acceptodds
Under review as a conference paper at ICLR 2027

When Do State-Based Diagnostics Transfer in Soft-Prompt Tuning?

Abstract

Understanding how adaptation changes language-model behavior is important for reliable evaluation and deployment. In soft-prompt tuning, we ask whether an internal diagnostic's advantage over answer loss transfers across prompt groups and output rules, and whether its ranking gains justify measurement costs. We propose a control-response evaluation framework to address these questions in frozen language models. First, in two Ministral–ECQA prompt pairs, removing continuation reverses the ranking of state and loss diagnostics under strict parsing, while preserving diagnostic scores and initially selected answers. This shows that detecting parseability changes can drive an apparent diagnostic advantage. Second, crossed prompt–question comparisons in Ministral–ECQA show larger changes in relative diagnostic performance across prompt groups than across question sets. The group contrast includes pair selection. Predictions fixed before confirmation reproduce this shift on fresh questions for the same prompts and improve on predicting no group difference. Pair-specific gains over a common-mean forecast remain uncertain, as do gains on unseen GSM8K prompts. Third, at the measured ECQA reference budgets, ranking gains do not offset checks lost to scoring, and direct checks outperform screening with the three original diagnostics on average. The framework provides a reproducible procedure for interpreting diagnostic advantages, testing their transfer, and deciding whether internal measurement is worth its evaluation cost.

Primary area:
interpretability and explainable AI
Keywords:
large language model evaluation, soft-prompt tuning, state-based diagnostics, diagnostic transfer, output-format sensitivity, cost-aware evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gurBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether the advantage of hidden-state-based diagnostic scores over answer loss transfers in soft-prompt tuning, where the task is to predict which questions change correctness status between two prompts.

It reports three main findings. Firstly, in two Ministral–ECQA prompt pairs, the state scores' advantage comes from whether outputs pass strict parsing, and the ranking reverses once outputs are constrained. Secondly, a crossed design shows that this difference varies mainly across prompt groups and little across question sets. Lastly, after charging measurement costs, direct checks are more efficient than screening with the diagnostic scores. The experiments are rigorous and honestly reported. However, in independent validation the proposed diagnostic underperforms simple baselines such as answer loss, JS divergence and a linear probe, and the transfer and cost conclusions are also negative.

The paper has no positive result that supports the value of the framework, so I lean toward a weak reject, rating 2.

Critical strengths

This paper workflow and presentation is relatively complete and professional. 1.The experimental procedure is rigorous. Predictions are locked before the confirmation data are generated, the evaluation uses fresh questions and independently trained prompts, multiple comparisons are Bonferroni-corrected, and undefined AUCs are never filled in. 2. Negative results are reported honestly. There is few overclaim in paper. 3. The constrained-output intervention in section 3 is cleanly designed. The diagnostic scores and the first answer letter are unchanged and only the continuation is removed, yet the state-minus-loss gap changes by −0.672 and −0.845. This shows that a diagnostic's apparent advantage can come from the output format.

Critical weaknesses

There are several weakness. From my perspective, I hope you can polish this work carefully.

Firstly, the proposed diagnostic underperforms simple baselines. In table 1, the state AUCs are 0.460, 0.549 and 0.622, below chance on Ministral. Answer loss reaches 0.834, 0.731 and 0.895. JS divergence reaches 0.850, 0.761 and 0.885. An ordinary linear probe also exceeds CF-PED on all three models.

Secondly, the control-response components of the method contribute thin. The norm-only control and the random-subspace control perform about the same as the full state/readout score, like parseability-channel AUCs of 0.785 and 0.791 versus 0.787 on the discovery set. The selected rank is 32, the upper limit of the candidate set, and the held-out residual energy is still 0.55–0.90. The gradient and subspace features therefore show no measurable effect, and the need for a "control-response framework" isn't supported.

