acceptodds
Under review as a conference paper at ICLR 2027

Target-Distribution Matching for Activation Steering: Paired Causal Residual Transport

Abstract

Activation steering changes a frozen language model’s behavior, but behavioral scores and hidden-state distances do not specify which output distribution it should produce. For instruction-defined control, the model supplies this target: its next-token distribution under the target instruction, given the same content and prefix. We introduce Paired Causal Residual Transport (PCRT), which predicts the same-content residual left by a fixed editor and adds it after final normalization, before the vocabulary head. The target instruction supplies training supervision and is absent at deployment. Within the learned residual subspace, the conditional mean minimizes squared output-head error, which bounds probability error before numerical rounding. On Gemma 2 2B, PCRT reduces 16-token sequence Kullback–Leibler (KL) divergence by 35.7% to 54.4% for Activation Transport (AcT) and Linear End-to-end Activation Steering (LinEAS), outperforming matched different-content residuals on all 64 test contents. AcT repair reduces 32-token sequence KL by 69.6% to 74.2% on Qwen 2.5, Qwen 3, and Llama 3.1. On five public instruction-following categories, direct correction lowers macro sequence KL by 22.2% and raises strict compliance by 44.1 percentage points. A fixed editor’s remaining distributional error can therefore provide a learnable correction target, without retraining the editor.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Activation Steering, Large Language Models, Target-Distribution Matching, Residual Correction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DhDUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Paired Causal Residual Transport (PCRT), which learns to correct the remaining activation errors of existing steering methods using paired source-target states. The method uses output-head geometry and conditional residual prediction to better match the distribution induced by a target instruction. Experiments demonstrate substantial reductions in sequence KL divergence across multiple models.

Critical strengths
  1. The paper provides a well-defined distribution-level objective for evaluating and improving activation steering.
  2. The experiments across multiple backbones and steering methods demonstrate consistent reductions in target-distribution errors. The ablation studies and statistical analyses are reasonably thorough.
Critical weaknesses
  1. The main approach relies on established residual learning and linear prediction techniques. Moreover, replacing output-head geometry with Euclidean fitting produces nearly identical results, raising questions about the necessity of the proposed formulation.
  2. Insufficient justification for post-editor correction. Tables 19–20 show that direct Source + PCRT achieves lower sequence KL than AcT + PCRT across all three external backbones. The paper does not adequately establish why correcting an existing editor is preferable to learning a direct intervention.
Questions for authors

1.Under what conditions does post-editor PCRT offer advantages over direct source correction? 2.What are the practical benefits of the proposed output-head geometry and conditional predictor over simpler residual correction methods?

4 comments

Thank you for the positive assessment and for two precise questions. Changed text in the revised PDF is blue, and L refers to its line numbers. Tables 19–20 of the submission are now Tables 21–22.

Q1. Under what conditions does post-editor PCRT offer advantages over direct source correction?

Your reading of Tables 19–20 is correct, and the revision makes direct correction the recommended default (L95–102, L463–470). We can now state a condition and the evidence for it.

In our data the editor has to start closer to the target than the unedited source. Under our settings AcT starts farther on all 64 test contents of every additional backbone, at 2.0 to 6.8 times the source KL. PCRT removes a similar or larger share of its error (69.6% to 74.2%, against 45.8% to 71.8% for the source), yet the post-editor path ends above the direct path on 64, 62 and 64 of 64 contents (new Table 40). The same holds at all four training budgets (Table 28) and in the five IFEval categories. In none of our comparisons does a candidate that starts farther end closer.

Starting closer is not enough, however. A response-aligned transport editor starts closer than the source in four question answering and sentiment cohorts. After correction it matches direct correction or exceeds it by one item in joint success with the KL-fitted correction, and the result is mixed with the closed-form correction. KL favors each path in four of eight fits (Appendix S.3, Tables 42–43). On IFEval the post-editor path gains 9 items and loses 4 on one model, and gains 13 and loses 4 on the other, against direct correction (p = 0.27 and 0.049, unadjusted).

We therefore do not claim a general post-editor advantage (L518–525). Its use is repair. When an editor is already fixed, PCRT lowers its sequence KL by 35.7% to 54.4% on Gemma 2 with two editors and by 69.6% to 74.2% on three backbones without retraining it, and it raises joint success for an editor with usable output in all four cohorts (Table 4).

Q2. What are the practical benefits of the output-head geometry and the conditional predictor over simpler residual correction?

