acceptodds
Withdrawn from ICLR 2027

Reflection Pretraining Enables Token-Level Self-Correction in De Novo Peptide Sequencing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

De novo peptide sequencing infers a peptide's amino-acid sequence directly from its tandem mass spectrum (MS/MS). State-of-the-art models decode the peptide autoregressively, one residue at a time, and cannot revise a residue once it has been emitted: a single misread residue, caused for instance by a missing fragment ion, corrupts the whole peptide. We propose reflection pretraining, which augments the amino-acid vocabulary with a reflection token <reflect> and trains the decoder to revise its own residues during generation. De novo sequencing is unusually well suited to this: every training spectrum is paired with a gold-label peptide, so the position and identity of an injected error are known exactly, and the spectrum itself (the fragment-ion ladder and the precursor mass) contains the evidence needed to verify each residue. During pretraining, we inject synthetic residue errors into target peptides and require the model to emit <reflect> followed by the correct residue. Errors are drawn either uniformly or from later residues of the same peptide, which mimics the residue-order confusions typical of real spectra. Gradients are blocked at error positions, so the injected error serves as context rather than as a learning target. On the 9-species benchmark, reflection-pretrained Transformers detect and correct their own mistakes at inference time and improve amino-acid and peptide-level precision over a standard Transformer and, on average, over domain-specific de novo methods, without any task-specific architectural modules. The same mechanism regularizes training and lets experts insert <reflect> at suspect residues. Code, data and pre-trained model weights can be found in the Anonymous GitHub.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Biology, de novo peptide sequencing, ai for biology, sequence modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SH3zOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets de nove peptide sequencing with transformer models. A key gap the paper observe is that current model generates the sequence in one step, and lack mechanism mimicing the "thinking" behavior of modern autoregressive model like LLM. This is partly due to the nature of peptide sequence, where thinking cannot be explicitly represented. Hence, the authors propose a way to artificially construct the "thinking" sequence, or as they termed <reflect> tokens to assemble a thinking process for these transformers. Specifically, they construct miss-by sequence and insert a <reflect> tokens on where the miss happens, urging the model to think twice on the last token. The empirical shows that after such training, the model is able to obtain much better results.

Critical strengths

The empirical evidence is strong. Practicaly, this method is easy to implement, and no RL needed for training like existing thinking model for other autoregressive model (although RL seems highly reasonable, where one first warm up with this <reflect> format, and use RL to keep improving).

Critical weaknesses
  • The concept is not new. This reminds me of [1], where mistakes are injected during pretraining and see if the model can realize that by itself. Moreover, compare to existing approaches, this one seems particularly limited, having <reflect> right after the potential mistakes means that the model can only realize mistakes happend in the last token. The model might be capable of detecting mistakes few steps earlier, but his formulation blocked that. Since you need to do data construction anyway, more general method like injecting multiple <reflect> to represent longer reflect process can all be considered. I think the current idea is nice, but just too many loose end on why this is "the" approach.

  • From the paper, the self-correction does not fire very much, this fundamentally discredit the claim. Also even the among the fired ones, only few actually resolve the mistake.

[1] Ye, Tian, et al. "Physics of language models: Part 2.2, how to learn from mistakes on grade-school math problems." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors
  • Do you have a more finegrained statistics on mistakes and <reflect> fire rate?

  • Can you try just train something like multiple <reflect> tokens?

  • Is RL applicable? I feel the current stage is like the warm up stage, and an RL might bring up the success.

Reviewer_m5T4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Autoregressive de novo peptide decoders cannot revise an emitted residue, so one early error ruins the peptide. The paper adds a ⟨reflect⟩ token and pretrains on gold peptides with one injected error, followed by ⟨reflect⟩ and the correct residue. Errors are either uniform (RPRE) or order swaps drawn from later residues (RPLE), and the loss on the error token is masked (Sec. 3.2). On the 9-species benchmark, this raises average peptide recall over Casanovo from 0.521 to 0.600 (Table 2).

