acceptodds
Under review as a conference paper at ICLR 2027

Readout and Relay: Temporal Functional Geometry in Autoregressive Language Models

Abstract

Autoregressive transformers use each hidden state both to predict the next token and to support later predictions. We ask how these two roles coexist within a single representation. We identify a compact current-output-sensitive subspace, the readout, and decompose future credit into a component that reuses readout directions and a current-insensitive relay outside this subspace. Three patterns emerge. First, future credit substantially reuses the readout; within these reused directions, current and future credit can align or oppose, giving rise to synergy and conflict. Second, the relay is temporally selective: the component associated with a given future loss mainly affects the corresponding future prediction, with much smaller effects at other positions. For predictions farther in the future, a smaller fraction of future credit lies in the readout, while a larger fraction is carried by the relay. Third, across post-training stages, the balance between readout reuse and temporally selective relay changes little, while current-output sensitivity becomes concentrated in fewer directions. This suggests that post-training sharpens which directions matter for the current prediction without substantially reorganizing how hidden states support present and future predictions. Building on these findings, we introduce Temporal Proximal Credit (TPC), a lightweight training rule that reduces credit from later positions along directions important for the current prediction, without changing inference. TPC improves supervised fine-tuning across two model families, and a matched control shows that both moderating future credit and directing that moderation toward current-sensitive directions contribute to the gains.

Primary area:
interpretability and explainable AI
Keywords:
Language Model, Representation learning, Supervised fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_wo8TOpenReview ↗
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
  1. Have you evaluated TPC with multiple random seeds?

  2. Could you clarify the MMLU-Pro accuracy calculation? Table 16 lists 560 questions, but the reported 58.71% does not match this denominator.

  3. How often is the additional generation step in Appendix K.2 triggered, and does it affect the results?

5 comments
AuthorsOpenReview ↗

[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)

All training results use a single seed, so the stability of the 0.50 and 0.95 point gains over Uniform is unclear.

Yes. We retrained all methods with four seeds.

Table R1. Greedy accuracy, mean ± std over seeds.

ModelMethodMATH500MinervaMMLU-ProBBHMuSRZebraLogicAvg
Qwen3-4B-BaseVanilla SFT82.25 ± 0.7552.67 ± 2.2559.36 ± 0.9483.30 ± 0.4453.57 ± 0.7941.55 ± 0.9862.12 ± 0.43
DFT77.55 ± 1.3455.97 ± 1.3959.21 ± 1.1276.43 ± 2.5751.88 ± 1.0439.15 ± 0.9660.03 ± 0.34
Uniform82.15 ± 0.6056.25 ± 2.2959.64 ± 0.5883.50 ± 0.8753.90 ± 1.1342.08 ± 0.5762.92 ± 0.63
TPC83.00 ± 0.9157.08 ± 2.3560.61 ± 1.5783.87 ± 0.5154.46 ± 1.0342.80 ± 0.8163.64 ± 0.81
Llama-3.2-3B-InstructVanilla SFT54.65 ± 0.8125.55 ± 1.9343.25 ± 0.4961.23 ± 0.8648.05 ± 0.6024.50 ± 0.7842.87 ± 0.54
DFT52.55 ± 0.1924.63 ± 1.8341.57 ± 0.6260.53 ± 1.7548.38 ± 0.3324.83 ± 0.4242.08 ± 0.42
Uniform54.70 ± 0.8427.57 ± 1.5043.25 ± 1.1960.93 ± 1.0748.28 ± 1.1425.85 ± 0.4443.43 ± 0.44
TPC55.65 ± 0.6428.31 ± 1.5644.14 ± 1.0561.07 ± 1.0749.57 ± 2.0425.15 ± 0.9843.98 ± 0.78

Table R2. Sampled accuracy on competition math with k=8k=8, (P: Pass@8, A: Avg@8).

ModelMethodAIME25 PAIME25 AAIME26 PAIME26 ABeyondAIME PBeyondAIME A
Qwen3-4B-BaseVanilla SFT27.50 ± 1.6717.92 ± 0.5931.67 ± 4.3013.54 ± 0.7216.50 ± 1.295.19 ± 0.75
DFT22.50 ± 5.0010.42 ± 1.4421.67 ± 3.338.85 ± 0.6311.00 ± 2.165.16 ± 0.57
Uniform26.67 ± 2.7217.81 ± 0.8623.33 ± 3.8511.98 ± 0.2120.00 ± 1.636.16 ± 0.28
TPC30.00 ± 2.7217.81 ± 1.1528.33 ± 3.3314.27 ± 0.4016.75 ± 0.965.50 ± 0.49
Llama-3.2-3B-InstructVanilla SFT5.00 ± 1.920.63 ± 0.246.67 ± 2.720.94 ± 0.524.50 ± 1.000.59 ± 0.12
DFT6.67 ± 4.710.94 ± 0.799.17 ± 5.001.25 ± 0.904.25 ± 0.500.78 ± 0.19
Uniform7.50 ± 5.001.56 ± 1.105.00 ± 1.920.73 ± 0.405.75 ± 0.501.00 ± 0.27
TPC12.50 ± 4.191.98 ± 0.8610.83 ± 3.191.46 ± 0.424.75 ± 2.220.78 ± 0.26

Table R3. Sampled accuracy on competition math with k=16k=16, (P: Pass@16, A: Avg@16).