The geometry has a definitional role. G=W⊤WG=W^\top W identifies which state error the vocabulary head can see, gives the probability bound, and separates span coverage from prediction error. It does not improve accuracy in our experiments. Euclidean fitting changes mean head error by at most 0.55% and both intervals cross zero, which the main text now states (L412–415). Appendix Q.3 explains why this is expected. For fixed features, unrestricted multi-output least squares has the same normal equations under any positive-definite output weighting.

The conditional predictor does help. The revision adds its comparison on the additional backbones with a paired residual mean, which is the simpler correction with the same supervision (Table 39, L407–410).

BackboneAcT KLPaired meanConditionalConditional better
Qwen 2.5 7B127.9114.138.964/64
Qwen 3 4B203.7154.453.064/64
Llama 3.1 8B177.3101.845.863/64

All three simultaneous 95% intervals for the difference exclude zero. On the primary Gemma 2 task the gain over the mean is 19.6% for AcT and 14.2% for LinEAS, with intervals [1.289, 1.873] and [0.798, 1.351]. On the second attribute there is no gain, and the paper says so. A different-content residual with the same predictor loses on all 64 contents for both editors. The benefit therefore depends on the task and the model, and it is largest where the editor's error is largest.

The revision also adds behavioral evidence for both paths (L452–479) and a worked example (Table 1). We would be glad to answer further questions.

This follows up on both of your questions with a new run of the estimator exactly as submitted, now with three training replicates (Table 4 and Appendix S.4 of the second revision). The editor is a response-aligned transport editor whose outputs stay fluent, and the tasks are question answering, sentiment classification and the 73 IFEval items.

On the conditional predictor (Q2), the run compares it with the simpler correction on outputs. A paired residual mean fitted on the same pairs changes task success by at most three items, both after the editor and on the unedited source. The conditional predictor raises success in all 18 model, task and replicate cohorts in both settings. After the editor on Qwen 2.5, for example, exact answers in the instructed format go from 0 to between 35 and 42 of 48 questions. Against the paired mean, all 36 comparisons are significant after Holm correction (Table 45). We expect the reason to be that the instruction changes some response positions, such as the opening of the JSON object, much more than others, which a constant shift cannot express. The saved measurements do not resolve KL by position, so we state this as an expectation and not as a result.

On when post-editor correction helps (Q1), this editor starts closer to the target than the unedited source in all 18 cohorts. After correction it ends with lower KL than direct correction in 13 of them and has higher task success in 11, lower in 3 and equal in 4. None of these differences is significant (Tables 45 and 47). This sharpens the condition in our first reply. In our data an editor has to start closer to the target for the post-editor route to end closer, but starting closer does not by itself give a measurable advantage, so we still claim none. What the run does show is that post-editor correction reliably repairs an editor that is already in place, with higher task success than the editor alone in all 18 cohorts and significantly so in 17.

Does this answer both questions? If the revision resolves the two weaknesses in your review, we would be grateful if you could say so in the thread, because the other two reviews raise the same two points.

Thank you for your response. The comparison with the paired residual mean also addresses my concerns about the conditional predictor.

Thank you for returning to the thread and for confirming this. Your two questions led to the comparison in Table 39 and to the three-replicate study in Table 4, and the paper is better for them.

If any concern from your review remains open, please tell us and we will answer it today. If none does, we would be grateful if you would consider whether your rating still reflects your assessment of the revised paper.

Reviewer_i1yeOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Paired Causal Residual Transport (PCRT), a method that learns to correct activation steering outputs toward the distribution induced by a target instruction. The method predicts residuals between the outputs of a fixed editor and the target representations using paired trajectories. Experiments show reductions in sequence-level KL divergence across multiple models, along with improvements in instruction-following performance on selected tasks.

Critical strengths
  • The paper introduces a clear distribution-level objective for measuring how closely activation steering reproduces target-instructed behavior.
  • The experiments include useful controls that distinguish the effects of paired residual supervision and candidate-dependent prediction, rather than reporting KL improvements alone.
Critical weaknesses
  • Limited methodological novelty. The proposed method is essentially a learned linear correction of paired hidden-state residuals. Its main theoretical results largely follow from standard properties of regression and probability distributions, making the contribution beyond existing learned steering approaches unclear.
  • Insufficient practical motivation. PCRT learns to reproduce the distribution induced by a target instruction, but it is unclear when this approach is preferable to directly providing the instruction or using existing prompt distillation methods, especially given the need for task-specific paired supervision.
  • Unclear necessity of post-editor correction. The paper also shows that direct source-to-target correction achieves substantial KL reductions. This raises the question of why an existing editor is needed before learning an additional residual correction.
  • Limited behavioral evaluation. Most post-editor results focus on KL reduction, while the IFEval compliance improvements come from direct correction. It remains unclear whether correcting an existing editor leads to meaningful improvements in downstream performance.
Questions for authors
  • Could the authors clarify the methodological differences between PCRT and existing paired-prompt steering methods? What is the main contribution beyond learning a linear correction from paired hidden states?
  • In what scenarios would PCRT be preferable to directly using the target instruction or applying prompt distillation, considering the additional training requirements?
  • Given that direct source-to-target correction is already effective, what practical benefits does the initial activation editor provide?
  • Could the authors provide further analysis of whether KL reductions translate into downstream task performance improvements?
4 comments

