acceptodds
Under review as a conference paper at ICLR 2027

Same Delta, Different Effects: Reusing Post-Training Updates after Continual Pretraining

Abstract

Adding a post-training delta after continual pretraining (CPT) can recover instruction following. Does a similar instruction gain mean that the update corrects the same failures at the same cost to language adaptation? We study a fixed delta, defined as the weight difference between a post-trained model and its base, along a Qwen3.5-9B Tibetan CPT trajectory. Overall instruction gains remain similar, while the language-modeling penalty in bits per byte (BPB) grows roughly fourfold. The corrected prompts also change, even within a fixed set that all receiving checkpoints fail. Scaling reveals a related distinction between preserving scores and preserving individual improvements. At roughly half the BPB cost, a scaled update retains 96% of the full update's net IFEval gain but corrects only 81% of the prompts newly corrected by the full update. Its net IFBench gain retention is 31%. Restricting transfer to lower-half feed-forward blocks shows no clear advantage over scaling at matched cost. An external Llama-3.1-8B/Basque comparison also finds instruction gains with a BPB penalty, although that penalty decreases relative to the base point. These results show why successful reuse requires more than recovering an average instruction score: the same update can change which prompts improve and its cost to language-model fit.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Continual pretraining, Post-training transfer, Model merging, Instruction following, Language adaptation

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_LaJVOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work examines adding a fixed post-training weight delta after continued pretraining. It measures instruction-following gains, language modeling penalties, and changes in corrected prompts along a Qwen3.5-9B Tibetan trajectory. Scaling, selective transfer, and a matched-cost control are complemented by a Llama-3.1-8B Basque comparison. Similar aggregate gains can conceal different behavioral effects, and selective transfer shows no clear advantage over matched-cost scaling.

Critical strengths
  1. Controlled comparisons: Measures each update’s gains and costs against the checkpoint receiving it, and compares selective transfer with full-vector scaling at matched language-modeling cost. This helps distinguish benefits of parameter selection from simply applying a smaller update.
  2. Prompt-level analysis reveals changes hidden by aggregate scores.
  3. Relevant baselines, including comparisons among base, post-trained, continually pretrained, and delta-merged models.
Critical weaknesses
  1. The paper lacks a CPT→fresh post-training baseline, leaving unclear how delta reuse compares with directly post-training the adapted model.
  2. Limited task coverage: English behavioral evaluation concentrates on IFEval and IFBench, without dedicated coding or mathematical-reasoning benchmarks. Although Tibetan tasks include mathematics, broader capability preservation remains insufficiently explored.
  3. Target-language instruction results are inconclusive and frequently reaching the generation limit.
  4. Scaling: Only 8B–9B models are tested, leaving scale dependence unknown.
  5. Reproducibility and robustness: Some training details are unavailable or inconsistent. Confidence intervals capture evaluation-sample uncertainty but do not establish robustness across independent training runs
Questions for authors
  1. Do the trajectory and matched-cost findings replicate across model families?
  2. Do the Tibetan instruction-following conclusions change when stopping behavior is checked and generation limits are increased?
3 comments
AuthorsOpenReview ↗

Matched pipeline comparisons, coding gains, and independent training strengthen the evidence

Thank you for the positive assessment and the concrete questions. We added a matched-configuration pipeline comparison, independent training runs, coding and math evaluations, and longer Tibetan generation. The new results clarify both when reuse works and how to choose an adaptation pipeline.

W1. Full reuse retains IFBench gains under matched training. Revised Table 4 compares three Qwen3.5-9B 500M endpoints trained with one configuration and seed, evaluated in one runtime. Brackets denote 95% paired bootstrap intervals.