Critical strengths
  • Simple drop-in change with consistent gains. Peptide precision improves on all nine species with no architectural change (Table 1). The placeholder-token control rules out sequence length as the cause.
  • Error realism matters. RPLE beats RPRE at the same injection ratio (e.g., Human 0.494 → 0.535; Table 1), so the content of the injected errors is not incidental.
Critical weaknesses
  1. The reported gains are not attributable to the proposed self-correction mechanism, and the control needed to separate the two is missing.

    • In Table 3, only 4.5% of sequences emit ⟨reflect⟩ in the 90%-error model.
    • Only 12.7% of reflections correct an error, while 19.5% replace a residue with another error or turn a correct residue into an error.
    • Yet peptide precision improves by 9.0 points on Mouse (0.443 → 0.533) and 13.0 points on Human (0.433 → 0.563).

    Even if every reflecting sequence had been flipped from wrong to right, reflection could account for at most about 4.5 points. Most of the gain must therefore come from sequences that never emit ⟨reflect⟩. A more plausible explanation is that error-injected prefixes act as a regularizer or exposure-bias correction. The paper mentions a “counter-overfitting” effect but never isolates it. The placeholder-token ablation only rules out a sequence-length effect. The necessary control is error injection with identical gradient blocking but without ⟨reflect⟩, i.e., the correct residue follows the error directly; scheduled sampling or standard regularizers would also be informative. This is consequential because the title, abstract, and theoretical framing all attribute the improvement to token-level self-correction.

  2. The reported comparisons are not reliable enough to support the headline numbers.

    • The “Standard Pre-train Transformer” baseline in Table 1 is identical to the Casanovo (beam = 1) entries in Table 2 to three decimals, across all nine species, for both metrics. It is unclear whether this baseline was retrained under the authors’ configuration.
    • The configuration itself is inconsistent. Section 4.1 states 512-dimensional embeddings, a learning rate of 5×10⁻⁴, and a 10k-step warmup. Appendix F states 400 dimensions, 4×10⁻⁴, batch size 1600, a one-epoch warmup, and 30 epochs. The same values are labeled peptide precision in Table 1 and peptide recall in Table 2.
    • InstaNovo’s Tomato peptide recall (0.732) equals its Tomato AA precision, which looks like a transcription error. If the value is correct, InstaNovo outperforms the proposed method on Tomato (0.668).
    • Table 2 mixes baselines from different training protocols (species-wise cross-validation vs. MassIVE-KB training). It also compares beam sizes 3 and 5 against beam-1 baselines.
    • Figure 5 reports overfitting from about 100 steps. At batch size 1600 that is roughly 160k spectra, far less than one epoch over 30M PSMs. The no-injection curve also sits at a different level from step 0, which suggests the losses are computed over different targets and are not comparable.

    Because the paper’s case rests almost entirely on these empirical comparisons, I cannot accept the reported gains as like-for-like improvements.

Questions for authors

Clarifications

  1. Please decompose the peptide-precision gain into sequences with and without ⟨reflect⟩. In addition, what precision does the reflection-pretrained model achieve if ⟨reflect⟩ is suppressed at inference (logit set to −∞)? If most of the gain survives suppression, the improvement comes from training-time regularization rather than self-correction.
  2. What are the results of the reflect-free control described above: error injection with identical gradient blocking, followed directly by the correct residue?
  3. Was the Table 1 baseline retrained by the authors under the same configuration as the reflection models? Which configuration (Section 4.1 or Appendix F) produced each row of Tables 1 and 2?
  4. Please clarify:
    • whether the peptide metric is precision or recall, and at what coverage;
    • whether the InstaNovo Tomato value is correct;
    • which 9-species version (v1 or v2) is reported;
    • the training protocol of each baseline in Table 2.
  5. In Figure 5, over which targets is validation loss computed for each model? Are ⟨reflect⟩ tokens and injected positions included, and are all models evaluated on identical validation targets?
  6. Ye et al. (ICLR 2025, Physics of Language Models Part 2.2) studied pretraining on erroneous solution steps immediately followed by their corrections, including whether masking is needed on the erroneous tokens and whether such data can be deferred to the fine-tuning stage. Backspace actions were also proposed in SequenceMatch (Cundy & Ermon, 2024). How does the proposed method differ beyond the application domain? The claim to be the first to incorporate reflection into pretraining should be revised accordingly.
  7. Please report variance across random seeds. Several per-species differences are small; for example, R.bean peptide precision improves by only 0.8%.

