acceptodds
Under review as a conference paper at ICLR 2027

Trajectory-to-Go: Forecasting Changes in Language Model Predictions

Abstract

A language model's intermediate prediction can change little from one step to the next yet still differ substantially from its final output. We ask whether the last two steps help predict how much remains to change. Trajectory-to-Go (T2G) summarizes their magnitudes and changes in direction in both the model's predictions and its hidden states, then learns to forecast disagreement with a chosen later output. Its main readout, Risk-T2G, combines fourteen trajectory and confidence inputs in a two-head error forecast. On held-out Huginn-0125 and two Ouro scales, it reduces error-ranking loss (EAURC) by 7.7–22.3% over confidence-based predictors with matched training objectives and capacity, and outperforms a learned current-state probe in all three models. The fitted forecasts generalize to C4 and LAMBADA without adaptation. Across four ordinary transformers up to 7B parameters, T2G reduces EAURC by 10.8–21.1% over matched confidence. A state-augmented extension improves on either trajectory or state features alone in all eight model–lens settings. Stopping studies evaluate KL-constrained allocation; an answer-change extension tests decision preservation and execution cost. Recent trajectory history thus helps distinguish predictions that look stable from those close to a later output.

Primary area:
foundation or frontier models, including LLMs
Keywords:
looped transformers, language models, adaptive computation, intermediate readouts, selective prediction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PS7BOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether recent changes in a language model’s intermediate predictions can predict disagreement with a later output. The authors introduce Trajectory-to-Go (T2G), which summarizes two successive updates in output distributions and hidden states using magnitude and direction features. The main predictor, Risk-T2G, combines two regression heads to forecast terminal KL divergence. Experiments on recurrent-depth and conventional transformers report improvements in error ranking over confidence-based controls, together with corpus-transfer and state-augmentation results. The paper also evaluates KL-constrained computation allocation and introduces Task-T2G, a separate answer-change classifier for early exiting on MMLU. The authors conclude that recent trajectory information can help distinguish locally stable predictions from predictions that are close to a later output.

Critical strengths
  1. The paper provides a concrete formulation for predicting disagreement with a later model output. The reported results offer some evidence that recent trajectory features improve terminal-error ranking over matched confidence controls in the evaluated settings.

  2. The evaluation includes several model families and corpus-transfer checks, providing useful information about where the features help. The paper also acknowledges several limitations, including model-dependent feature contributions and the lack of a demonstrated latency advantage over learned controls.

Critical weaknesses
  1. The distinction from prior methods is not sufficiently discussed.

    • CATs (Schuster et al., 2021) already learns intermediate-to-final prediction agreement from prediction history, but it is not discussed.

    • The paper discusses CALM (Schuster et al., 2022) in Section 2 and provides a CALM-inspired state baseline for the main KL prediction task in Appendix C.3. However, it does not sufficiently explain the relationship between Task-T2G and CALM’s learned consistency classifier.

  2. The experiments do not fully establish the method’s contribution.

    • The specific contribution of the proposed geometric representation is not adequately isolated. The forecasting evaluation lacks a generic history predictor with the same prediction target, observation window, and comparable capacity. Improvements over confidence predictors demonstrate that trajectory information is useful, but do not establish the contribution of this particular geometric encoding relative to other ways of representing the same history.

    • The contribution of the complete Task-T2G feature combination is unclear. On Ouro-2.6B in Table 15, Task-T2G has a higher AURC (10.90) than inverse margin (10.13), linear confidence (10.28), the variant without the gate (10.48), and the variant without hidden features (10.62). The authors acknowledge model-dependent feature contributions and distinguish exit coverage from AURC. Nevertheless, the complete combination provides no ranking advantage over these simpler alternatives in the reported point estimates.