Thank you for the careful review and for noting the distribution-level objective and the controls. We answer the four questions in order. Changed text in the revised PDF is blue, and L refers to its line numbers.

Q1. How does PCRT differ from existing paired-prompt steering, and what is the contribution beyond a linear correction?

PSR trains activation-conditioned rank-one interventions at every decoder block by gradient descent and replaces the prompt starting from the source. PCRT is a single affine map at the output interface, fitted in closed form, and its input can be the unedited source or the output of any fixed editor. Context distillation updates the model's weights, while PCRT changes no weight and no internal activation (L125–131, Table 8).

We agree that the estimator is simple and that its identities are standard, and the revision says this plainly (L501–507). The contribution lies elsewhere. First, we define a same-content, same-prefix target and an estimator of sequence KL against it that applies to any editor. Second, we show that what editors leave behind is structured and is not a constant offset. The revision adds a result on this point. On three additional backbones a paired residual mean removes 10.8% to 42.6% of AcT sequence KL, whereas the conditional predictor removes 69.6% to 74.2% and is better on 64, 64 and 63 of 64 contents (Table 39). Correct pairing also beats matched different-content residuals on all 64 contents for both editors. Third, the criterion is diagnostic, as described under Q3.

Q2. When is PCRT preferable to the target instruction or to prompt distillation?

When the instruction can simply be supplied, prompting is the reference itself and we measure no speed gain (1.949 s against 1.944 s), so PCRT has no advantage there. It is also not the most accurate way to remove a prompt. Matched PSR reaches 65.8% compliance against 52.0% for PCRT, and an all-layer parameter student matches or exceeds a terminal correction in all four cohorts we tested (Appendix S.5).

We see three situations where PCRT fits (L508–517). The first is a deployment where the source prompt and the weights must stay fixed and only a controller on activations can be attached. The second is auditing or repairing an existing steering editor against the instruction it is meant to imitate, without retraining it. The third is when supervision and compute are small. Eight paired contents remove 62.8% of the source KL, and the fit is closed form and takes about one second once the states are collected.

Q3. What does the initial editor provide?

Nothing that is required. Direct correction is our recommended default (L463–470). The criterion shows why. AcT is farther from the target than the unedited source on all 64 contents of each additional backbone, and the candidate that starts farther ends farther (Table 40). In our data an editor helps only if it starts closer to the target than the source, and even then it does not always help (L518–525). Post-editor PCRT is meant for an editor that is already in place.

Q4. Do KL reductions translate into downstream performance?

For direct correction they do in the study we have (L452–461, Table 41). Strict compliance rises from 7.9% to 52.0%, which closes 70.8% of the gap to the target prompt. Thirty-one of 73 items go from failing to passing and none from passing to failing. KL and compliance order the four arms in the same way (source, PCRT, PSR, target prompt). Compliance rises in the four categories where KL falls and does not rise in the one where KL rises. Items that start passing have a mean KL reduction of 15.4, against 3.1 for the others (stratified one-sided permutation p = 0.029).

For post-editor correction the answer depends on whether the editor's own output is usable. The AcT configuration in our first extension collapses before any correction is applied (none of its 73 outputs ends naturally), so its higher format score is not answer recovery, and the paper says so. With a response-aligned transport editor that does not collapse, correction improves the task outcome in every cohort (Table 4, Appendix S).

Joint answer-and-format successEditor alone+ closed-form+ KL-fittedTarget prompt
Qwen 2.5, QA0/4814/4837/4844/48
Qwen 2.5, sentiment11/6460/6457/6459/64
Qwen 3, QA0/4831/4844/4843/48
Qwen 3, sentiment53/6461/6461/6461/64

Answer exact match changes by at most three items. On the 73 IFEval items, format compliance rises from 17 to 33 and from 11 to 34 (17 gains against 1 loss, and 24 against 1), and the text is not degenerate (Table 44). The KL-fitted correction is an addition to the method in the revision (L256–259).