ModelMethodAIME25 PAIME25 AAIME26 PAIME26 ABeyondAIME PBeyondAIME A
Qwen3-4B-BaseVanilla SFT34.17 ± 1.6717.60 ± 0.6535.00 ± 3.3313.75 ± 0.4522.75 ± 0.506.00 ± 0.27
DFT27.50 ± 1.6711.25 ± 0.8224.17 ± 1.679.01 ± 0.9115.75 ± 2.065.50 ± 0.46
Uniform34.17 ± 4.1918.18 ± 0.6728.33 ± 1.9213.02 ± 0.6325.00 ± 1.835.88 ± 0.22
TPC34.17 ± 3.1917.60 ± 0.4035.83 ± 5.0014.58 ± 0.7822.25 ± 0.965.48 ± 0.38
Llama-3.2-3B-InstructVanilla SFT14.17 ± 6.871.25 ± 0.5411.67 ± 1.921.04 ± 0.517.00 ± 1.410.63 ± 0.16
DFT11.67 ± 3.330.94 ± 0.2114.17 ± 3.191.51 ± 0.316.50 ± 1.290.75 ± 0.11
Uniform14.17 ± 1.671.67 ± 0.458.33 ± 4.300.73 ± 0.3610.25 ± 0.961.08 ± 0.27
TPC13.33 ± 5.441.20 ± 0.6013.33 ± 2.721.30 ± 0.268.00 ± 0.820.84 ± 0.04

Table R4. TPC − Uniform in greedy Avg, paired by seed.

ModelSeed 1Seed 2Seed 3Seed 4Mean ± std
Qwen3-4B-Base+0.50+0.58+0.88+0.90+0.72 ± 0.21
Llama-3.2-3B-Instruct+0.96+0.41+0.07+0.76+0.55 ± 0.39

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.

AuthorsOpenReview ↗

[Part 2/3] Response to Reviewer wo8T

2. Sensitivity to rank, shrinkage, and trust region (Weakness 2)

How fine tuning varies with rank, λ0\lambda_0, and ϵTR\epsilon_{\mathrm{TR}} is not shown; Appendix B studies output reconstruction only.

  • Rank r=64r = 64. 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 btb_t, so it leaves training essentially unchanged (Appendix H.2).
  • λ0=0.25\lambda_0 = 0.25. With μ^i\hat\mu_i 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.
  • ϵTR=0.1\epsilon_{\mathrm{TR}} = 0.1. 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.

  • λ0\lambda_0 enters only through si=1/(1+λ0μ^i)s_i = 1/(1+\lambda_0\hat\mu_i), and since 0≤1−si<10 \le 1-s_i < 1, the correction never exceeds the part of btb_t inside the basis: χ≤ϕ\chi \le \sqrt{\phi} for any λ0\lambda_0 (Appendix H.2). The median change of btb_t is therefore below about 21% and 19% for any λ0\lambda_0, against 5.2% and 6.7% at λ0=0.25\lambda_0 = 0.25, 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.

ModelBasisOverlap with TPC basisCorrection inside site ReadoutCosine with TPC correction
Qwen3-4B-BaseTPC, E[atat⊤]\mathbb{E}[a_t a_t^\top]10.411
Current output Jacobian0.580.350.69
Current loss gradient0.470.290.66
Uniform–0.100.18
Llama-3.2-3B-InstructTPC, E[atat⊤]\mathbb{E}[a_t a_t^\top]10.751
Current output Jacobian0.610.600.81
Current loss gradient0.530.710.89
Uniform–0.110.16

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.

AuthorsOpenReview ↗

[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 (%).