    • The practical contribution of Task-T2G is also limited by the reported comparisons. The authors explicitly acknowledge that they have not demonstrated a latency advantage over the learned control. On Ouro-1.4B in Table 14, the simpler inverse-margin policy also has a lower answer-change rate (1.46% versus 1.86%), higher accuracy (66.21% versus 66.11%), and greater time savings (14.23% versus 13.02%) than Task-T2G. Higher exit coverage therefore does not establish a better accuracy–consistency–latency trade-off. The paper does not report paired confidence intervals for these baseline differences. Moreover, omitting fixed pass 3 leaves the contribution of adaptive stopping relative to reduced fixed computation unresolved.

Questions for authors
  1. Supplementary materials. The package described in the reproducibility statement on page 10 is absent from the submitted materials. Please provide the stated code, prompt hashes, saved fits and thresholds, evaluation scripts, results, and table checks in an anonymous artifact accessible to reviewers.

  2. Relationship to CATs. Please compare T2G with CATs (Schuster et al., 2021) in model architecture, prediction target, historical inputs, and predictor design.

  3. Relationship to CALM. Please compare Task-T2G with CALM’s learned consistency classifier (Schuster et al., 2022) in architecture, input features, and prediction target.

  4. History baseline. Please add a CATs-inspired history predictor using ordinary prediction statistics and projected hidden states from the same three intermediate readings, without T2G’s engineered trajectory features.

  5. Fixed-depth comparison. Please add fixed pass 3 to the MMLU evaluation on both Ouro models.

7 comments
AuthorsOpenReview ↗

Response to Reviewer PS7B

Thank you for the detailed review; the controls you asked for are now in the paper. Tables and sections refer to the revised PDF. Table 14 of the submission is now Table 6 and Table 15 is Table 19. Brackets are paired 95% intervals from 2,000 resamples of whole prompts or questions.

Q1. Supplementary package

The anonymous T2G_supplementary.zip accompanies the revision. It contains the code and evaluation scripts, prompt hashes, saved fits and thresholds, and the results of the looped-model forecasting, control and stopping experiments. Its table check (python check_tables.py, standard library only) confirms that 318 printed entries of Tables 1, 2, 5, 6, 12, 16, 19 and 20 match the saved results, and lists the result file and field behind each entry. Per-example arrays are recomputed from the public models with the included scripts. The transfer and layer-wise tables (Tables 3 and 4) are not part of the check.

W1, Q2, Q3. CATs and CALM

CATsCALM classifierRisk-T2GTask-T2G
ModelFrozen fine-tuned Transformer; a prediction head is trained at every layerEncoder-decoder trained with a loss at every decoder layer, then frozen while the classifier is fittedFrozen looped model, or frozen layer-wise model read through a lensFrozen Ouro
TargetWhether the current prediction equals the final oneWhether the current token equals the top-layer tokenKL from a later output distribution to the current one (continuous)Whether the pass-3 answer differs from pass 4
InputsCurrent hidden state, current prediction and confidence, earlier predictionsCurrent hidden state14 scalars: confidence, and the sizes and directions of the last two updates in output and hidden space15 scalars: the same construction on the four answer tokens, hidden motion, Ouro's gate
PredictorSmall MLPLinear classifierTwo regression heads, 119 quadratic terms eachLinear logistic classifier
Exit ruleCalibrated, with a consistency guaranteeCalibrated, with a consistency guaranteeThreshold on held-out prompts, empirical2% answer-change target on 507 calibration questions, empirical

Risk-T2G differs from CATs in the target (a continuous distance between distributions, not agreement of a task prediction), in the setting (successive passes of a recurrent model, which share its output head) and in how history enters (sizes and directions of two updates, not the states and predictions themselves). Task-T2G has the same kind of target as CATs and CALM and, like them, leaves the model unchanged; it differs in its trajectory inputs and in the recurrent setting. Section 2 now discusses both.