This follows up on your four questions with one new experiment in the second revision.

Q4 asked whether correcting an existing editor improves downstream performance. Our first reply relied mainly on a KL-fitted version of the correction with one seed. We have now run the estimator exactly as submitted, with three training replicates (Table 4, Appendix S.4). The editor is a response-aligned transport editor whose own output stays fluent.

Task success, range over three replicatesEditor aloneEditor + PCRTTarget prompt
Qwen 2.5, QA (of 48)035 to 4244
Qwen 2.5, sentiment (of 64)11 to 1259 to 6059
Qwen 2.5, instructions (of 73)17 to 1837 to 3949
Qwen 3, QA (of 48)038 to 4143
Qwen 3, sentiment (of 64)53 to 5761 to 6261
Qwen 3, instructions (of 73)11 to 1335 to 3647

QA and sentiment need the exact answer in the instructed JSON format, and instructions use the official strict format check. Success rises in all 18 cohorts, and 17 of the gains are significant after Holm correction. The exception is one Qwen 3 sentiment replicate in which the editor alone already reaches 57 of 64. Answer exact match changes by at most three items, so the gain is in the instructed format and not at the cost of the answers.

On Q1, a constant correction fitted on the same pairs changes success by at most three items in every cohort. The linear map therefore has to read the state, and an offset does not do the job.

On Q2 and the training requirements you mention, the revision now reports measured cost (Appendix S.5). On one A100, sampling the supervision, collecting the paired states and solving the closed-form fit took 1.7 to 7.9 seconds per task for direct PCRT. Training the all-layer parameter student took 41 to 76 seconds per task on the same type of GPU. The student is more accurate on Qwen 2.5 question answering (44 against 33 to 37 of 48) and within three items on the other three cohorts. The choice is between a fit that takes seconds and changes no weight, and a more accurate one that trains every layer.

On Q3, the post-editor and direct routes do not differ significantly in any of the 18 cohorts, so the editor remains optional.

Your review rates soundness as minor or fixable and asks four questions. Do these answers resolve them? If one does not, please tell us which, and we will address it before the discussion closes on 18 November.

One addition before the discussion closes. Reviewer DhDU has confirmed in this forum that the comparison with the paired residual mean addresses the concern about the conditional predictor. The same comparison is part of our answer to your first question. A constant correction fitted on the same pairs changes task success by at most three items, while the correction that reads the state raises it in all 18 cohorts (Table 4), so the error left to correct is not a fixed offset.

If any of your four questions is still open, one line saying which would let us answer it today.

Thanks for the detailed response and additional experiments. Some of my concerns have been addressed, but my remaining concerns are mainly related to the novelty of the work.

I am still not fully convinced that PCRT offers a sufficiently distinct contribution beyond existing paired-prompt steering and linear correction methods, particularly given that it also requires gradient-based optimization.

Additionally, I find the scenarios suggested by the authors somewhat unrealistic, especially in settings where activation steering requires retaining the source prompts and model weights.

Reviewer_2u8DOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Paired Causal Residual Transport (PCRT), a method to improve existing activation steering methods by matching the output distribution induced by a target instruction.

Critical strengths

The authors evaluate the method across multiple model families and two existing activation steering methods. Although I want to warn that the clarity of the writing prevents me from evaluating the quality of the proposed methodology a lot.

Critical weaknesses
  1. The main paper is very difficult to follow, and the problem setup is not clearly introduced. I find the presentation unnecessarily complicated, especially given that the underlying problem and method seem relatively straightforward. Figure 1 introduces many mathematical symbols without explaining them, and the example illustration does not help clarify the setup. The actual source and target instructions are also not clearly presented in the main text. It took me quite some effort to understand what the input is, what behavior the authors want to achieve, and why an additional repair step is needed. These are the fundamentals of the paper and should be presented in a very straightfoward way.

I think the authors should first explain the problem using a concrete example, clearly distinguish the source, target, candidate, and repaired models, and then introduce the mathematical formulation. These are essential parts of the problem setup and should not require reading the appendix.

  1. I am not fully convinced by the motivation for target-distribution matching. The paper argues that the target instruction defines the desired output distribution, but I am not sure why matching this particular distribution should be the objective of activation steering.

  2. Another question I have is why we need the existing editor in the first place. If we already have paired source and target trajectories for learning the correction, why not directly learn an intervention from the source to the target?