Finally, the transfer and practical-value conclusions are all negative. The improvement of pair-specific forecasts over the uniform mean has intervals that include zero. On GSM8K, the full forecast's MAE (0.097) is worse than predicting zero (0.087). In the cost analysis, the state scores never beat direct inspection in any of the 54 pairs. The gain of the prompt update in Section 6.1 over a random objective also has an interval that includes zero. The paper has no positive result that supports the value of your proposed framework.

I hope authors can give a detailed explanation. If you clarify these questions, I will update my score.

Questions for authors

Please see the weakness section. Thank you for your efforts.

1 comment

Thank you for your careful comparison of CF-PED with simpler baselines and for asking which improvements matter in practice.

The submitted CF-PED and its baselines. On independent ECQA pairs, the submitted diagnostic scored 0.460/0.549/0.622 AUC on Ministral/Granite/Qwen. Loss reached 0.834/0.731/0.895 and JS divergence 0.850/0.761/0.885. The linear probe also beat CF-PED on all three (Table 1). These results make loss and JS the first benchmarks for a state score intended to find questions whose credit will change.

What can be improved? Swapping prompts leaves the change label identical. The original features included signed terms that depend on endpoint order. We compared ordered and swap-invariant features with local and integrated gradients, selected on development data, and froze the rule. On 32 endpoint-disjoint pairs per cell, the table gives the equal-pair mean AUC gain over a rank-32, local ordered refit of the CF-PED structure.

Task / modelΔ\Delta AUC [adjusted interval]
ARC–Ministral 8B+0.101 [0.049, 0.164]
ARC–Granite 8B+0.113 [0.003, 0.227]
ARC–Qwen 9B+0.139 [0.031, 0.261]
ECQA–Ministral 8B†+0.112 [0.049, 0.172]
BoolQ–Ministral 8B†+0.038 [−0.007, 0.075]

† Supplementary confirmations (Table 4). Intervals use 20,000 shared, crossed question/pair bootstrap draws and correction over all 96 comparisons. The symmetric combination improved ranking over the refit in all three ARC families and on ECQA. BoolQ's interval includes zero. On ARC–Ministral it also beat exact margin and the linear probe after correction (Appendix R, Table 37). Loss and JS still had higher mean AUC in all five cells.

Which design choice mattered? At fixed selected rank, changing local ordered to local symmetric features raised mean AUC from .805 to .902, .590 to .701, .726 to .848, .661 to .766 and .888 to .926, in table order. Switching the symmetric features from local to integrated gradients changed mean AUC by at most .004. These component comparisons are descriptive. The corrected gains above apply to the whole rule. In the original discovery set, parseability AUC was .787 for CF-PED, .785 for the norm alone and .791 for a random subspace. The original rank reached its ceiling of 32, with residual fractions .5545 (Granite–ECQA), .8264 (Ministral–ECQA) and .8998 (Ministral–GSM8K). All five new cells reached the expanded ceiling of 256, with residuals .285–.543. On BoolQ the rank-matched random basis scored .927 versus .926 for the selected rule. For a developer, the local, swap-invariant score is therefore a useful baseline for testing what integration or fitting a basis adds at the same rank and data budget.

What transfers? Forecasts locked before the third panel had lower point MAE than the common mean in all five cells. A known pair's previous state-minus-loss AUC gap can thus inform its forecast on fresh questions. Historical improvement intervals over the common mean span zero. For unseen GSM8K prompts, the full forecast had MAE .097 versus .087 for predicting a zero gap. Gains from the prompt-update objective over random were 1.13 [−5.75, 8.00] and 1.31 [−1.25, 4.88] percentage points. Both gains and losses are compatible with these intervals. Choosing a training objective needs its own test.

When does ranking pay for itself? The submission replayed costs on 54 pairs. We now timed four full strategies on two prespecified pairs per cell using panel 0. Each entry gives the changes found by direct checking on the first/second pair at three calibrated budgets (Appendix S).