MethodMMLU-ProBBHMuSRZebraLogicAvg
Vanilla SFT25.617.719.872.233.8
Uniform29.315.922.671.034.7
TPC31.720.06.167.831.4
  • 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

  1. Over four seeds, TPC improves on Uniform in every seed under greedy decoding; sampled results show no consistent winner (Tables R1–R4).
  2. Rank, λ0\lambda_0, and ϵTR\epsilon_{\mathrm{TR}} follow from measurements made before training, and the update stays within a fixed bound for any λ0\lambda_0 and changes little when the metric itself is replaced (Table R5).
  3. 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.
  4. MMLU-Pro uses 50 questions per subject (700 in total).
  5. 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

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

This paper studies how a hidden state in an autoregressive transformer supports both the current prediction and later ones. It identifies a current-output-sensitive READOUT subspace and splits each future gradient into a shared component and a RELAY component. In particular, the shared part is Qw=PtgwQ_w = P_t g_w, while the RELAY part lies outside the READOUT. By using fixed-trajectory interventions on Qwen3-4B and the OLMo-2 lineage, the authors find that future credit partly reuses the READOUT, while RELAY acts mainly at its own target horizon and carries a large share of the target effect as the horizon grows. Across the OLMo-2 checkpoints, post-training mostly concentrates current sensitivity rather than changing this organization. The authors propose TPC motivated by those findings, which shrinks key/value-path credit along a fixed sensitivity basis during SFT. Overall, TPC gives average gains over vanilla SFT and a norm-matched Uniform control on two models.

Critical strengths
  • The paper offers a useful functional view of how current and future predictions share a representation. It combines gradient decomposition with hidden-state interventions to examine when different components affect predictions. In this way it connects the geometry to model behavior and seems useful for work on latent planning and multi-token objectives.
  • The analysis includes several useful controls, such as donor future gradients, horizon-matched permutation null, isotropic random directions, and suffix identity checks. Also, the claims are tested against what the construction alone would give, for example, the target KL of the true RELAY is several hundred times that of its donor. The authors also distinguish empirical findings from effects expected by construction. All those checks make the measured temporal patterns convincing.
  • TPC applies its correction within the ordinary backward pass and leaves inference unchanged. The Uniform control is a good design choice because it separates how much future credit is reduced from where it is reduced. I also appreciate that the authors acknowledge that moderation alone appears to explain most of the gain over SFT.
Critical weaknesses
  • The evidence for temporal selectivity is limited by the offsets measured in my opinion. Temporal bills include only k∈{0,1,2,4}k \in \{0,1,2,4\}, so the large target shares in Table 7 are normalized over a sparse set of positions. Since RwR_w comes from the target-loss gradient, a strong target effect is partly expected. The donor controls are useful, but effects at unmeasured nearby positions remain a bit unclear.
  • The added benefit of TPC over Uniform is not that clear. Based on what’s reported in Table 3, the average gains are 0.50 percentage points on Qwen and 0.95 on Llama, while these are from a single training seed. The authors do acknowledge this limitation, but consistency across benchmarks does not establish stability across training runs. Uniform already recovers more than half of TPC's gain over SFT, so uncertainty in the remaining gain matters when assessing the value of targeting current-sensitive directions. It is also unclear whether the shared learning rate was tuned for vanilla SFT.
Questions for authors
  • I do understand that it would require large amount of compute resource to run multiple times, but would it be possible to report the mean and std over multiple seeds for table 3, for vanilla SFT, Uniform and TPC since it would provide the reader with a clear idea of how stable the gains are across training runs, even 3 seeds on one of the two models would already help. Also, could you report RELAY bills over a short contiguous window, for example k=0,…,8k=0,\ldots,8? This would clarify whether the target peak remains narrow when nearby offsets are included.
  • How were λ=0.25\lambda = 0.25, ϵTR=0.1\epsilon_{\mathrm{TR}} = 0.1 and the learning rate chosen? Was there any selection done on the evaluation benchmarks? Also, a small sweep over λ\lambda would help show whether the gain is stable.
4 comments
AuthorsOpenReview ↗

[Part 1/3] Response to Reviewer faeo

Dear Reviewer faeo

First of all, thank you for the careful review and the positive assessment of our controls. We restate each concern in one line and answer it below.

1. Selectivity over contiguous offsets (Weakness 1, Question 1)

Bills cover only k∈{0,1,2,4}k \in \{0,1,2,4\}, so target shares are normalized over sparse offsets and nearby effects are unclear.

We measured the bill at every offset k=0,…,8k=0,\dots,8 for w=1,…,8w=1,\dots,8, with the same donor control.

Table R1. Relay bill over offsets k=0,…,8k=0,\dots,8.

wwPeak at t+wt+w (true)Peak at t+wt+w (donor)Margin over other offsets
10.73–0.880.27–0.3229–79×
20.62–0.740.18–0.1911–14×
30.54–0.670.11–0.148–12×
40.51–0.620.09–0.108–14×
50.48–0.590.07–0.1010–13×
60.47–0.550.06–0.099–13×
70.43–0.510.06–0.077–8×
80.46–0.540.06–0.077–10×

Peak: share of anchors whose bill peaks at t+wt+w (chance 1/9).