EndpointTibetan BPBIFEvalIFBench
Base→CPT.377661.222.0
Base→CPT + Δ.388674.327.0
Post→CPT.377775.821.7

Full reuse has similar IFEval scores to Post→CPT (−1.5 pp [−4.4, 1.5]) and higher IFBench scores (+5.3 pp [1.7, 9.3]). Seed 42 Full also exceeds the same Post→CPT reference (+7.3 pp [3.3, 11.3]). Adding Δ costs .011 BPB, lowers 1 MiLiC direction and improves 2; Post→CPT has no significant MiLiC differences from Base→CPT. Together with the α=.75 dose result, this shows why choosing a pipeline from IFEval alone can miss a loss on IFBench (Sec. 6; App. C.3). This comparison tests adaptation order; we did not train a fresh post-CPT SFT baseline.

W2. Reuse also improves coding. At the original 500M receiver, HumanEval+ accuracy rises from 77.4% at CPT to 83.5% with Δ, the observed Post score: +6.1 pp [1.2, 11.6]. Changes on GSM8K and MBPP+ are +0.6 pp [−0.7, 1.9] and +1.1 pp [−2.4, 4.2]. These evaluations extend the evidence beyond instruction constraints (Sec. 7; App. C.7).

W5. Matched configurations and independent seeds address robustness. The original Base→CPT and Post→CPT used different training configurations. We trained the new Base→CPT runs with the Post→CPT configuration. Revised Table 4 replaces the submitted pipeline table; the cross-configuration results are labeled in App. C.4. Sec. 3.3 and App. A.1 document both recipes. The original run had no warmup; the 200-step field was not applied.

The two new seeds yield 500M BPB costs of .0110 and .0111, with all 12 instruction-gain intervals above zero. On fixed common-failure prompts, checkpoint differences in correction status exceed seed differences by 11.8 pp [6.6, 17.5] on IFEval and 6.8 pp [3.9, 9.7] on IFBench (App. C.1). This directly supports receiver dependence beyond the observed training variation. Code, configurations and Tibetan prompt sets are included in the supplement.

W3 / Q2. Longer Tibetan generation and paired testing clarify the transfer gap. We checked stopping and regenerated the original 500M endpoints at 8,192 tokens. Of 1,062 outputs capped at 3,072, 1,034 remain capped; no strict score changes by more than 1.3 pp. On matched English–Tibetan prompt pairs, the English Full−CPT gain exceeds the Tibetan gain by 11.9 pp [3.7, 20.1] on IFEval and 16.5 pp [2.5, 30.4] on IFBench. English uses 1,536 tokens. Both intervals exclude zero despite the larger Tibetan budget. Thus, English gains cannot substitute for target-language evaluation (Sec. 4.2; App. C.6).

Q1. Replication and matched-cost interpretation. Four Qwen trajectories and an external Llama CPT endpoint now show positive instruction gains. Llama's cost is 0.72× its Post−Base cost, and its corrected prompts change with the receiver (Sec. 4.3). The calibrated Lower comparison is specific to the original Qwen endpoint. In Llama×Basque, Lower is 8.5 pp IFEval above an interpolation between α=.25 and α=.5 at its BPB. Those doses are not a calibrated control. We therefore retain the general requirement for a matched-cost comparison, while keeping the null location result specific to the tested endpoint (Sec. 7; App. C.5).

W4. Model scale. The main evidence covers Qwen3.5-9B and Llama-3.1-8B. A 2B extension did not reproduce the robust instruction gain after CPT; Sec. 7 states the 8–9B scope and App. C.8 reports the extension.

Reviewer_LaJVOpenReview ↗

Response to Authors

Thank you for the detailed response and additional experiments. The coding and math evaluations address the concern of limited evaluations. The receiver-dependence findings remain informative.

However, two important concerns remain:

  1. Missing CPT→fresh post-training baseline: The paper does not establish how delta reuse compares with directly post-training the adapted model, which is the standard approach. This baseline is important for assessing the practical quality–cost trade-off and supporting pipeline-selection conclusions.
  2. Generality remains limited - the 2B experiment doesn't reproduce the gains. The trajectory analysis and calibrated matched-cost findings remain specific to Qwen; their replication across families is not established.

