8K gain from 17.7% to 22.3%.
  • Theoretical justification (Q1): SERF-P is equivalent to label smoothing with the teacher's prediction (Appendix A.7).

We would be happy to discuss any remaining questions. Thank you again for taking the time to review our paper.

The Authors

Comments on the paper

1 comment

We thank all reviewers for their careful and constructive feedback. We have uploaded a revised manuscript in which all changes are highlighted in blue. In summary:

New experiments

  • Smoothing distribution (Appendix B): SERF-P outperforms both uniform label smoothing and a teacher computed from independent noise, showing that the gains come from the shared-noise teacher rather than from smoothing alone.
  • Time reweighting vs. distillation (Appendix B): reweighting the time distribution alone does not improve over FLM, while SERF-L with a uniform student time retains nearly its full gain.
  • Same-noise-level teacher, t=st = s (response to Reviewer K6iV, W3 (iv)): removing the noise gap loses most of the gain of SERF-L and makes SERF-P fall below the baseline, confirming that the cleaner teacher is key.
  • Additional training seeds (response to Reviewer K6iV, W3 (ii)): two additional seeds of SERF-L on OWT and on GSM8K, all clearly above the baseline.
  • Retrained S-FLM training curve (response to Reviewer K6iV, W3 (iv)): SERF-L outperforms a retrained S-FLM baseline at every training step up to 100K.
  • Extended temperature sweep (response to Reviewer j2UR, W4): the entropy-matched Gen. PPL is now interpolated at every sampling budget, and the results are robust to the fitting function.

New analyses and appendix sections

  • Label-smoothing view of SERF-P (Appendix A.7): the SERF-P objective is equivalent to label smoothing with the teacher's prediction.
  • Relation to prior work (Appendix A.4): a point-by-point comparison with SRA, Self-Flow, LayerSync and consistency training.
  • Implementation details (Appendices A.2 and A.3): the construction of shared-noise teacher–student pairs for ELF and S-FLM.

Clarifications: baseline checkpoints and evaluation (response to Reviewer K6iV, Q2), the sampling temperature under top-1 decoding (response to Reviewer K6iV, Q3), and EMA sampling (response to Reviewer j2UR, Q3).

Taken together, these controls consistently indicate that the gains of SERF come from distillation from a cleaner teacher, rather than from smoothing, time reweighting or plain self-distillation. We address each reviewer's points in detail in the individual responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.