Margin: KL at t+wt+w ÷ the largest KL at any other offset k≥1k \ge 1.

  • As you expected, the peak is broader than the shares in Table 7 suggest, especially at large ww.
  • Still, the effect at t+wt+w is at least 7× that at any other offset, adjacent ones included.
  • Donor Relays peak at t+1t+1 for every ww and are weaker at t+wt+w.

The target remains the dominant offset when nearby positions are included, so we read the selectivity as relative rather than strict.

AuthorsOpenReview ↗

[Part 2/3] Response to Reviewer faeo

2. Stability of the gain over Uniform (Weakness 2, Question 1)

The gains of 0.50 and 0.95 points over Uniform come from a single seed.

We retrained all methods with four seeds.

Table R2. Greedy accuracy, mean ± std over seeds.

ModelMethodMATH500MinervaMMLU-ProBBHMuSRZebraLogicAvg
Qwen3-4B-BaseVanilla SFT82.25 ± 0.7552.67 ± 2.2559.36 ± 0.9483.30 ± 0.4453.57 ± 0.7941.55 ± 0.9862.12 ± 0.43
DFT77.55 ± 1.3455.97 ± 1.3959.21 ± 1.1276.43 ± 2.5751.88 ± 1.0439.15 ± 0.9660.03 ± 0.34
Uniform82.15 ± 0.6056.25 ± 2.2959.64 ± 0.5883.50 ± 0.8753.90 ± 1.1342.08 ± 0.5762.92 ± 0.63
TPC83.00 ± 0.9157.08 ± 2.3560.61 ± 1.5783.87 ± 0.5154.46 ± 1.0342.80 ± 0.8163.64 ± 0.81
Llama-3.2-3B-InstructVanilla SFT54.65 ± 0.8125.55 ± 1.9343.25 ± 0.4961.23 ± 0.8648.05 ± 0.6024.50 ± 0.7842.87 ± 0.54
DFT52.55 ± 0.1924.63 ± 1.8341.57 ± 0.6260.53 ± 1.7548.38 ± 0.3324.83 ± 0.4242.08 ± 0.42
Uniform54.70 ± 0.8427.57 ± 1.5043.25 ± 1.1960.93 ± 1.0748.28 ± 1.1425.85 ± 0.4443.43 ± 0.44
TPC55.65 ± 0.6428.31 ± 1.5644.14 ± 1.0561.07 ± 1.0749.57 ± 2.0425.15 ± 0.9843.98 ± 0.78

Table R3. Sampled accuracy on competition math with k=8k=8, (P: Pass@8, A: Avg@8).

ModelMethodAIME25 PAIME25 AAIME26 PAIME26 ABeyondAIME PBeyondAIME A
Qwen3-4B-BaseVanilla SFT27.50 ± 1.6717.92 ± 0.5931.67 ± 4.3013.54 ± 0.7216.50 ± 1.295.19 ± 0.75
DFT22.50 ± 5.0010.42 ± 1.4421.67 ± 3.338.85 ± 0.6311.00 ± 2.165.16 ± 0.57
Uniform26.67 ± 2.7217.81 ± 0.8623.33 ± 3.8511.98 ± 0.2120.00 ± 1.636.16 ± 0.28
TPC30.00 ± 2.7217.81 ± 1.1528.33 ± 3.3314.27 ± 0.4016.75 ± 0.965.50 ± 0.49
Llama-3.2-3B-InstructVanilla SFT5.00 ± 1.920.63 ± 0.246.67 ± 2.720.94 ± 0.524.50 ± 1.000.59 ± 0.12
DFT6.67 ± 4.710.94 ± 0.799.17 ± 5.001.25 ± 0.904.25 ± 0.500.78 ± 0.19
Uniform7.50 ± 5.001.56 ± 1.105.00 ± 1.920.73 ± 0.405.75 ± 0.501.00 ± 0.27
TPC12.50 ± 4.191.98 ± 0.8610.83 ± 3.191.46 ± 0.424.75 ± 2.220.78 ± 0.26

Table R4. Sampled accuracy on competition math with k=16k=16, (P: Pass@16, A: Avg@16).