Task / model10%25%50%
ARC–Ministral2/22/36/5
ARC–Granite3/69/1920/41
ARC–Qwen2/312/920/11
ECQA–Ministral2/510/816/12
BoolQ–Ministral0/12/58/10

At 10% and 25%, screening exhausted the budget before completing any check under both prompts. At 50%, ARC–Granite loss found 0/2 changes. ARC–Qwen loss, JS and the selected score found 20/11, 15/9 and 11/6. Direct checking matched or beat every screen in all 30 observed pair–budget conditions. On ARC–Qwen, even the strongest screen tied direct checking on both pairs at 50%.

The historical Ministral–ECQA replay explains the trade-off. At the reference budget, loss screening's higher hit rate added 20.51 discoveries per pair on average, while fewer completed checks cost 43.56. An evaluator can measure both terms under their own time budget to see whether screening pays (Eq. 7).

A separate supplementary GSM8K preflight timed one historical pair on six calibration questions. Direct checking took 31.83 s and loss screening 33.15 s. Both checked all six without a per-policy deadline and each found one correctness change. Outputs were 23–111 tokens under a 512-token cap. These are full-pass timings. The corresponding long-answer decision requires comparing changes found in equal time.

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

The paper compares hidden-state diagnostics, including the proposed method CF-PED, with the change in answer loss as predictors of which questions change outcome when a frozen language model switches between two trained soft prompts. Experiments cover ECQA on three model families and GMS8K with MATH on one. The authors find that the apparent advantage of the state diagnostic comes from changes in answer parseability, that its performance relative to loss depends more on the choice of prompt pairs than on the questions, and that direct inspection of outputs is more cost-effective than screening with any diagnostic.

Critical strengths

The decomposition of risk changes into parseability and correctness channels and the constrained-output intervention are useful for anyone who evaluates probes against parsed model outputs. They show that changing only the output rule reverses the diagnostic ranking, while the scores and the initial answers remain the same. The analysis is statistically correct, and its results are reported transparently.

Critical weaknesses

The main findings are not shown to generalize beyond one benchmark and a small number of prompt pairs. The central claims rest on a single multiple-choice benchmark, ECQA. The GSM8K and MATH extensions each use one model and on both benchmarks the diagnostics perform near change, so neither reproduces the ECQA results. THe key results also depend on very few prompt pairs. The reversal in Section 3 comes from two Ministral pairs that share an endpoint, and the reference group of prompt pairs in Section 4 contains only three fixed pairs.

The proposed diagnostic is weaker than simple baselines. On independent pairs, CF-PED reaches AUC 0.46-0.62, compared with 0.73-0.90 for loss. A linear probe on hidden states and JS divergence on output logits also outperform it. In the authors’ own cost replay, direct inspection finds more behaviour changes than any state diagnostic, and the paper does not identify a setting in which such a diagnostic would justify its cost.

Questions for authors
  1. Can the diagnostic ranking and the prompt-group effect be reproduced on at least one benchmark other than ECQA, using several models including larger ones?
  2. The gradient terms of CF-PED are a left-endpoint linearization of the readout change, while the exact readout change is available from the same logits as the gold-margin baseline and performs better (Table 7). What does the linearized decomposition add, and how large is the linearization error at the observed prompt displacements?
  3. The paper presents its contribution as an evaluation framework, but the supplied artifact replays the statistics from saved scores and labels. Do the authors plan to release the end-to-end pipeline, including prompt training, hidden-state extraction, diagnostic fitting and cost accounting, in a form that others can apply to new models and tasks?
1 comment

Thank you for asking which findings generalize and how others can reproduce them.

Q1. What transfers beyond ECQA? Our 8B–9B confirmations cover ARC-Challenge on three model families and BoolQ on Ministral. Each cell has 64 independently trained prompts in 32 endpoint-disjoint pairs. We pool two disjoint 300-question panels per pair and weight pair AUCs equally. A third 300-question panel tests a forecast locked before generation. This separates transfer to new prompts from forecasting a known pair on new questions.

