- Rating
- 2
- Confidence
- 3(−1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes SVDoRA, an SVD-based PEFT method that uses outcome supervision to select spectral directions. After a warmup stage, generated calibration answers receive task-specific verification labels and are used to select task-specific generation directions and task-specific verification directions through simulated annealing. A later stage also reweights the training loss using a signal derived from these two direction sets.
The method is evaluated on commonsense reasoning, mathematical reasoning, code generation, and factual QA, and is compared with several SVD-based PEFT baselines. The paper also includes selector ablations and linear probes intended to test whether the selected verification directions contain correctness-related information.
Critical strengths
-
Using outcome supervision to guide spectral direction selection is an interesting extension of existing SVD-based PEFT methods.
-
The paper evaluates the idea across several task families and attempts to validate the proposed verification interpretation with probing experiments.
Critical weaknesses
-
The method builds closely on existing SVD-based task-specific direction approaches such as LoRA-DASH and DisLoRA. Its main novelty is the use of outcome-supervised annealing to jointly select spectral directions with separate generation and verification roles. This is a meaningful extension, but the overall methodological novelty remains relatively incremental.
-
The ablations do not clearly justify the complexity of the full pipeline. On Qwen2.5-1.5B commonsense reasoning, SASS reaches 79.32% compared with 78.88% for the reported correctness-agnostic selector. Moreover, replacing the Phase-3 GVSCR weighting with simpler alternatives changes the eight-task mean by at most 0.3 percentage points. These results suggest that the incremental benefit of some components of the full pipeline is modest.
-
The evidence for a distinct verification role of TSVD remains limited. At the deployed probe dimension, TSVD achieves 0.635 accuracy compared with a 0.624 majority baseline, while its AUROC only ties the strongest null-space control. This supports the presence of correctness-related information but does not clearly establish a verification-specific mechanism. In addition, code-generation calibration uses normalized text similarity rather than execution outcomes, so the outcome supervision is not defined consistently across task families.
Questions for authors
-
How much of the performance gain remains in a simplified variant that retains outcome-guided spectral direction selection but removes the separate TSGD/TSVD role design and uses uniform cross-entropy instead of GVSCR? This would help isolate which parts of the full pipeline are necessary.
-
Can the authors provide stronger evidence that TSVD captures verification-specific information rather than features that are merely correlated with answer correctness? For example, would selectively removing or replacing TSVD directions affect correctness discrimination more than matched control directions?
5 comments
Response to Reviewer sUz3 (Part 2/3): Selector Ablation and Simplified Variant Without Separate TSGD/TSVD Roles (W2 & Q1)
2. Selector Ablation: SASS vs. VUS and Controlled Outcome-Label Ablation (E1) (W2)
In Table A3 (tab:b10-selection), Random selection () is the information floor under the same hard-partitioned parameterization (Prop. 5(iii)), over which SASS-v2 () gains , while VUS (, single seed) is our own second-order Fisher selector (Eq. A4) that already beats random by . Because the VUS and SASS-v1 () rows in Table A3 are single-seed and SASS-v1 also differs in annealing energy form and singular direction shortlist construction, we ran a controlled label ablation (E1) within the identical SASS-v2 pipeline, which this version of SASS is same to the main table in the paper (Qwen2.5-1.5B-Instruct, , hard partition, uniform CE), comparing aligned labels (E1-A), shuffled labels (E1-B), and removed labels (E1-C, ):
Table R1. Controlled correctness-label ablation (E1) within the identical SASS-v2 selection pipeline (Qwen2.5-1.5B-Instruct, , , hard partition, uniform CE).
| Selection Arm | MedQA Acc (%, 2 seeds) | MedQA Cal. | StrategyQA Acc (%) | StrategyQA Cal. | GSM8K Acc (%) | GSM8K Cal. |
|---|---|---|---|---|---|---|
E1-A: Aligned correctness labels () | 48.47$\pm$0.55 (49.02/47.92) | 0.558 | 66.81 (153/229) | 2.211 | 44.35 (585/1319) | 0.641 |
E1-B: Shuffled correctness labels () | 48.74$\pm$0.35 (49.10/48.39) | 0.333 | 66.81 (153/229) | 0.352 | 46.70 (616/1319) | 0.419 |
E1-C: Labels removed () | 47.80$\pm$0.04 (47.76/47.84) | 0.201 | 67.25 (154/229) | 0.579 | 43.37 (572/1319) | 0.142 |
What Table R1 Shows: (1) Comparing E1-A against E1-C (labels removed), outcome-label supervision is a clear source of performance gain: it drives much higher TSVD calibration separation across all three tasks ( vs. on MedQA; vs. on StrategyQA; vs. on GSM8K) and improves test accuracy by on MedQA ( vs. , consistent across both seeds: on s42, on s123) and on GSM8K ( vs. ). (2) Comparing E1-A against E1-B (shuffled labels), aligned labels are essential to achieve high correctness separation ( vs. ; vs. ; vs. ), whereas on downstream test accuracy under uniform CE the difference between aligned and shuffled order is less pronounced ( vs. on MedQA, vs. on StrategyQA, vs. on GSM8K), as keeping the non-zero separation terms () active in the annealing energy already regularizes direction selection relative to dropping them in E1-C.