Q4 and W2. A CATs-inspired history predictor

Table 2 (new). The generic predictor sees the same three readings without engineered trajectory features: entropy, maximum probability and margin at each reading, both successive KL values, and the three hidden states after a random projection and PCA fitted on training prompts. It has the same target, losses, penalties and two-head rule as Risk-T2G, and 119 terms per head. EAURC ×1,000, lower is better:

ModelGeneric historyRisk-T2GGeneric − T2G
Huginn8.9798.822.157 [.019, .294]
Ouro-1.4B20.85119.2371.614 [.295, 3.031]
Ouro-2.6B12.87612.416.460 [−.278, 1.253]

This predictor uses the hidden coordinates linearly, whereas CATs combines them in an MLP. We therefore also fitted the same two 32-unit networks to each input set, with weight decay and training length chosen by four-fold validation within the fitting prompts:

ModelNeural, ordinary historyNeural, T2G inputsHistory − T2G
Huginn9.4078.863.544 [.368, .714]
Ouro-1.4B20.76719.2641.503 [.145, 2.886]
Ouro-2.6B13.99112.2241.767 [.818, 2.803]

The T2G inputs give lower EAURC in five of the six comparisons with intervals above zero (1.7% and 7.7% under regression; 5.8%, 7.2% and 12.6% under the networks, which have 1,026 parameters against 3,522). Under regression on Ouro-2.6B the two are comparable (the interval includes zero). These controls test the representation at the same target, window and comparable capacity; they do not reproduce the complete CATs system with its trained heads.

AuthorsOpenReview ↗

Response to Reviewer PS7B

W2. Is the full Task-T2G combination needed?

Task-T2G is one calibrated score; the paper does not claim that each of its inputs is necessary, and Appendix F now gives the paired intervals. Control minus Task-T2G (AURC ×1,000; positive favors Task-T2G):

ControlOuro-1.4BOuro-2.6B
Inverse margin2.00 [.55, 3.43]−.77 [−1.89, .36]
Linear confidence1.60 [.37, 2.83]−.62 [−1.65, .46]
Confidence + gate.41 [−.21, 1.08]−.19 [−1.20, .87]
Without hidden motion.37 [−.14, .91]−.28 [−.98, .40]
Without gate.54 [−.12, 1.20]−.42 [−.85, −.04]

On Ouro-1.4B, Task-T2G ranks better than inverse margin and linear confidence (intervals above zero) and is comparable to the three closer variants. On Ouro-2.6B (AURC 10.90) the controls have slightly lower point AURC; four of the five intervals include zero, and the variant without the gate ranks better by .42 [.04, .85].

W2. Inverse margin, paired intervals and latency

Task-T2G minus the comparator at the calibrated operating points, in percentage points (Appendix F, Table 20, new):

ComparatorModelExit rateAnswer changeTime saved
Inverse marginOuro-1.4B4.98 [3.52, 6.54].39 [−.10, .98]−1.21 [−2.32, .01]
Inverse marginOuro-2.6B3.32 [1.76, 4.88].49 [−.10, 1.07].20 [−.67, 1.22]
Confidence + gateOuro-1.4B−.29 [−1.46, .78]−.10 [−.59, .39]−1.30 [−2.14, −.49]
Confidence + gateOuro-2.6B−.59 [−1.95, .78]−.29 [−.78, .20].35 [−.40, 1.17]

Task-T2G exits on more questions than inverse margin at both scales (intervals above zero), with comparable answer change. Its scoring cost offsets the extra exits, so the time saved is comparable to inverse margin at both scales (both intervals include zero). On Ouro-1.4B the point estimates of answer change and time favor inverse margin, accuracy differs by one question in 1,024, and confidence + gate is faster. The MMLU study supports a different claim from latency: the forecast can be fitted to a task decision and calibrated (about 98% of pass-4 answers preserved at 75–81% exits). The gains we claim for trajectory inputs are in forecasting later change (Tables 1–4) and in iterations saved under a KL limit against learned confidence (Table 5); on MMLU the revision states that confidence and the gate remain competitive.

Exit and change use 1,024 questions; timing uses a 128-question subset at batch 1 with scoring included.