The above limitations leave the broader practical conclusions insufficiently supported. Therefore, I change me score to a weak acceptance to a weak reject.

Thank you for clarifying the two concerns behind your revised assessment. We have now completed three additional experiments that directly address these concerns: fresh post-CPT SFT, a directly calibrated matched-cost comparison on Llama, and an independent 2B training run. The new evidence substantially strengthens both the practical and cross-model conclusions.

1. Direct comparison with fresh post-CPT SFT.

We evaluated both approaches starting from the identical Qwen3.5-9B Tibetan CPT checkpoint. Fresh SFT updated all parameters for one epoch on 50,000 instruction examples, including Tibetan, Chinese, and English data (1,671 steps; 57.48M processed tokens). Delta reuse applied the existing Post−Base update without additional training.

ConditionIFEval ↑IFBench ↑Tibetan BPB ↓
CPT alone61.18%22.00%0.37761
CPT → fresh SFT66.73%21.00%0.39240
CPT + full delta74.31%27.00%0.38858

Delta reuse outperforms fresh SFT on both instruction benchmarks: +7.58 pp on IFEval (paired 95% CI [2.96, 12.01]) and +6.00 pp on IFBench [0.33, 11.33]. It also achieves better Tibetan BPB (0.38858 versus 0.39240).

This directly addresses the missing-baseline concern and establishes a concrete quality–cost advantage: given an available donor Post model, delta reuse achieves stronger instruction following and better Tibetan language-modeling fit than fresh post-CPT SFT, without an additional 57.48M-token training run.

2. Matched-cost findings now extend to Llama.

We have replaced the earlier Llama interpolation with a directly evaluated, calibrated full-delta control. At the Llama-3.1-8B/Basque CPT endpoint, we calibrated the full-delta coefficient (α = 0.3496) against the lower-half MLP update using separate Basque text, then froze it before evaluation.

Lower − calibrated scalingQwen-9B / TibetanLlama-8B / Basque
IFEval, pp [95% CI]+2.03 [−1.48, 5.37]+11.65 [7.39, 15.90]
IFBench, pp [95% CI]+1.00 [−2.67, 4.67]+1.00 [−2.33, 4.67]

This directly addresses the concern that the matched-cost analysis had not been tested beyond Qwen. The Llama experiment reveals a significant IFEval benefit from lower-half selective transfer over full-vector scaling at closely matched language-modeling cost, whereas the corresponding Qwen comparison shows no clear advantage.

The finding is stronger than the earlier interpolation: parameter placement can yield additional instruction-following benefits beyond simply reducing the update magnitude, and the relative benefit differs across adaptation settings. Together with the receiver-dependent prompt corrections already observed in Llama, this extends key behavioral and matched-cost analyses beyond the original Qwen experiments.

3. Positive instruction transfer also occurs at 2B.

We completed an independent Qwen3.5-2B Tibetan CPT run at 500M tokens. In this run, full delta reuse improves IFEval by +2.22 pp [0.06, 4.38] and IFBench by +1.33 pp [0.15, 2.51], with both paired 95% confidence intervals excluding zero.

This provides new positive evidence at the smaller model scale, addressing the earlier impression that the 2B setting does not benefit from delta reuse. Although the gains are smaller than those observed at 8–9B and differ from the first 2B run, the independent experiment demonstrates that instruction transfer is not confined to the larger models.

Overall, the new experiments address both principal concerns motivating the score reduction. We now provide (i) the requested direct fresh-SFT baseline, demonstrating an instruction-quality and training-cost advantage for reuse; (ii) a calibrated matched-cost comparison in a second model family, revealing a significant selective-transfer benefit; and (iii) positive instruction-transfer evidence at 2B. These findings reinforce the paper's central contribution: the effectiveness of a reused post-training update depends not only on its average instruction gain but also on its receiving model, parameter placement, and language-adaptation cost.

We hope these additional results address the concerns underlying your revised assessment and warrant reconsideration of the score.