Questions for authors

See weakness above.

2 comments

Thank you for the detailed comments on presentation. They were right, and most of the revision responds to them. Changed text is blue, and L refers to line numbers in the revised PDF.

The setup is hard to follow, and the instructions are not in the main text.

Section 3.1 now explains the task in words before any symbol appears (L138–142). It then prints the complete source prompt and states the one change that turns it into the target prompt (L143–152). A new Table 1 follows a single held-out scene, "A group of people are each holding umbrellas on a rainy day," through the paths with the model's actual outputs.

PathWhat runsOutput begins
Sourceprompt with "Style: literal"(the deployment input)
Targetsame prompt with "Style: vivid fantasy", offline only"with a symphony of raindrops, each one a tiny hammer striking the cobblestones."
Candidatesource prompt + AcT editor"to weave a tapestry of light and shadow, mirroring the celestial dance of the heavens."
Repairedcandidate + PCRT"with a symphony of rain, each droplet a tiny, shimmering, emerald green jewel. The umbrellas, …"

The picture in Figure 1(a) only stands for an example content, and the caption now says so. The worked example is Table 1. Figure 1 also uses plain labels instead of symbols and marks the target branch as used in training only. The introduction now ends with a list of contributions (L95–102).

PCRT takes paired runs of the same content under the source and the target instruction and learns an affine correction to the final hidden state, so that the model without the target instruction produces next-token distributions close to those it produces with it.

Why should matching this particular distribution be the objective?

The task we study is to reproduce the effect of a specific instruction when that instruction is not supplied. For that task the model run with the instruction is the specification, token by token, and forward KL to it is the excess log loss of the imitation (Appendix Q.1). We see two reasons to prefer it to an attribute score.

First, an attribute score can reward the wrong thing. A fantasy lexicon fires on 42.2% of AcT outputs but on only 7.8% of the target prompt's own outputs, because AcT overshoots the style and loses the scene, as Table 1 shows. After repair, source-caption recall rises from 0.266 to 0.411 for AcT and from 0.321 to 0.434 for LinEAS (L153–157, Appendix M).

Second, the distribution tells us something an attribute score cannot. In our settings AcT is farther from the target than the unedited source on all 64 test contents of each of three further backbones (Table 40).

Lower KL also goes with better behavior. With direct correction, strict IFEval compliance rises from 7.9% to 52.0%, 31 of 73 items go from failing to passing, and none goes the other way (L452–461). We do state that the target distribution is a reference to imitate and not a guarantee of answer quality (L150–152).

Why do we need the existing editor in the first place?

We do not, and the revision says so (L95–102, L463–470). Learning the intervention directly from the source is PCRT with the identity as the editor, and it is our recommended default. It reaches lower final KL than the post-editor path on all three additional backbones and at all four training budgets (Tables 40 and 28).

Post-editor PCRT answers a different question. If an editor is already in place, how far is it from the instruction it is meant to imitate, and can the gap be closed without retraining it? There PCRT removes 35.7% to 54.4% of sequence KL for two editors on Gemma 2 and 69.6% to 74.2% for AcT on three backbones. With an editor whose output is usable, it also raises joint answer-and-format success in all four tested cohorts (Table 4).

We hope the revised Section 3.1 makes the method easy to assess, and we are glad to clarify anything that is still unclear.

We are writing once more because the discussion closes on 18 November and we would like to use the remaining time well.

Your third point was that, with paired source and target runs available, one could learn the intervention directly from the source without an editor. That is what the paper now recommends, and the second revision tests it on outputs. With the estimator as submitted, learning directly from the source raises task success in all 18 model, task and replicate cohorts. On question answering, exact answers in the instructed JSON format go from 0 of 48 to between 33 and 42. Strict IFEval passes go from 7 to between 33 and 37 of 73 on Qwen 2.5 and from 4 to between 29 and 35 on Qwen 3 (Table 4, and the table in our note to all reviewers).

Two answers from you would help us. Is the setup now clear from Section 3.1 and Table 1, which give the full prompts and one worked example before any notation? And is there a soundness concern apart from presentation? The review rates soundness 1 but names no error, so at present we have nothing specific to answer. If there is one, please tell us and we will address it before the discussion closes.

Comments on the paper

2 comments

We thank all three reviewers for their time. We have uploaded a revised PDF in which every sentence added or changed since submission is shown in blue. Line numbers below (L) refer to this revised PDF.

The reviews raised three shared points, and the revision answers each of them.