Task / modelSelected state scoreLossJS divergenceExact marginLinear probe
ARC / Ministral 8B.904.941.967.789.844
ARC / Granite 8B.704.752.797.654.639
ARC / Qwen 9B.848.969.971.683.865
ECQA / Ministral 8B†.766.845.882.716.741
BoolQ / Ministral 8B†.926.941.951.726.898

† Supplementary confirmation (Table 4).

Loss and JS had higher mean AUC in all five cells. The adjusted loss advantage excludes zero on ARC–Ministral, ARC–Qwen and ECQA–Ministral. With scores fixed, truncating to the native first token turned a state lead into a loss lead in 1, 6, 0, 0 and 0 of 32 pairs, in table order. Truncation also reversed the ranking outside ECQA. These descriptive counts use identical rows where both first tokens were legal (19,138/19,200 on ARC–Granite, all elsewhere). Constrained decoding is a separate intervention.

All five prompt-group contrasts were inconclusive (adjusted intervals include zero). We paired adjacent seeds in one cohort and adjacent selection-set losses in another (16 pairs each). The historical comparison used three Ministral–ECQA pairs screened for loss equivalence by interval. Its two highlighted reversals shared a prompt. The new groups test a different prompt population.

In all 30 timed pair–budget conditions, direct checking found at least as many changes as every screen (two prespecified pairs per cell, Appendix S). Scoring left less time to check these short answers.

Q2. What does linearization add, and how large is its error? As you note, endpoint logits give the exact change in gold log-odds. The local gradient separates signed contributions inside and outside the fitted subspace, showing when they reinforce or cancel. For affine choice logits, write rr for gold log-odds, gsg_s for its source gradient and Δh\Delta h for the state displacement. The log-partition identity gives

Δr−gs⊤Δh=−KL(ps−y∥pt−y),\Delta r-g_s^\top\Delta h=-\mathrm{KL}(p_s^{-y}\Vert p_t^{-y}),

where p−yp^{-y} normalizes the non-gold choices. The tangent overestimates the signed change by exactly this KL, computable from the two endpoints.

On historical ECQA (6,400 rows/model), median/p95 absolute errors in nats were 0.1672/1.0009, 0.2527/1.3413 and 0.0892/0.6418 for Ministral/Granite/Qwen. Signs disagreed on 365/393/244 rows. Historical GSM8K uses full-vocabulary KL averaged over reference positions. Across 3,200 rows, median/p95 was 0.023756/0.046098 nats/token, with 1,068 sign disagreements. Binary BoolQ has one non-gold choice, hence zero analytic remainder (maximum numerical residual 7.54×10−67.54\times10^{-6} nats over 28,800 rows). Its AUCs still differ. We can check the tangent and the numerical integral against the exact readout. Predicting a change in credit requires testing against behavioral labels.

Against a rank-32, local ordered CF-PED refit, the development-selected symmetric combination gained +.101/+.113/+.139/+.112/+.038 AUC. Four intervals exclude zero after correction over 96 comparisons. BoolQ's is [−.007, .075]. The combined features ranked credit changes better than exact margin in four cells (all except ARC–Granite) and better than the probe in one (ARC–Ministral), after correction.

Q3. Can others run and check it? The original attachment replayed 52 intervals from saved scores and labels. The anonymous package includes pinned inputs and documented code for score replay, training, extraction, fitting, freezing, evaluation, timing and reporting. A real Ministral run trained eight prompts (two updates each), saved 60 raw records, timed eight policy executions and ran CPU replay through reporting. Its one-class AUC is undefined.

Per confirmation cell, production replay rebuilt 86,400 feature/target rows and 950,400 scores across three bases and three panels. Independent rank/count code checked 135 point statistics on 144,000 distinct cell/pair/question rows. The 20,000-draw intervals use production replay. We include a raw example, index weights and large tensors, and demonstrate how to check readout position and choice mapping on a fifth model and a new three-choice item on CPU (Appendices R–T).