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

The authors study how adding a fixed post-training delta after continual pretraining (CPT) affects instruction following along a Qwen3.5-9B Tibetan CPT trajectory. Their scaling analysis finds that, at approximately half the BPB cost, a scaled update retains 96% of the full update’s net IFEval gain but corrects only 81% of the prompts newly corrected by the full update relative to the CPT receiver. Applying the delta only to lower-half feed-forward blocks shows no clear instruction-following advantage over scaling the full delta at matched BPB cost. A limited Llama-3.1-8B/Basque comparison finds that the BPB penalty decreases relative to the penalty at the original base model, unlike the increase observed in the Qwen/Tibetan study.

Critical strengths
  • The prompt-level analysis is useful because it examines individual outcomes beyond average scores. At approximately half the BPB cost, the scaled update retains 96% of the full update’s net IFEval gain but only 81% of its newly corrected prompts. Moving from the scaled update to the full delta fixes 28 IFEval prompts and breaks 25, showing that a small change in average score can hide substantial changes in which examples improve.

  • The selective-transfer experiment uses a fair control: the authors compare the lower-half MLP update with a scaled full delta at matched BPB cost. This tests whether selecting particular parameters provides a benefit beyond simply reducing the update’s strength. The tested lower-half MLP update shows no clear instruction-following accuracy advantage over scaling the full delta.

Critical weaknesses
  • The experiments cover only language-adaptation CPT on Tibetan and Basque. It remains unclear whether the findings extend to domain-adaptation CPT, such as code.

  • The authors evaluate only Qwen/Tibetan and Llama/Basque. As they acknowledge, this confounds model family and language, making it difficult to explain the different changes in BPB penalty relative to the original base point. The CPT training conditions also differ. Moreover, the detailed prompt-level analysis is conducted only for Qwen/Tibetan, leaving its generality across settings untested.

  • Target-language instruction-following gains for Tibetan are not clearly established, as the confidence intervals include zero. For Basque, target-language instruction following is not directly evaluated in this paper.

Questions for authors

Do you have CPT results for Qwen/Basque or Llama/Tibetan that could help distinguish language effects from model-family effects, ideally under comparable training conditions?

What accounts for the inconclusive Tibetan instruction-following results? Could you also provide a direct evaluation of instruction following in Basque?

3 comments
AuthorsOpenReview ↗

Requested Qwen×Basque experiment completed; prompt-level receiver effects extend across settings

Thank you for highlighting the prompt-level findings and the matched-cost control. We completed the Qwen×Basque experiment requested in Q1 and extended the correction analysis beyond Qwen×Tibetan. These additions strengthen the evidence that the receiver changes both the cost and the behavior of a reused update.

Q1 / W2. Qwen×Basque holds model and CPT recipe fixed. We trained Qwen3.5-9B on 500M Basque tokens from HPLT v2, filtered against BL2MP and EusReading. Two additional Tibetan runs use the same model, training configuration and token budget. The table reports CPT+Δ minus CPT; brackets are 95% intervals.

CPT tokensBPB cost of ΔIFEval gain (pp)IFBench gain (pp)
10M.041713.1 [9.1, 17.2]11.3 [6.3, 16.7]
250M.025515.2 [10.9, 19.4]8.3 [3.7, 13.3]
500M.025314.6 [10.5, 18.5]10.3 [5.0, 15.7]

All gains are positive. At 500M, the cost is 0.62× the Basque Post−Base cost, compared with 0.81× for both Tibetan seeds. Basque CPT raises BL2MP from 71.2% to 91.1% and EusReading from 55.7% to 68.8%; adding Δ yields 90.3% and 71.9%. Sec. 4.3 and App. C.2 now report this same-model comparison.

The new Tibetan and Basque runs both have endpoint costs below their Post−Base costs. The original Tibetan run uses a different configuration and reaches 4.0×. Language alone therefore does not explain the opposite cost directions in the submission. The revised finding is that cost depends on the receiver, with fourfold growth confined to the original trajectory (Sec. 4.1, Table 2).