First, the setup was hard to follow. Section 3.1 now starts with one concrete example before any notation (L138–157 and Table 1). There is one frozen model. The source prompt asks for a literal description of a scene. The target prompt is the same prompt with "Style: vivid fantasy" and is used only offline. The candidate is the source run, optionally passed through an existing editor, and the repaired output is the candidate plus PCRT, a correction added after the final normalization. Table 1 shows the actual outputs for one held-out scene. The AcT candidate writes fantasy but loses the rain and the umbrellas, and the repaired output brings them back.

Second, all three reviewers asked why an existing editor is needed. It is not, and the paper now says so directly (L95–102, L463–470). PCRT applied to the unedited source is our default. The target distribution itself shows why. In our settings AcT is farther from the target than the unedited source on all 64 test contents of each of the three additional backbones (new Table 40), and the candidate that starts farther also ends farther. Post-editor PCRT remains useful as a repair when an editor is already fixed.

Third, the reviewers asked whether the method is more than a constant shift and whether lower KL means better behavior. We added results on both. On the three additional backbones a paired residual mean removes 10.8% to 42.6% of AcT sequence KL, whereas the conditional predictor removes 69.6% to 74.2% and is better on 64, 64 and 63 of 64 contents (Table 39). For direct correction, strict IFEval compliance rises from 7.9% to 52.0%, and 31 of the 73 items go from failing to passing while none goes the other way (L452–461, Table 41). For post-editor correction we added question answering and sentiment tasks with an editor whose own output is usable. Joint answer-and-format success is 0/48, 11/64, 0/48 and 53/64 for the editor alone. It rises to 14/48, 60/64, 31/48 and 61/64 with the closed-form correction, and to 37/48, 57/64, 44/48 and 61/64 with a KL-fitted version of the correction that the revision adds to the method (L256–259). IFEval format compliance rises from 17 to 33 and from 11 to 34 of 73 items (Table 4, Appendix S).

We have also tried to be exact about what the paper does not show. PCRT is not faster than simply supplying the target prompt (1.949 s against 1.944 s). Matched PSR and an all-layer parameter student, which train components at every layer, are more accurate than a single map at the output (L508–517, Appendix S.5). We do not claim that correcting an editor is better than correcting the source, and the output-head metric gives no measured accuracy gain over Euclidean fitting. The theory uses standard identities. What we contribute is the same-content, same-prefix target, an estimator of sequence KL against it that applies to any editor, a closed-form correction at the output, and evidence that the error editors leave is structured and can be measured and reduced.

Replies to each reviewer follow.

We have uploaded a second revision. It adds one experiment on a point that all three reviews raise in different words, namely whether the correction changes what the model actually outputs, with and without an existing editor.

We ran the estimator exactly as submitted, with no tuned quantity, on question answering (SQuAD), sentiment classification (SST-2) and the 73 IFEval items, on Qwen 2.5 7B and Qwen 3 4B, with three training replicates. For the first two tasks a success is the exact answer in the instructed JSON format, and for the third it is a pass of the official strict format check. The source prompt scores 0 on the first two tasks because it answers in free text.

Task success, range over three replicatesSource promptDirect PCRTEditor aloneEditor + PCRTTarget prompt
Qwen 2.5, QA (of 48)033 to 37035 to 4244
Qwen 2.5, sentiment (of 64)058 to 6011 to 1259 to 6059
Qwen 2.5, instructions (of 73)733 to 3717 to 1837 to 3949
Qwen 3, QA (of 48)040 to 42038 to 4143
Qwen 3, sentiment (of 64)06153 to 5761 to 6261
Qwen 3, instructions (of 73)429 to 3511 to 1335 to 3647

The editor here is a response-aligned transport editor whose own output stays fluent. Three results follow from the table. Direct correction of the source raises success in all 18 model, task and replicate cohorts. Correction after the editor also raises success in all 18, and 17 of these gains are significant after Holm correction. A constant correction fitted on the same pairs changes success by at most three items in every cohort, so the gain comes from reading the state and not from adding an offset. The direct and post-editor routes do not differ significantly in any cohort, which agrees with our earlier answer that an editor is optional.

The new results are in Table 4 and Appendix S.4 (Tables 45 to 47). Every line, table and appendix number quoted in our first replies still points to the same place. The post-editor numbers quoted there, such as 14/48 and 37/48, came from a rank-32 least-squares fit and a KL-fitted version with one seed. They are now in Tables 42 to 44, and Table 4 reports the submitted estimator.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.