ModelMethodAIME25 PAIME25 AAIME26 PAIME26 ABeyondAIME PBeyondAIME A
Qwen3-4B-BaseVanilla SFT34.17 ± 1.6717.60 ± 0.6535.00 ± 3.3313.75 ± 0.4522.75 ± 0.506.00 ± 0.27
DFT27.50 ± 1.6711.25 ± 0.8224.17 ± 1.679.01 ± 0.9115.75 ± 2.065.50 ± 0.46
Uniform34.17 ± 4.1918.18 ± 0.6728.33 ± 1.9213.02 ± 0.6325.00 ± 1.835.88 ± 0.22
TPC34.17 ± 3.1917.60 ± 0.4035.83 ± 5.0014.58 ± 0.7822.25 ± 0.965.48 ± 0.38
Llama-3.2-3B-InstructVanilla SFT14.17 ± 6.871.25 ± 0.5411.67 ± 1.921.04 ± 0.517.00 ± 1.410.63 ± 0.16
DFT11.67 ± 3.330.94 ± 0.2114.17 ± 3.191.51 ± 0.316.50 ± 1.290.75 ± 0.11
Uniform14.17 ± 1.671.67 ± 0.458.33 ± 4.300.73 ± 0.3610.25 ± 0.961.08 ± 0.27
TPC13.33 ± 5.441.20 ± 0.6013.33 ± 2.721.30 ± 0.268.00 ± 0.820.84 ± 0.04

Table R5. TPC − Uniform in greedy Avg, paired by seed.

ModelSeed 1Seed 2Seed 3Seed 4Mean ± std
Qwen3-4B-Base+0.50+0.58+0.88+0.90+0.72 ± 0.21
Llama-3.2-3B-Instruct+0.96+0.41+0.07+0.76+0.55 ± 0.39

With greedy decoding, TPC has the best average on both models and improves on Uniform in every seed, so the remaining gain is small but consistent across training runs. On the sampled competition sets, the differences among Vanilla SFT, Uniform, and TPC stay within the variation across seeds.

AuthorsOpenReview ↗

[Part 3/3] Response to Reviewer faeo

3. Choice of hyperparameters (Weakness 2, Question 2)

How were λ0\lambda_0, ϵTR\epsilon_{\mathrm{TR}}, and the learning rate chosen, and was anything selected on the evaluation benchmarks?

  • Learning rate. It was not tuned for any method, Vanilla SFT included. All methods share the same optimizer settings, so they differ only in how credit is assigned. None of the hyperparameters below was selected on the evaluation benchmarks.
  • λ0=0.25\lambda_0 = 0.25. With si=1/(1+λ0μ^i)s_i = 1/(1+\lambda_0\hat\mu_i) and μ^i\hat\mu_i 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.
  • ϵTR=0.1\epsilon_{\mathrm{TR}} = 0.1. 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.

Sensitivity of the update. To see how much these choices matter, we measured how the TPC update itself responds to them on held out credits.

  • λ0\lambda_0 enters only through sis_i, and since 0≤1−si<10 \le 1-s_i < 1, the correction never exceeds the part of btb_t inside the basis: χ≤ϕ\chi \le \sqrt{\phi} for any λ0\lambda_0 (Appendix H.2). The median change of btb_t is therefore below about 21% and 19% for any λ0\lambda_0, against 5.2% and 6.7% at λ0=0.25\lambda_0 = 0.25, 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 R6. TPC basis vs. bases built without future loss.

ModelBasisOverlap with TPC basisCorrection inside site ReadoutCosine with TPC correction
Qwen3-4B-BaseTPC, E[atat⊤]\mathbb{E}[a_t a_t^\top]10.411
Current output Jacobian0.580.350.69
Current loss gradient0.470.290.66
Uniform–0.100.18
Llama-3.2-3B-InstructTPC, E[atat⊤]\mathbb{E}[a_t a_t^\top]10.751
Current output Jacobian0.610.600.81
Current loss gradient0.530.710.89
Uniform–0.110.16

Summary

  1. Over contiguous offsets, the effect at the target remains at least 7× that at any other offset, and donor Relays show no horizon (Table R1).
  2. Over four seeds, TPC improves on Uniform in every seed under greedy decoding; sampled results show no consistent winner (Tables R2–R5).
  3. The update stays within a fixed bound for any λ0\lambda_0 and changes little when the metric itself is replaced (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

AuthorsOpenReview ↗

To Reviewer faeo

Dear Reviewer faeo,

We hope this message finds you well. As the discussion deadline approaches, we would appreciate any further feedback you may have on our responses. If any concerns remain unresolved or any points need clarification, we would be happy to address them.

Thank you again for your time and careful consideration of our work.

Best regards,

The authors

Reviewer_8cHbOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper defines Readout using the current-output Jacobian, and defines shared credit and Relay through the projections of future loss gradients inside and outside Readout. It uses interventions on fixed trajectories to measure the temporal localization of influence along different directions. The authors find that future gradients partially reuse directions that are currently sensitive, while long-range effects depend more heavily on Relay. Based on this, they propose TPC, which attenuates, during fine-tuning, the component of future key/value credit along currently sensitive directions. TPC achieves modest average gains over Vanilla SFT and Uniform on Qwen3-4B-Base and Llama-3.2-3B-Instruct.