3. Simplified Variant Without Separate TSGD/TSVD Roles and With Uniform Cross-Entropy (E3) (Q1 & W2)
In Q1, you asked how much gain remains in "a simplified variant that retains outcome-guided spectral direction selection but removes the separate TSGD/TSVD role design and uses uniform cross-entropy instead of GVSCR." On 8-task commonsense, disabling GVSCR (Off, Table A4) achieves ( over the random floor), and setting (Table A5) changes the 8-task mean by (, single seed). On MedQA (Qwen2.5-1.5B-Instruct, seed 42, uniform CE), we trained this exact simplified variant (E3) with so all 8 directions are selected purely by the outcome-verified objective without a separate TSGD role:
Table R3. Simplified variant without separate TSGD/TSVD roles and with uniform CE (E3) vs. two-role anchor (E1-A) on MedQA (Qwen2.5-1.5B-Instruct, , hard partition, seed 42, greedy decoding).
| Variant | Direction Budget | Energy Weights | Phase-3 Loss | Cal. | MedQA Test Acc (%) | vs. Two-Role |
|---|---|---|---|---|---|---|
Two-role SVDoRA with uniform CE (E1-A anchor) | Off () | 0.5326 | 49.02 (624/1273) | — | ||
Simplified variant without separate TSGD/TSVD roles (E3) | Off () | 0.5794 | 48.47 (617/1273) | -0.55,pp (-7 items) |
In this single-seed comparison, E3 reaches () vs. (, a point-estimate difference) for the two-role anchor (E1-A), showing that outcome-guided selection with uniform CE and a single role retains nearly all MedQA accuracy while the two-role split adds .
Response to Reviewer sUz3 (Part 3/3): GVSCR Design Motivation, Hyperparameter Sweep, and Literature Comparison (W2)
(Note: For a quick summary of all experiment codes E1-A/B/C, E2a/b/c, E3, and E7/E8, see Table R0 in General Comment Part 1.)
4. GVSCR Design Motivation, Hyperparameter Sweep (E7), and Literature Comparison (E8) (W2)
We agree that on commonsense (Table A4), GVSCR does not improve accuracy over Off. Its architectural motivation was not to act as a standalone accuracy booster, but to close the loop between Phase-2 selection and Phase-3 optimization: without GVSCR, TSVD directions extracted at are updated only passively under uniform CE, never invoking their correctness-correlated readout during training. GVSCR was designed as a step-to-step training loop that consumes the TSVD readout signal at step to modulate updates so step builds on weights that absorbed prior feedback (without claiming inference-time self-verification). To evaluate Phase-3 weighting on MedQA and zero-shot MMLU-Pro (Qwen2.5-1.5B-Instruct, seed 42), we swept 21 GVSCR configs on MedQA dev (E7) and compared against 4 literature methods (E8: Focal, GCE, SPL [Jiang et al., ACM MM 2014], and JTT, each tuned over 3 dev configs under matched warmup/normalization):
Table R6. Phase-3 loss reweighting comparison on MedQA and zero-shot MMLU-Pro (Qwen2.5-1.5B-Instruct, , ; test columns report mean$\pm$std across 3 nucleus evaluations at , decoding RNG seeds 42/123/7 on a single trained checkpoint).
| Phase-3 Weighting Method | Dev Configs | Selected Hyperparameters (on MedQA dev) | MedQA Dev (Greedy) | MedQA Test (%) | MMLU-Pro Overall (%) | MMLU-Pro Health (%) | MMLU-Pro Biology (%) |
|---|---|---|---|---|---|---|---|
| Uniform CE (Off) | 1 | (E1-A) | 46.54 (592/1272) | 48.94$\pm$0.33 | 26.53$\pm$0.12 | 33.54$\pm$0.15 | 48.77$\pm$0.35 |
| Legacy (TSVD-norm) | 1 | 47.09 (599/1272) | 47.71$\pm$0.21 | 25.16$\pm$0.09 | 29.71$\pm$0.65 | 47.05$\pm$0.29 | |
| Focal Loss | 3 | (from ) | 47.09 (599/1272) | 49.12$\pm$0.13 | 26.22$\pm$0.03 | 29.46$\pm$0.62 | 51.28$\pm$0.47 |
| Generalized CE (GCE) | 3 | (from ) | 47.88 (609/1272) | 46.24$\pm$0.26 | 25.77$\pm$0.05 | 31.30$\pm$0.70 | 47.88$\pm$0.29 |
| Self-Paced Learning (SPL) | 3 | (from ) | 48.66 (619/1272) | 47.58$\pm$0.27 | 27.12$\pm$0.01 | 32.36$\pm$0.15 | 53.09$\pm$0.13 |
| Just Train Twice (JTT) | 3 | (from ) | 46.70 (594/1272) | 46.19$\pm$0.61 | 25.52$\pm$0.11 | 30.52$\pm$0.30 | 48.58$\pm$0.07 |
| GVSCR (Paper Default) | 1 | , mid_last | 47.01 (598/1272) | 48.65$\pm$0.26 | 26.40$\pm$0.11 | 31.70$\pm$0.30 | 50.49$\pm$0.41 |
| GVSCR (Dev-Tuned) | 21 | (gvscr_lambda_warmup_steps=200) | 47.56 (605/1272) | 49.86$\pm$0.32 | 27.22$\pm$0.12 | 32.11$\pm$0.12 | 48.68$\pm$0.39 |
Findings: On MedQA test (E8), GVSCR (Dev-Tuned) (, dev ) reaches , beating Uniform CE (, vs. std sum), GVSCR Default (, ), and all four literature methods (closest is Focal at , vs. std sum; note the 21-vs-3 dev config asymmetry). On MMLU-Pro, it reaches (SPL , Uniform CE ). On 1.5B math (E4), lowered GSM8K$+-1.45\text{ pp}$** ( vs. ).
Response to Reviewer sUz3 (Part 1/3): TSVD Directional Interventions, Answer-Span Probes, and Entropy Fusion (W3 & