- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies how hidden states support current and future predictions in autoregressive language models through a Readout/Relay decomposition and temporal interventions. Based on these findings, the authors propose Temporal Proximal Credit (TPC), which selectively attenuates future-path gradients along directions sensitive to the current prediction. Experiments on two model families show improvements over Vanilla SFT and several controls.
Critical strengths
-
Practical improvements. In Table~3, TPC improves the six-benchmark average over Vanilla SFT by 1.19 and 2.74 percentage points on Qwen3-4B-Base and Llama-3.2-3B-Instruct, respectively, achieving the highest average among the compared methods on both models.
-
Informative analysis and controls. Temporal interventions provide a useful functional perspective on current--future computation. The correction-norm-matched Uniform control helps distinguish general gradient attenuation from directional modification.
Critical weaknesses
-
Limited evidence of robustness across training runs. All training results use a single seed. Although the overall gains over Vanilla SFT are encouraging, the additional gainsover Uniform are 0.50 and 0.95 percentage points. Their stability across training runs remains unclear; paired multi-seed results would strengthen the empirical conclusion.
-
Hyperparameter sensitivity and training costs are insufficiently characterized. The paper does not show how fine-tuning performance varies with rank, shrinkage strength, or the trust-region bound. Appendix~B studies output reconstruction rather than training outcomes. Calibration time, training throughput, and additional peak memory are also not quantified, leaving the practical cost and tuning requirements unclear.
Questions for authors
-
Have you evaluated TPC with multiple random seeds?
-
Could you clarify the MMLU-Pro accuracy calculation? Table 16 lists 560 questions, but the reported 58.71% does not match this denominator.
-
How often is the additional generation step in Appendix K.2 triggered, and does it affect the results?
5 comments
[Part 2/3] Response to Reviewer wo8T
2. Sensitivity to rank, shrinkage, and trust region (Weakness 2)
How fine tuning varies with rank, , and is not shown; Appendix B studies output reconstruction only.
- Rank . The Readout preserves the current output at rank 64 and saturates there (Table 4). A rank 1 metric changes a median of less than 0.05% of , so it leaves training essentially unchanged (Appendix H.2).
- . With normalized to mean 1, a mode of average sensitivity keeps 80% of its future credit, and a mode four times as sensitive keeps 50%. Since shared credit both agrees and conflicts with current credit (§3.3), we chose a value that limits this credit rather than removes it.
- . It caps the change at each token at 10% of the SFT gradient, so TPC stays a small perturbation of SFT and every mode keeps its sign (Appendix I). Uniform uses the same cap.
To see how much these choices matter, we measured how the TPC update itself responds to them on held out credits.
- enters only through , and since , the correction never exceeds the part of inside the basis: for any (Appendix H.2). The median change of is therefore below about 21% and 19% for any , against 5.2% and 6.7% at , and the trust region caps it further.
- Even replacing the metric itself with bases built without any future loss leaves the correction largely along the TPC correction.
Table R5. TPC basis vs. bases built without future loss.
| Model | Basis | Overlap with TPC basis | Correction inside site Readout | Cosine with TPC correction |
|---|---|---|---|---|
| Qwen3-4B-Base | TPC, | 1 | 0.41 | 1 |
| Current output Jacobian | 0.58 | 0.35 | 0.69 | |
| Current loss gradient | 0.47 | 0.29 | 0.66 | |
| Uniform | – | 0.10 | 0.18 | |
| Llama-3.2-3B-Instruct | TPC, | 1 | 0.75 | 1 |
| Current output Jacobian | 0.61 | 0.60 | 0.81 | |
| Current loss gradient | 0.53 | 0.71 | 0.89 | |
| Uniform | – | 0.11 | 0.16 |
3. Training cost (Weakness 2)
Calibration time, throughput, and peak memory are not quantified.
The metric is estimated once before fine tuning. Fine tuning processes 20k sequences for two epochs, so calibration adds about 1.3% of the fine tuning passes. During training, TPC needs no additional model, optimizer state, or forward pass: it adds one extra autograd branch of a single attention input and a rank 64 projection per token at one layer. Peak GPU memory is therefore similar for Vanilla SFT, Uniform, and TPC. Uniform and TPC are slightly slower than Vanilla SFT because they adjust credit at the rule site, but the difference is small. Inference is unchanged.
4. MMLU-Pro accuracy (Question 2)
Table 16 lists 560 questions, but 58.71% does not match this denominator.
Thank you for catching this. The evaluation uses 50 questions per subject, 700 in total; Table 16 and Appendix K.1 state 40 per subject (560) by mistake. Every MMLU-Pro entry in Table 3 is computed over the 700 questions; for example, 58.71% is 411/700.
[Part 3/3] Response to Reviewer wo8T
5. The additional generation step in Appendix K.2 (Question 3)
How often is the additional generation step triggered, and does it affect the results?
The step exists so that every response can be scored even when it does not follow the answer format. It runs only when a response contains no \boxed{}, either because the model stated its answer in another form or, more often, because it reached the 4096 token limit. We then append Answer:\boxed{ and let the model continue greedily from its own output; and a continuation without a valid answer counts as incorrect. The step is identical for all methods.
Table R6. Share of items that needed the step (%).
| Method | MMLU-Pro | BBH | MuSR | ZebraLogic | Avg |
|---|---|---|---|---|---|
| Vanilla SFT | 25.6 | 17.7 | 19.8 | 72.2 | 33.8 |
| Uniform | 29.3 | 15.9 | 22.6 | 71.0 | 34.7 |
| TPC | 31.7 | 20.0 | 6.1 | 67.8 | 31.4 |
- The rate is similar across methods. It is highest on ZebraLogic, where long puzzles often exceed the token limit.
- Answers obtained through the step are mostly wrong, with similar accuracy across methods.
- The step raises accuracy by a similar amount for every method, and least for TPC, so it does not account for TPC's gain.
Summary
- Over four seeds, TPC improves on Uniform in every seed under greedy decoding; sampled results show no consistent winner (Tables R1–R4).
- Rank, , and follow from measurements made before training, and the update stays within a fixed bound for any and changes little when the metric itself is replaced (Table R5).
- Calibration is a one time cost of about 1.3% of the fine tuning passes; peak memory is similar to Vanilla SFT, training is only slightly slower, and inference is unchanged.
- MMLU-Pro uses 50 questions per subject (700 in total).
- The additional generation step recovers answers from responses that miss the answer format; it is applied identically, and adds the least accuracy to TPC (Table R6).
We hope these responses have adequately addressed the concerns raised. Should any points remain unclear or raise further questions, we are happy to discuss them and will respond promptly. We welcome any additional feedback the reviewer may have.
Best regards
All authors
[Part 1/3] Response to Reviewer wo8T
Dear Reviewer wo8T
First of all, thank you for the careful review and for checking the evaluation details closely. We restate each concern in one line and answer it below.
1. Robustness across training runs (Weakness 1, Question 1)
Yes. We retrained all methods with four seeds.
Table R1. Greedy accuracy, mean ± std over seeds.
Table R2. Sampled accuracy on competition math with k=8, (P: Pass@8, A: Avg@8).
Table R3. Sampled accuracy on competition math with k=16, (P: Pass@16, A: Avg@16).
Table R4. TPC − Uniform in greedy Avg, paired by seed.
With greedy decoding, TPC has the best average on both models and improves on Uniform in every seed, a small but consistent gain. On the sampled competition sets, the differences among Vanilla SFT, Uniform, and TPC stay within the variation across seeds.