Suggestions for future work

  1. Analyze where ⟨reflect⟩ fires (e.g., at positions with missing fragment ions), and whether RPLE specifically reduces order-swap errors.
  2. Evaluate human-in-the-loop correction on a larger set of residues flagged without access to gold labels.
  3. Substantially shorten or formalize the expressiveness argument in Section 2 and Appendices B–E. A single token that revokes the preceding residue does not provide the unbounded intermediate computation that the cited CoT results rely on.
Reviewer_noKDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a token-level self-correction method for de novo peptide sequencing, called "Reflection Pretraining". Specifically, this method introduces a special <REFLECT> token and employs a dynamic error injection mechanism - including random amino acid substitution and lookahead-based perturbations - while masking the training loss at the damaged positions. Benchmark tests on nine species demonstrate that this method performs well in terms of accuracy at the amino acid level and peptide segment level, and is highly competitive.

Critical strengths
  1. This method introduces an explicit error correction mechanism without any modification to the model architecture. By leveraging peptide sequence supervisory information and mass spectrometry data, it alleviates the problem of cumulative error propagation that occurs during the self-recursive sequence generation process.
  2. The proposed method improves the prediction accuracy at the amino acid and peptide segment levels. By comparing it with the existing cutting-edge methods in this field and conducting relevant ablation experiments, the effectiveness of this method has been verified.
Critical weaknesses
  1. In the paper, only 4.5% of the sequences triggered the "reflection" mechanism during the reasoning process, and among these sequences, only 12.7% successfully corrected the erroneous residues. Was the performance improvement more due to the robustness brought about by introducing errors during the training process rather than the self-correction in the reasoning stage? Moreover, the ablation experiments in the paper did not distinguish between these two factors; it is recommended to compare the performance when the "reflection" mechanism is enabled and disabled while keeping the training scheme unchanged.
  2. If considering any finite-length sequence, the original sequence set itself is already a countably infinite set. Introducing special tokens neither changes the cardinality of this set nor proves that the model has stronger computational capabilities. Therefore, the authors must clearly specify what operational capabilities these tokens specifically introduce. Additionally, the authors should demonstrate whether the "reflection" mechanism truly meets the applicable conditions of the related "Chain-of-Thought" (CoT) theory, rather than merely claiming its theoretical advantages based on superficial similarity in form.
  3. Section 4.1 of the paper states that the embedding dimension is 512, the learning rate is 5×10−45\times10^{-4}, and the warm-up steps are 10,000; however, the corresponding values listed in Appendix F are 400, 4×10−44\times10^{-4}, and a training epoch. This inconsistency is perplexing - are these different configurations used for different experiments? To ensure the reproducibility of the paper, it is necessary to clearly determine which set of configurations was actually used to obtain the main results.
  4. RPLE selects a residue from the subsequent positions of the sequence to replace the residue at the current position, while keeping the original residues at the subsequent positions unchanged. This is a one-way substitution rather than an exchange between two positions; this means that the total "quality" (or composition) of the sequence may not remain unchanged. This raises a question: How does this mechanism align with the original intention of "maintaining the total quality"? The underlying logic is still unclear.
  5. This paper uses artificial perturbations to simulate prediction errors, but it does not compare these perturbations with the errors generated by the model in practical applications (including aspects such as error type, position, and frequency). If there are significant differences between the two, then training based on these synthetic errors is likely to be difficult to effectively help the model correct various common errors in practical applications. Can further analysis be provided to explain why the selected perturbation method is suitable for simulating the errors that occur in the actual prediction process?
Questions for authors

Refer to the weaknesses

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

Related papers

Discussion (0)

Sign in to comment.