W2. Similar gains can correct different prompts in both model families. In Llama×Basque, among 246 IFBench prompts failed by both Base and CPT, Δ corrects 44 at Base and 43 at CPT, but only 29 are shared. On the new Qwen trajectories, correction status changes for 50/194 and 52/199 common-failure IFBench prompts in Tibetan, and 50/202 in Basque. The new Tibetan runs also show checkpoint differences above seed differences: +11.8 pp [6.6, 17.5] on IFEval and +6.8 pp [3.9, 9.7] on IFBench. The identity of the corrected prompts therefore adds information that aggregate gains miss (Secs. 4.1, 4.3; Apps. C.1, C.5).

W3 / Q2. Longer decoding reveals persistent repetition; paired testing establishes a language gap. At an 8,192-token limit, 1,034 of 1,062 outputs capped at 3,072 remain capped. No strict score changes by more than 1.3 pp. Capped outputs compress to a median 2% of their size, versus 20–30% for terminated outputs, consistent with repetitive loops.

We compared gains on English prompts and their Tibetan translations with the same constraints. At the original 500M endpoint, English Full−CPT gains exceed Tibetan gains by:

Paired subsetEnglish−Tibetan gain (pp, 95% CI)
IFEval, 294 pairs+11.9 [3.7, 20.1]
IFBench, 79 pairs+16.5 [2.5, 30.4]

English uses 1,536 tokens and Tibetan 8,192. Both intervals exclude zero. English instruction gains are therefore not reproduced at a comparable magnitude in Tibetan, even with the larger budget (Sec. 4.2; App. C.6).

W1. Scope of the contribution. We study language-adaptation CPT across three model–language settings. Domain-adaptation CPT remains untested, as stated in Sec. 7. Within the tested scope, the added same-model comparison and prompt-level replication strengthen the main contribution: successful reuse must be assessed through both instruction gains and compatibility with the receiving model.

To complement the Llama IFBench result in our first reply, English IFEval provides a second test of prompt-level receiver dependence. We select the 370 prompts that both Llama Base and the Basque CPT model initially fail. This fixes their initial correctness on every selected prompt. Adding the same delta gives:

ReceiverPrompts corrected by ΔCorrected only at this receiver
Base25445
Basque CPT model24334

The two updated models correct 209 prompts in common. Another 45 are corrected only at Base and 34 only at CPT. Thus, 79 prompts change correction status, while the net difference is only 11 prompts. Similar correction counts conceal substantial differences in which failures the update fixes.

Together with the IFBench result, this extends the prompt-level finding to two English instruction benchmarks in the Llama/Basque setting (Appendix C.5). It strengthens the contribution beyond the Qwen/Tibetan analysis: the receiver affects which problems the same update corrects.

Thank you for emphasizing the importance of the matched-cost control and its generality. We have now extended this analysis to Llama-3.1-8B/Basque with a directly evaluated, calibrated full-delta baseline, replacing the interpolation used in our earlier response.

1. Matched-cost analysis now reveals a significant selective-transfer advantage on Llama.

At the Llama/Basque CPT endpoint, we calibrated the full-delta coefficient to match the lower-half MLP update on 256 separate Basque documents, then froze α = 0.3496 before evaluation. The calibration BPB gap was only 0.000021; on held-out Basque passages, Lower had 0.00105 higher BPB.

On English IFEval, Lower passes 258/541 prompts (47.69%), versus 195/541 (36.04%) for calibrated full-vector scaling. This yields a +11.65 percentage-point advantage (paired 95% CI [7.39, 15.90]). On IFBench, the difference is +1.00 pp [−2.33, 4.67].

This result goes beyond confirming the earlier interpolation. It demonstrates that selective parameter placement can provide substantial instruction-following benefits beyond simply reducing the full update's magnitude.