Critical strengths
  1. The problem formulation and measurement pipeline are valuable. The paper distinguishes between “future information being decodable from a hidden state” and “which prediction a given direction actually changes,” then analyzes these questions step by step using the current-output Jacobian, future-gradient projections, fixed-trajectory interventions, and donor controls.
  2. TPC’s control design is clear. Uniform uses the same correction norm and changes only the correction direction, offering an initial way to distinguish the overall effect of reducing future credit from the additional effect of placing the adjustment along currently sensitive directions.
Critical weaknesses
  1. The mechanistic explanation is constrained by the construction and approximations. R_w is the projection of the target future gradient onto the orthogonal complement of Readout, so the target-position effect is partly built into the definition. The Temporal Bill also fixes the token trajectory. In addition, TPC uses a path-gradient proxy that may not cleanly isolate current sensitivity and adjusts only first-hop key/value future credit. The paper therefore has not fully established that this training rule faithfully implements the earlier Readout/Relay decomposition.
  2. Evidence for the method’s gains and generalizability is limited. TPC is evaluated on two models, with one training seed each, using mathematical-reasoning data and a single intermediate layer. Its average gains over Uniform are 0.50 and 0.95 percentage points, with no variance estimates or significance analysis. The post-training conclusion is also based on just one OLMo-2 lineage.
Questions for authors
  1. How robust are the projection results to the choice of Readout? Please report controls using a random rank-64 subspace and activation PCA, or explain why these controls are not appropriate. Also clarify whether “future energy” denotes Ωw2\Omega_w^2, a Fisher-weighted quantity, or ordinary gradient energy.
  2. Please explain how the Temporal Bill under a fixed token trajectory relates to genuine autoregressive generation, and provide a perturbation-magnitude sweep or, at minimum, a behavioral validation without a fixed trajectory.
  3. Please construct a sensitivity matrix using only current-loss gradients or the current-output Jacobian and compare it directly with the E[a_t a_t^\top] matrix used by TPC. Could the large amount of future-loss information in a_t itself be driving the performance gains?
5 comments
AuthorsOpenReview ↗

[Part 1/4] Response to Reviewer 8cHb

Dear Reviewer 8cHb

First of all, we sincerely thank you for your insightful feedback and highlighting the strengths of our work. We address your concern below and list the revisions at the end.

1. The target effect of RwR_w is built into its definition (Weakness 1)

RwR_w comes from the target gradient, so its effect at t+wt+w is expected.

We agree that a first-order effect at t+wt+w is built in. The construction does not fix (i) whether RwR_w also acts at nearby offsets, or (ii) whether any gradient projected the same way behaves alike. We therefore measured the bill at every offset k=0,…,8k=0,\dots,8 and compared each RwR_w with a donor Relay, built the same way from another problem's future gradient (Appendix F).

Table R1. Relay bill over offsets k=0,…,8k=0,\dots,8.

wwPeak at t+wt+w (true)Peak at t+wt+w (donor)Margin over other offsets
10.73–0.880.27–0.3229–79×
20.62–0.740.18–0.1911–14×
30.54–0.670.11–0.148–12×
40.51–0.620.09–0.108–14×
50.48–0.590.07–0.1010–13×
60.47–0.550.06–0.099–13×
70.43–0.510.06–0.077–8×
80.46–0.540.06–0.077–10×

Peak: share of anchors whose bill peaks at t+wt+w.

Margin: KL at t+wt+w ÷ the largest KL at any other offset k≥1k \ge 1.

  • True Relays peak at their own offset well above chance, and their bill centroid follows the horizon.
  • Donor Relays peak at t+1t+1 for every ww and are weaker at t+wt+w, so the projection alone creates no horizon.
  • The effect at t+wt+w is at least 7× that at any other offset, adjacent ones included.

The selectivity therefore comes from the pairing with each position's own future loss, not from the projection.

2. The temporal bill fixes the token trajectory (Weakness 1, Question 2)

How does a bill on a fixed trajectory relate to autoregressive generation?

The bill measures the direct effect of a perturbation with later tokens held fixed. In generation this is the full effect until a generated token changes; after that, cascades add to it. We check two things.

Perturbation size. Over a 100× range of ϵ\epsilon, the selectivity of the Relay is unchanged. Up to the value used in the paper (ϵ=0.01\epsilon=0.01), the measured change in the target loss is within 4% of its first order prediction.

Table R2. Perturbation size sweep.