Q5. Fixed pass 3

Now in Table 6, on the same 1,024 questions (%):

ModelPolicyAnswer changeExitAccuracy
Ouro-1.4BFixed pass 311.7210066.21
Task-T2G1.8674.6166.11
Fixed pass 40066.50
Ouro-2.6BFixed pass 37.5210074.41
Task-T2G1.6680.7674.80
Fixed pass 40075.00

Fixed pass 3 changes 7.52–11.72% of pass-4 answers, four to six times the 2% target that Task-T2G meets. It exits on every question, so it saves more time: 23.45% and 24.07%, against 12.58% and 18.75% for Task-T2G in the same timing run (Appendix F; the paired run behind Table 6 did not include fixed pass 3). Its accuracy is similar because the changed answers gain and lose correctness about equally (a net loss of 3 and 6 questions). The adaptive policies are for the case where agreement with the full model's answer is the requirement.

Thank you for the clarifications, supplementary material, and additional experiments. The history baseline and fixed-pass-3 comparison address the missing experiments in my review. The following concerns remain.

Q1.

The supplementary package contains the stated categories of materials. However, the paired-timing script requires per-question latency arrays and a provenance file that are absent from the package. The table checks verify agreement with saved summaries, but these missing inputs prevent independent recomputation of the reported runtime intervals. Please provide the missing inputs or complete instructions for generating them and rerunning the analysis.

Q2/Q3.

The response clarifies the methodological differences. Based on this comparison, I find Task-T2G particularly close to CATs and CALM in its prediction objective and stopping procedure (Schuster et al., 2021; Schuster et al., 2022). Its trajectory inputs and recurrent setting distinguish the implementation, but a direct experimental comparison is needed to establish their contribution to answer-consistency prediction. The CATs-inspired history baseline and CALM-inspired state probe are evaluated on terminal-KL prediction and do not address this comparison. Please compare Task-T2G with a CATs-inspired history consistency classifier and a CALM-style current-state classifier, using the same pass-3/pass-4 answer-disagreement target and data splits.

Q4.

The neural comparisons support the contribution of the trajectory representation across the three evaluated models, with relative EAURC reductions of 5.8–12.6%. These results evaluate the input representations under the same neural architecture and should be distinguished from the improvements obtained by the original regression-based Risk-T2G predictor. For that predictor, the relative reductions are approximately 1.7% on Huginn and 7.7% on Ouro-1.4B, while the Ouro-2.6B difference has an interval containing zero. The additional experiments address the missing baseline, but the gains vary substantially across models, leaving it unclear whether regression-based Risk-T2G offers a consistent advantage across models.

W2.

Task-T2G has not established a clear overall advantage over simpler adaptive policies. Its local improvements in risk ranking and exit coverage provide limited evidence of a better practical stopping policy once answer consistency and execution cost are considered. This limits the contribution of the task-specific stopping extension and reduces the breadth of the benefits demonstrated by the paper.

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

T2G predicts terminal KL error from fourteen features of output and hidden-state motion. Risk-T2G combines log-error and mean-error heads. Experiments on Huginn-0125, two Ouro models, and ordinary transformers show better error ranking than matched confidence controls. The paper also evaluates frozen forecasts on C4 and LAMBADA, KL allocation, and Task-T2G for MMLU stopping.

Critical strengths

Risk-T2G improves EAURC over Risk confidence by 8.6%, 22.3%, and 7.7% on the three looped models, with matched objectives and parameter counts (Table 1). Frozen forecasts retain gains on C4 and LAMBADA. T2G + state also improves on either predictor alone in all eight ordinary-transformer model-lens settings (Table 4).

Critical weaknesses

Mean-only has lower EAURC than Risk-T2G on both Ouro models, although Risk-T2G has higher global rank correlation (Table 1). In the MMLU experiment, Task-T2G and confidence + gate give similar exits and time savings at their separately calibrated operating points (Table 14). The one-update, magnitude-history, and full variants use 20, 44, and 119 terms per head (Appendix C.2), so some gains may come from a larger readout. The main Risk confidence comparison does match capacity.

Questions for authors

Do history and direction still help when readout capacity is matched? Can you compare stopping performance against Mean confidence on the fresh 512-prompt cohort?