Importantly, the same calibrated comparison produces different outcomes across the two settings:

Lower − calibrated full scalingQwen-9B / TibetanLlama-8B / Basque
IFEval, pp [95% CI]+2.03 [−1.48, 5.37]+11.65 [7.39, 15.90]
IFBench, pp [95% CI]+1.00 [−2.67, 4.67]+1.00 [−2.33, 4.67]

The Qwen comparison shows no clear placement advantage, whereas Llama shows a significant advantage on IFEval. Thus, the direct cross-family comparison strengthens our central finding: the effectiveness of a transferred update depends on the receiving model and adaptation setting, and selective transfer can outperform uniform scaling under calibrated language-modeling cost.

This complements our earlier Llama prompt-level replication: among 370 common-failure IFEval prompts, 79 change correction status between Base and the Basque CPT receiver, despite only an 11-prompt difference in total corrections. Together, these findings extend both the matched-cost and receiver-dependent behavioral analyses beyond Qwen/Tibetan.

2. Basque-language evaluation (Q2).

We additionally evaluate Basque capabilities directly through BL2MP, EusReading, and BPB. On Llama/Basque, full delta reuse improves EusReading from 40.6% to 56.5% (+15.9 pp [10.4, 21.5]), while BL2MP remains comparable (92.5% versus 92.2%). These measurements provide complementary evidence about target-language capabilities after reuse.

Our IFEval and IFBench comparisons use English prompts consistently across model families; we report them separately from Basque-language task performance rather than treating English instruction gains as direct measurements of Basque-prompt instruction following.

Overall, the new calibrated experiment directly addresses your concern about extending the matched-cost analysis to a second model family. More importantly, it reveals a previously unobserved positive result: selective transfer can substantially outperform full-vector scaling at calibrated language-modeling cost. Combined with the replicated prompt-level receiver dependence, this provides stronger evidence for the paper's main contribution beyond the original Qwen/Tibetan setting.

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

This paper studies what happens when a fixed post training delta, the weight difference between a post trained model and its base, is reused across checkpoints of a continual pretraining trajectory in a low resource language setting. Along a Qwen 9B Tibetan CPT trajectory the English instruction gains from reuse stay similar while the bits per byte cost grows roughly fourfold, and the identity of the corrected prompts shifts even among prompts that every receiving checkpoint fails. The paper varies the update scale and a lower half parameter mask, using a matched cost control that calibrates a scaled full update to the same calibration bits per byte before making any selective transfer claim, and an external Llama Basque setting replicates the gain with cost pattern.

Critical strengths

(1)The matched cost comparison is the right idea, since masking changes both the location and the amount of an update. (2)Paired bootstrap intervals are reported throughout, the external Basque setting replicates the core pattern under a different model family, and the paper is honest about its unresolved comparisons, reporting wide intervals and capped outputs plainly.

Critical weaknesses

(1)The paper reports only one CPT trajectory and does not repeat training with different random seeds. Because there is no seed replication, the "receiver dependence" claim really rests on just two model/language settings, which is too few to generalize from. Likewise, the reported fourfold cost growth comes from this single trajectory, so it reads as an anecdote about one run rather than a reliable finding about how costly reuse actually is.

(2)The evidence for language adaptation currently relies only on bits-per-byte, but whether instruction following works in the target language is still unresolved. The confidence intervals span zero, so the effect is not statistically clear; moreover, over half of the Tibetan prompted responses hit the generation length cap, meaning a large part of the evaluation is truncated. As it stands, the paper cannot say whether instruction following actually recovers in the language of use.

3)The contribution is measurement and diagnosis rather than a new method. The measurement is careful, but a careful study of one trajectory and two settings is closer to a workshop report than a full ICLR paper. I recommend weak rejection. Before resubmitting, I encourage the authors to replicate across multiple seeds, extend across more model and language settings, and resolve the truncated target language instruction evaluation.

Questions for authors

Would two or three additional CPT seeds on the same Tibetan trajectory materially change the spread of the bits per byte penalty across checkpoints? Does raising the generation cap resolve the Tibetan prompted instruction comparison, and if the cap reflects a real deployment constraint, which endpoint would you actually recommend to a practitioner today? If the authors can address my concerns well in their rebuttal, I would be willing to raise my score.