ϵ\epsilonΔLt+w\Delta L_{t+w} ÷ first order predictionPeak at t+4t+4 (true Relay)
0.0011.000.48–0.54
0.0030.99–1.000.50–0.60
0.010.96–0.990.51–0.62
0.030.86–0.960.50–0.61
0.10.61–0.880.49–0.60

The model's own text. On each OLMo2 checkpoint's own greedy continuations, the fixed trajectory is the one the model generates. There the Relay still leaves the current loss unchanged (∣ΔLt∣≤1.2×10−6|\Delta L_t| \le 1.2\times10^{-6}), and its bill centroid still grows with ww:

Relay bill centroid (OLMo-2 Instruct)w=1w=1w=2w=2w=4w=4w=8w=8
Teacher forced text1.822.202.852.97
Own greedy text1.932.232.752.78
AuthorsOpenReview ↗

[Part 2/4] Response to Reviewer 8cHb

3. Whether TPC implements the decomposition (Weakness 1, Question 3)

TPC uses an approximate current path credit and only first hop key/value credit, and the future loss content of ata_t may drive the gains.

We agree that TPC is an inexpensive surrogate, not an exact implementation, which would need a 64 row Jacobian at every token. What TPC keeps from the analysis is the rule itself: shrink future credit along directions the current prediction relies on and pass the rest, applied where later positions first read the hidden state. Appendix H shows that both parts hold:

  • btb_t is almost purely future credit: self credit is only 1.6–6.4% of ∥bt∥\|b_t\|.
  • The TPC correction lies 41–76% inside the site Readout, against 9–11% for Uniform.

To test whether the future loss content of ata_t drives this (Q3), we built two current only bases, which use no future loss: one from the current output Gauss–Newton matrix E[Jx⊤FtJx]\mathbb{E}[J_x^\top F_t J_x] and one from the current loss gradient E[at(0)at(0)⊤]\mathbb{E}[a^{(0)}_t a^{(0)\top}_t]. We then applied the rule with each basis to the same held out credits.

Table R3. TPC basis vs. current only bases.

ModelBasisOverlap with TPC basisCorrection inside site ReadoutCosine with TPC correction
Qwen3-4B-BaseTPC, E[atat⊤]\mathbb{E}[a_t a_t^\top]10.411
Current output Jacobian0.580.350.69
Current loss gradient0.470.290.66
Uniform–0.100.18
Llama-3.2-3B-InstructTPC, E[atat⊤]\mathbb{E}[a_t a_t^\top]10.751
Current output Jacobian0.610.600.81
Current loss gradient0.530.710.89
Uniform–0.110.16

The current only bases share about half of the TPC basis, about 20× chance, and also place the correction in the site Readout. More directly, the corrections they produce point largely along the TPC correction (cosine 0.66–0.89), whereas Uniform does not (0.16–0.18). Removing future loss information therefore leaves TPC acting in nearly the same directions: what it shrinks is set by current sensitivity, and E[atat⊤]\mathbb{E}[a_t a_t^\top] is a cheap way to find those directions.

AuthorsOpenReview ↗

[Part 3/4] Response to Reviewer 8cHb

4. Robustness of the training gains (Weakness 2)

One seed per model, and gains over Uniform of 0.50 and 0.95 points without variance estimates.

We retrained all methods with four seeds.

Table R4. Greedy accuracy, mean ± std over seeds.

ModelMethodMATH500MinervaMMLU-ProBBHMuSRZebraLogicAvg
Qwen3-4B-BaseVanilla SFT82.25 ± 0.7552.67 ± 2.2559.36 ± 0.9483.30 ± 0.4453.57 ± 0.7941.55 ± 0.9862.12 ± 0.43
DFT77.55 ± 1.3455.97 ± 1.3959.21 ± 1.1276.43 ± 2.5751.88 ± 1.0439.15 ± 0.9660.03 ± 0.34
Uniform82.15 ± 0.6056.25 ± 2.2959.64 ± 0.5883.50 ± 0.8753.90 ± 1.1342.08 ± 0.5762.92 ± 0.63
TPC83.00 ± 0.9157.08 ± 2.3560.61 ± 1.5783.87 ± 0.5154.46 ± 1.0342.80 ± 0.8163.64 ± 0.81
Llama-3.2-3B-InstructVanilla SFT54.65 ± 0.8125.55 ± 1.9343.25 ± 0.4961.23 ± 0.8648.05 ± 0.6024.50 ± 0.7842.87 ± 0.54
DFT52.55 ± 0.1924.63 ± 1.8341.57 ± 0.6260.53 ± 1.7548.38 ± 0.3324.83 ± 0.4242.08 ± 0.42
Uniform54.70 ± 0.8427.57 ± 1.5043.25 ± 1.1960.93 ± 1.0748.28 ± 1.1425.85 ± 0.4443.43 ± 0.44
TPC55.65 ± 0.6428.31 ± 1.5644.14 ± 1.0561.07 ± 1.0749.57 ± 2.0425.15 ± 0.9843.98 ± 0.78

Table R5a. Sampled accuracy on competition math with k=8k=8 (P: Pass@8, A: Avg@8).