4 comments
AuthorsOpenReview ↗

Response to Reviewer RyNy

Thank you for the precise questions. Both requested comparisons are in the revised PDF, to which tables and sections refer. EAURC is ×1,000 and lower is better. Brackets are paired 95% intervals from 2,000 resamples of whole prompts or questions.

W1. Mean-only has lower EAURC than Risk-T2G on both Ouro models

Mean-only and Risk-T2G use the same 14 T2G inputs and differ only in the training target, so this comparison is between two objectives for the same inputs (Section 5.3, Table 8). The comparison that tests the inputs holds the objective fixed (119 terms per head in every row):

ObjectiveInputsHuginnOuro-1.4BOuro-2.6B
Mean errorConfidence9.56220.97311.665
Mean errorT2G8.85815.89511.234
Confidence − T2G.705 [.564, .876]5.078 [3.913, 6.757].431 [−1.396, 1.694]
Log + mean errorConfidence9.65224.77313.451
Log + mean errorT2G (Risk-T2G)8.82219.23712.416
Confidence − T2G.830 [.700, .989]5.536 [3.958, 7.729]1.036 [.213, 1.856]

Under either objective the T2G inputs give lower EAURC than confidence in all three models (7.4/24.2/3.7% and 8.6/22.3/7.7%). Five of the six intervals are above zero; the exception is the mean objective on Ouro-2.6B. Between objectives, mean-only has the lower EAURC on the Ouro models (by 17.4% and 9.5%) and Risk-T2G the higher rank correlation in all three models (.9500/.8083/.8357 against .9475/.7549/.7928). For stopping under an average-KL limit the revision recommends mean-only, on the evidence of the comparison you asked for in Q2.

Q2. Stopping against Mean confidence on the 512 held-out prompts

Table 5 (new). All predictors are frozen. The first column is the percentage of iterations avoided by Mean confidence; the others are additional percentage points.

Model / KL limitMean confidenceRisk-T2GMean-only T2GT2G + state
Huginn / .0255.17.46 [.15, .64].63 [.29, .84].45 [.21, .72]
Huginn / .0564.07.26 [.11, .49].41 [.27, .68].62 [.44, .85]
Ouro-1.4B / .0223.38.26 [−.64, 1.13]1.59 [1.05, 2.37]1.33 [.46, 2.31]
Ouro-1.4B / .0537.481.09 [.53, 1.57]1.46 [1.08, 1.91]1.98 [1.49, 2.57]
Ouro-2.6B / .0232.91−.56 [−1.88, .36].56 [−.21, 1.16].18 [−1.09, 1.05]
Ouro-2.6B / .0547.25.23 [−.33, .75].89 [.33, 1.47].76 [.19, 1.21]

Mean-only T2G and T2G + state save more iterations than Mean confidence in five of six settings with intervals above zero; Risk-T2G does so in three. Each score uses its best threshold under the same observed KL limit, selected again in every resample, so the table measures attainable savings; it excludes the cost of computing the scores.

With thresholds fixed in advance on calibration prompts (Table 16), mean-only T2G on Ouro-1.4B at .05 saves .81 [.62, 1.01] points more iterations than Mean confidence with .00169 [.00049, .00289] lower KL. At the other settings the two scores reach different realized KL, so their savings are not directly comparable.

AuthorsOpenReview ↗

Response to Reviewer RyNy

W3 and Q1. Do history and direction help at matched readout size?

A larger one-update readout (Table 2a, new). We expanded the five one-update inputs to degree four, which gives 125 terms per head against T2G's 119. Prompts, targets, penalties and the two-head rule are unchanged, and no feature is discarded.

ModelOne update, 125 termsT2G, 119 termsOne update − T2G
Huginn9.2948.822.472 [.366, .594]
Ouro-1.4B20.47919.2371.242 [.580, 1.939]
Ouro-2.6B12.11912.416−.297 [−1.157, .555]

On Huginn and Ouro-1.4B the gain persists against a one-update predictor with more coefficients: T2G is 5.1% and 6.1% lower, and each head improves separately. On Ouro-2.6B the two are comparable (the interval includes zero), as in Appendix C.2.

The same neural predictor on ordinary history (Table 2b, new). Two 32-unit networks with the same losses and combination rule are fitted either to ordinary three-reading history (confidence statistics, both step-to-step KL values, compressed hidden states) or to the 14 T2G inputs. Weight decay and training length are selected by four-fold validation within the 128 fitting prompts. EAURC is 9.407/20.767/13.991 on ordinary history and 8.863/19.264/12.224 on T2G inputs, with gaps of .544 [.368, .714], 1.503 [.145, 2.886] and 1.767 [.818, 2.803]. The T2G networks have 1,026 parameters against 3,522.

Direction separately (Table 12). To give history and direction identical heads, we reduced each feature set to 20 principal coordinates. Adding direction lowers EAURC on Huginn (9.284 to 8.961) and not on the Ouro models; with 10 coordinates the full feature set improves no model. This check equalizes head size by discarding feature information.

In summary, the two-update representation as a whole helps against a larger readout on Huginn and Ouro-1.4B, and on all three models under the neural predictor. The separate contribution of direction is clear on Huginn and model dependent elsewhere.

W2. Task-T2G and confidence + gate on MMLU

The MMLU study tests whether the forecast can be fitted to a task decision, and it can. Trained on answer changes and calibrated to a 2% target, Task-T2G exits at pass 3 on 74.61% and 80.76% of 1,024 test questions and changes 1.86% and 1.66% of pass-4 answers. It exits on 4.98 [3.52, 6.54] and 3.32 [1.76, 4.88] points more questions than inverse margin. Against confidence + gate, exit rate and answer change are comparable at the calibrated operating points. Task-T2G minus confidence + gate, in percentage points (Appendix F, Table 20):

ModelExit rateAnswer changeTime saved
Ouro-1.4B−.29 [−1.46, .78]−.10 [−.59, .39]−1.30 [−2.14, −.49]
Ouro-2.6B−.59 [−1.95, .78]−.29 [−.78, .20].35 [−.40, 1.17]

The time difference at 1.4B, where confidence + gate is faster, reflects its lower scoring cost. The claim for Task-T2G on MMLU is the calibrated task decision, not latency or accuracy; the revised discussion states that confidence and the gate remain competitive on this task, since KL can change without changing the leading answer.

AuthorsOpenReview ↗

Response to Reviewer RyNy

Thank you again for your review. Since our earlier reply, we have added two adapted consistency classifiers as MMLU baselines (Appendix F, Tables 21 and 22). At the same calibration target, Task-T2G ranked answer changes better and exited more often than both classifiers on both Ouro models. The result also held on 8,894 additional test questions per model.

We would be glad to clarify any remaining points before the discussion closes.

Reviewer_RyNyOpenReview ↗

Response to the Authors' Rebuttal

Thank you for your rebuttal and for addressing my concerns. After considering your responses, I have decided to maintain my original score.

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

The paper focuses on the recursive prediction setup where models repeatedly refine their predictions. It explores whether one can predict how much the model's final output differs from current intermediate predictions. The paper proposes training a regression model to predict the KL divergence between the final prediction and the intermediate states. Complicated feature engineering is adopted and elaborated. Experiments show that the proposed methods outperform confidence-based baselines in performance prediction.

Critical strengths
  • The proposed method is simple but effective.
  • Many baselines are compared in the experiments.
Critical weaknesses
  • The paper lacks clarity. Although the method is simple and the problem studied is not complex, the writing makes both seem more complicated than they are. I also don't see much value in elaborating on how the feature engineering is conducted with so many equations.
  • The contribution is somewhat incremental. It's a regression problem, so a regressor is trained, and it works.
Questions for authors
  • Can the proposed methods be applied to downstream tasks instead of perplexity tasks? Does the predictor still transfer across tasks?
1 comment
AuthorsOpenReview ↗

Response to Reviewer zufy

Thank you for the review. Tables and sections refer to the revised PDF.