2 comments
AuthorsOpenReview ↗

Independent seeds support receiver dependence; matched pipelines and longer Tibetan evaluation clarify reuse

Thank you for the concrete suggestions. We added independent training runs, a same-model language comparison, and longer Tibetan evaluation. The results strengthen both the reliability and the practical value of the receiver-dependent finding.

W1 / Q1. Two seeds reproduce the instruction gain. The new Qwen3.5-9B Tibetan runs use configuration 2, also used by Post→CPT. Values below are seed 17 / seed 42; gains compare CPT+Δ with CPT.

10M250M500M
IFEval gain (pp)12.4 / 11.619.8 / 18.513.1 / 10.7
IFBench gain (pp)10.0 / 9.36.3 / 9.05.0 / 8.3
BPB cost.0349 / .0313.0084 / .0087.0110 / .0111

All 12 gain intervals are above zero. The 500M BPB costs are 0.81× the Post−Base cost for both seeds, versus 4.0× in the original run. Qwen×Basque under the same configuration also has positive gains at all three checkpoints, with a 500M cost of 0.62×. The revision identifies receiver-dependent cost as the general finding and confines cost growth to the original trajectory (Sec. 4.1, Table 2; Apps. C.1–C.2).

W3. Receiver effects exceed seed variation. We fixed the prompts failed by all receivers of both new Tibetan runs. Correction differences between checkpoints exceed differences between seeds at the same checkpoint by 11.8 pp [6.6, 17.5] on IFEval and 6.8 pp [3.9, 9.7] on IFBench (Sec. 4.1; App. C.1). Thus, the receiving model affects which failures Δ corrects beyond the observed seed variation. Recovering an instruction score does not establish that the update fixes the same problems at the same adaptation cost. The dose and pipeline controls below show how this distinction affects the choice of reuse strategy.

Q3. Full reuse is preferable when English IFBench constraints matter. Revised Table 4 compares pipelines with one training configuration and seed at 500M, evaluated in one runtime:

EndpointTibetan BPBIFEvalIFBench
Base→CPT.377661.222.0
Base→CPT + Δ.388674.327.0
Post→CPT.377775.821.7

Full reuse has similar IFEval scores to Post→CPT (difference −1.5 pp [−4.4, 1.5]) and higher IFBench scores (+5.3 pp [1.7, 9.3]). Seed 42 Full exceeds the same Post→CPT reference by 7.3 pp [3.3, 11.3] on IFBench. Adding Δ costs .011 BPB, lowers 1 MiLiC direction and improves 2. The dose experiment gives a related result: α=.75 retains 96% of the full IFEval gain but 31% of the IFBench gain. These two interventions show why IFEval alone can give an incomplete recommendation (Secs. 5–6).

W2 / Q2. Longer decoding and paired prompts establish a language gap. We regenerated the original endpoints at 8,192 tokens. Each Tibetan prompt preserves its English original's constraints. Full−CPT gains at the original 500M endpoint are:

BenchmarkEnglish (1,536)Tibetan (8,192)English−Tibetan
IFEval (N=294)+12.9+1.0 [−5.1, 7.1]+11.9 [3.7, 20.1]
IFBench (N=79)+13.9−2.5 [−13.9, 8.9]+16.5 [2.5, 30.4]

Both paired differences exclude zero: the English gain is not reproduced at a comparable magnitude in Tibetan. Raising the cap leaves 1,034 of 1,062 previously capped outputs capped and changes no strict score by more than 1.3 pp. Repeated text dominates capped outputs (median compressed size 2%, versus 20–30% for terminated outputs). Sec. 4.2 and App. C.6 now report this distinction between persistent repetition and the measured language gap.

Language adaptation is also evaluated through MiLiC, beyond BPB. Δ lowers 8 of 11 directions at the original receiver but 1 of 11 at the new receiver. The task evidence reinforces the central result: the cost of the same update depends on where it is applied.

One additional result addresses your question about the target-language comparison: the English–Tibetan gain gap also persists under loose scoring. We use the same generated responses and prompt pairs as in our first reply. Only the scoring criterion changes from strict to loose.

Paired subsetEnglish gain minus Tibetan gain (pp, 95% CI)
IFEval, 294 pairs+11.2 [3.1, 19.4]
IFBench, 79 pairs+19.0 [6.3, 31.6]

Here, gain is the score of CPT+Δ minus CPT within each language. Both paired confidence intervals exclude zero. The comparison uses the original Qwen3.5-9B 500M endpoints, with 1,536 tokens for English and 8,192 for Tibetan.

Because the responses are fixed, this check shows that the measured gap persists when scoring is relaxed. It strengthens a practical implication of our study: recovering English instruction scores does not establish comparable benefits in the adapted language. Assessing update reuse therefore requires measuring those benefits directly, alongside the cost to language adaptation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

New evidence strengthens receiver-dependent reuse: independent seeds, Qwen×Basque, and matched pipelines

We thank the reviewers for recognizing the controlled comparisons and prompt-level analysis. We reuse the fixed weight difference Δ = Post − Base after CPT. The new evidence strengthens the central finding: instruction gains can persist while the adaptation cost and the corrected prompts change with the receiving model.

1. Replication beyond one trajectory. We added two 500M-token Qwen3.5-9B Tibetan runs and a Qwen×Basque run with the same training configuration. At 10M, 250M and 500M tokens, all 18 English IFEval and IFBench gains have 95% intervals above zero. The two Tibetan seeds yield BPB costs of .0110 and .0111 at 500M. Across four Qwen trajectories and an external Llama CPT endpoint, costs range from 0.6× to 4× the Post−Base cost in each language. The replicated finding is receiver-dependent cost; fourfold growth describes the original trajectory. See Secs. 4.1, 4.3 and new Table 2.

2. Changes in corrected prompts exceed seed variation. On a fixed set of prompts failed by all receivers of both new Tibetan runs, we compared correction differences across checkpoints with differences across seeds at the same checkpoint. Checkpoint differences are larger by 11.8 pp [6.6, 17.5] on IFEval and 6.8 pp [3.9, 9.7] on IFBench. Llama×Basque also shows similar correction counts but different corrected prompts. These results support receiver dependence beyond aggregate scores. See Sec. 4.1 and Apps. C.1, C.5.

3. Matched training gives a practical pipeline comparison. Base→CPT+Δ and Post→CPT have similar IFEval scores, but full reuse scores higher on IFBench by 5.3 pp [1.7, 9.3] and 7.3 pp [3.3, 11.3]. The two estimates compare the two Base→CPT seeds against the same Post→CPT reference. Adding Δ costs .011 BPB and lowers 1 of 11 MiLiC directions while improving 2. Together with the dose experiment, this shows why choosing a pipeline from IFEval alone can miss a loss on IFBench. See Sec. 6 and revised Table 4.

4. Target-language gains require separate evaluation. We regenerated Tibetan responses from the original 500M endpoints with an 8,192-token limit. On matched English–Tibetan prompt pairs, the English gain exceeds the Tibetan gain by 11.9 pp [3.7, 20.1] on IFEval and 16.5 pp [2.5, 30.4] on IFBench. English uses a 1,536-token limit. The larger Tibetan budget does not reproduce a comparable gain. See Sec. 4.2 and App. C.6.

5. Transfer extends beyond instruction benchmarks. At the original 500M receiver, Δ raises HumanEval+ accuracy from 77.4% to 83.5%, the observed Post score: +6.1 pp [1.2, 11.6]. GSM8K and MBPP+ are also reported. See Sec. 7 and App. C.7.

Revisions are marked in blue. Sec. 3.3 and App. A.1 document both training configurations; App. C provides the added experiments. The supplementary material contains code, configurations and Tibetan prompt sets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.