On Ministral–GSM8K, we selected a swap-invariant integrated-gradient rule on the development set. Independent replay reproduced its mean pair AUC of 0.474 across 16 pairs sharing 100 questions (1,600 rows, Appendix T). This development estimate is below the 0.5 random-ranking reference.

Reviewer_4Go6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

If I understood correctly, the paper investigates whether internal state-based diagnostic methods are better than directly comparing answer losses when evaluating different soft-prompt tuning methods. The paper also tests whether the conclusion holds across different soft-prompting setups.

Critical strengths

None

Critical weaknesses

The clarity is very poor. After reading the abstract, introduction, and the first few sections, I cannot understand what the paper is doing. For example, I got completely lost when reading the content between lines 16 and 27. Consider the sentence "First, in two Ministral–ECQA prompt pairs, removing continuation reverses the ranking of state and loss diagnostics under strict parsing, while preserving diagnostic scores and initially selected answers"; it's unclear to me what "continuation" and "strict parsing" are, or what it means to "preserve diagnostic scores and initially selected answers". The introduction is basically an extended paraphrase of the abstract and provides no extra clarity to me.

After spending roughly as much time as I spend reviewing other papers and still being unable to understand this paper, I decided to stop. The current form offers no positive value to human readers, in my opinion, so I recommend a clear rejection.

Questions for authors

How much AI assistance was involved in writing the paper? This should be disclosed in the "AI USE STATEMENT" section, but it is not currently addressed.

1 comment

Thank you for identifying lines 16–27. We had omitted the step that turns a model output into an evaluation label. The revised abstract, introduction and opening of the method use a recorded example before introducing the equations.

What are we trying to predict? Two prompts can earn the same overall accuracy on different questions. We ask which questions gain or lose credit when one trained soft prompt replaces another on the same frozen model. A soft prompt is a short sequence of learned input vectors. On each question, a prompt earns credit if its output passes the parser and gives the reference answer. The label is 11 when exactly one prompt earns credit. Each diagnostic assigns the question a score. One uses the absolute change in loss on the reference answer, another uses the model's hidden states. AUC measures how well each score ranks questions whose credit changes above those whose credit stays the same.

What changes for one real question? ECQA validation row 814 asks, “A person achieved success and was what by his peers?” The reference is B, “congratulated.” After trimming surrounding whitespace, the strict parser accepts a single uppercase letter A–E. The continuation is everything generated after the first answer token.

Prompt seedRecorded outputEarns credit?First token onlyEarns credit?
2102B. congratulatedNoBYes
2111BYesBYes

For seed 2102 the continuation is . congratulated. Removing it leaves both first answers equal to B, yet the pair label changes from 11 (one prompt earns credit) to 00 (both do). We reuse each diagnostic's original score for this question. The question thus moves between the two classes used to calculate AUC even though its scores stay fixed.

Across 200 historical questions and five prompts, all 1,000 native first tokens were legal choices. The 840 already parseable outputs kept their labels. Of the 160 newly parseable outputs, 124 had chosen correctly and lost credit for extra text alone. The remaining 36 had chosen incorrectly. In two historical pairs sharing one prompt, the state-minus-loss AUC gap shifted by −0.672 [−0.893, −0.457] and −0.845 [−1.069, −0.616]. The state score ranked credit changes better when the whole reply was graded. Loss ranked them better when only the first choice was graded.

What can an evaluator learn? On independent ARC pairs, truncation also reversed the selected state score's lead over loss in 1/32 Ministral pairs and 6/32 Granite pairs. Each of the other three completed model–task combinations had 0/32 such reversals. A benchmark designer can apply both scoring rules to the same outputs with legal first choices. For a one-letter interface, B. congratulated is a genuine format failure. For a benchmark of answer selection, its first choice is correct. Reporting both reveals how much of a diagnostic's apparent advantage survives when credit depends on the choice itself.

Sections 1–2 work through this example. Section 3 and Appendix R give the pair-level interventions. The paper then tests transfer to new prompts and questions and measures how many changes each strategy finds within a time budget.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.