W1. Clarity

The procedure is simple, and Section 3 now opens with it in plain terms: keep three successive predictions and hidden states, summarize the two updates between them by 14 numbers, and fit a small regression that predicts how far the current output is from a later one. The model's own later output supplies the training target. A table lists the 14 inputs: 2 for current confidence, 3 each for the size of the latest and of the previous change, and 3 each for direction in output space and in hidden space. An example shows why history matters: the probability sequences (.50, .54, .56) and (.52, .54, .56) have the same confidence and the same last step, but different earlier steps. The section keeps only the definitions needed to compute the inputs; the hidden-space formulas and the derivation are in Appendix B.

W2. "A regressor is trained, and it works"

The regressor is standard and is not our contribution. The contribution is what it is given, and the evidence is that the same regressor on the obvious alternatives is worse. EAURC ×1,000 (error-ranking loss; lower is better) on 256 held-out prompts for each of the three looped models (Table 1):

PredictorHuginnOuro-1.4BOuro-2.6B
Best published stopping rule10.79331.83515.122
Same regression on current confidence9.65224.77313.451
Same regression on the current hidden state12.87322.35016.028
T2G (14 inputs from two updates)8.82219.23712.416

T2G is 8.6%, 22.3% and 7.7% below the confidence regression, which has the same training losses and number of terms, and 18.3%, 39.6% and 17.9% below the best published rule. Every paired 95% interval is above zero.

Two new controls give the regressor ordinary history from the same three readings (entropy, maximum probability and margin at each reading, both step-to-step KL values, compressed hidden states) in place of T2G's 14 inputs (Table 2). With the same regression size, T2G is 1.7% and 7.7% lower on Huginn and Ouro-1.4B; on Ouro-2.6B the interval includes zero. With the same small neural predictor on both input sets, T2G inputs are 5.8%, 7.2% and 12.6% lower in the three models, with 1,026 parameters against 3,522. So "train a regressor" gives different results depending on its inputs, and a compact description of the last two updates is the better input.

Q. Downstream tasks

Yes. On five-shot MMLU, Task-T2G predicts whether Ouro's answer at pass 3 will change at pass 4, and exits at pass 3 when the predicted risk is low. It is fitted on 1,024 questions, its threshold is set on another 507 for at most 2% changed answers, and it is tested on 1,024 new questions (Table 6). Fitting uses the model's own answer changes, not correctness labels.

ModelExits at pass 3Answers changed from pass 4Accuracy: Task-T2G / pass 4Time saved
Ouro-1.4B74.61%1.86%66.11% / 66.50%13.02%
Ouro-2.6B80.76%1.66%74.80% / 75.00%18.85%

Time is measured at batch 1 against running all four passes and includes the cost of the predictor. The result shows that the approach carries over to a task decision: 75–81% of the questions exit one pass early, and about 98% of the full model's answers are kept.

Q. Does the predictor transfer?

Across corpora, unchanged. The predictors fitted on WikiText-2 are applied without refitting to C4 and LAMBADA. Their EAURC is 6.6–23.5% below the matching confidence regression, with paired intervals above zero in five of six comparisons (Table 3).

Across models, refitted. The same 14 inputs and fitting procedure, refitted for each model, lower EAURC by 10.8–21.1% against matched confidence along the layers of four ordinary transformers from 410M to 7B (Table 4).

Across tasks. A change of task changes the prediction target (KL to a later output for text, answer change for MMLU), so each task has its own fit by design; the construction of the trajectory inputs carries over.

On "perplexity tasks": the main experiments measure the KL between the model's current and later output distributions, not perplexity against reference text, so they need no labels.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of the revision

We thank the three reviewers. Two questions recur in the reviews: what the trajectory inputs add beyond a regressor of similar size on ordinary inputs, and what the forecasts are worth for stopping. The revised PDF answers both with new controls. Table and section numbers in all replies refer to the revised PDF.