ModelMethodAIME25 PAIME25 AAIME26 PAIME26 ABeyondAIME PBeyondAIME A
Qwen3-4B-BaseVanilla SFT27.50 ± 1.6717.92 ± 0.5931.67 ± 4.3013.54 ± 0.7216.50 ± 1.295.19 ± 0.75
DFT22.50 ± 5.0010.42 ± 1.4421.67 ± 3.338.85 ± 0.6311.00 ± 2.165.16 ± 0.57
Uniform26.67 ± 2.7217.81 ± 0.8623.33 ± 3.8511.98 ± 0.2120.00 ± 1.636.16 ± 0.28
TPC30.00 ± 2.7217.81 ± 1.1528.33 ± 3.3314.27 ± 0.4016.75 ± 0.965.50 ± 0.49
Llama-3.2-3B-InstructVanilla SFT5.00 ± 1.920.63 ± 0.246.67 ± 2.720.94 ± 0.524.50 ± 1.000.59 ± 0.12
DFT6.67 ± 4.710.94 ± 0.799.17 ± 5.001.25 ± 0.904.25 ± 0.500.78 ± 0.19
Uniform7.50 ± 5.001.56 ± 1.105.00 ± 1.920.73 ± 0.405.75 ± 0.501.00 ± 0.27
TPC12.50 ± 4.191.98 ± 0.8610.83 ± 3.191.46 ± 0.424.75 ± 2.220.78 ± 0.26

Table R5b. Sampled accuracy on competition math with k=16k=16 (P: Pass@16, A: Avg@16).

ModelMethodAIME25 PAIME25 AAIME26 PAIME26 ABeyondAIME PBeyondAIME A
Qwen3-4B-BaseVanilla SFT34.17 ± 1.6717.60 ± 0.6535.00 ± 3.3313.75 ± 0.4522.75 ± 0.506.00 ± 0.27
DFT27.50 ± 1.6711.25 ± 0.8224.17 ± 1.679.01 ± 0.9115.75 ± 2.065.50 ± 0.46
Uniform34.17 ± 4.1918.18 ± 0.6728.33 ± 1.9213.02 ± 0.6325.00 ± 1.835.88 ± 0.22
TPC34.17 ± 3.1917.60 ± 0.4035.83 ± 5.0014.58 ± 0.7822.25 ± 0.965.48 ± 0.38
Llama-3.2-3B-InstructVanilla SFT14.17 ± 6.871.25 ± 0.5411.67 ± 1.921.04 ± 0.517.00 ± 1.410.63 ± 0.16
DFT11.67 ± 3.330.94 ± 0.2114.17 ± 3.191.51 ± 0.316.50 ± 1.290.75 ± 0.11
Uniform14.17 ± 1.671.67 ± 0.458.33 ± 4.300.73 ± 0.3610.25 ± 0.961.08 ± 0.27
TPC13.33 ± 5.441.20 ± 0.6013.33 ± 2.721.30 ± 0.268.00 ± 0.820.84 ± 0.04

Table R6. TPC − Uniform in greedy Avg, paired by seed.

ModelSeed 1Seed 2Seed 3Seed 4Mean ± std
Qwen3-4B-Base+0.50+0.58+0.88+0.90+0.72 ± 0.21
Llama-3.2-3B-Instruct+0.96+0.41+0.07+0.76+0.55 ± 0.39

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, no method is consistenly best, so we do not claim a gain there.

The rule site is the middle layer chosen by the depth analysis (Appendix C); several layers are left for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for AC and Reviewers

Thank you to all reviewers for their time and constructive feedback.

Strengths noted. A functional view that separates what is decodable from a hidden state from which prediction a direction actually changes, supported by thorough controls, and a matched Uniform control that separates how much future credit is reduced from where it is reduced.

Key concerns and our responses

  • Single-seed gains (all reviewers): We retrained all methods with four seeds. Under greedy decoding, TPC beats Uniform in every seed on both models.
  • Relay selectivity by construction (8cHb, faeo): Selectivity holds over contiguous offsets and is absent for donor gradients. It is also robust to perturbation size and to the model's own greedy text.
  • TPC vs. the decomposition (8cHb): Bases built without any future loss yield largely the same correction.
  • Hyperparameters, cost, and evaluation (faeo, wo8T): We clarified each of these.

(Reviewer wo8T noted that the additional experiments strengthened their confidence in the empirical results and addressed most of their concerns, maintaining a positive rating of 3.)

We would greatly appreciate it if the remaining reviewers could revisit these results. Thank you again for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.