What changed

  1. Presentation (zufy). Section 3 now states the procedure in three sentences, lists the 14 inputs in a table and gives a two-sequence example before any formula. The hidden-space formulas and the derivation are in Appendix B.
  2. What the inputs add (RyNy, PS7B, zufy). New Table 2 compares Risk-T2G with a larger one-update predictor (125 terms per head against 119), with a generic history predictor on the same three readings (same target, 119 terms), and with ordinary history under the same small neural predictor. T2G has lower EAURC than neural ordinary history on all three models, and than both regression controls on Huginn and Ouro-1.4B (paired 95% intervals above zero).
  3. Stopping against learned confidence (RyNy). New Table 5 compares iterations saved on 512 further held-out prompts against Mean confidence, which has the same objective and size. Mean-only T2G saves more in five of six settings.
  4. MMLU (PS7B, RyNy). Table 6 adds fixed pass 3. New Table 20 gives paired runtime differences from one run, and Appendix F gives paired intervals for exits, answer changes and AURC. Calibrated to a 2% answer-change target, Task-T2G exits one pass early on 75–81% of questions and keeps about 98% of the full model's answers, which shows that the forecast target adapts to a task decision. The replies to PS7B and RyNy give the comparison with inverse margin and confidence + gate.
  5. Related work (PS7B). Section 2 discusses CATs and CALM. A CATs-style history predictor (Table 2) and a current-state probe in the manner of CALM (Table 1, Appendix C.4) are tested as controls.
  6. Artifact (PS7B). An anonymous supplementary ZIP accompanies the revision, with code, prompt hashes, saved fits and thresholds, results and checks of Tables 1, 2, 5, 6, 12, 16, 19 and 20.

Table numbers. Table 14 of the submission (MMLU policies) is now Table 6, and Table 15 (MMLU controls) is Table 19. Tables 2, 5 and 20 are new.

Statistics. Brackets are 95% intervals from 2,000 resamples of whole prompts or questions, with the fitted predictors fixed. The added controls reuse the evaluation prompts of the submission and did not change Risk-T2G.

Dear reviewers, please review the authors’ rebuttal and provide your responses as soon as possible. Your timely feedback is important for moving the discussion forward. Thank you!

AuthorsOpenReview ↗

Response to Reviewer zufy

Thank you for reading the revision, and for your positive view of the experiments and of the method's usefulness. We reply briefly to the one remaining point: that established knowledge already suggests the approach would work.

The method is simple on purpose. It uses 14 numbers and a small regression, and it does not change the model. The contribution is the finding behind it: how a prediction moved over its last two updates tells us how much it will still change. Confidence and the size of the last step leave much of this unexplained, and learned probes with much richer inputs do worse. Two results in the paper show it.

1. Learned probes with much richer inputs do worse. We fitted a probe to the same target on the current hidden state, which is what early-exit classifiers read. Its EAURC (lower is better) is 16–46% higher than T2G's in all three looped models, and every interval is above zero (Table 1, Appendix C.4). On MMLU, the CATs-inspired classifier reads the hidden states and confidences of all three passes. At the same answer-change target, it exits on 4.59 to 8.89 points fewer questions than Task-T2G with its 15 scalars (Table 21, added in the revision). The gain comes from knowing what to measure, and that is what the paper identifies.

2. A small last step does not mean the prediction is finished. Take the quarter of predictions whose last step is smallest. For these, adding T2G's output features to confidence and the last step lowers EAURC from .737 to .389 on Huginn and from 26.651 to 22.962 on Ouro-1.4B (Section 5.1, Table 9). Even among predictions with nearly the same last-step size, the fifth with the highest forecast ends four to six times farther from the final prediction than the fifth with the lowest forecast (Figure 2). Many predictions that look stable are not finished, and the forecast tells them apart.

The finding is not specific to one model. The same 14 inputs give lower EAURC than the same regression on confidence in three looped models, along the layers of four ordinary transformers from 410M to 7B, and on two other corpora without refitting (Tables 1, 3 and 4).

We therefore see the paper as establishing a property of how predictions evolve across iterations and layers: recent motion predicts how much change remains. The method is kept simple so that this property can be measured and used for stopping. We hope this helps in weighing the contribution, and we thank you again for the constructive review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.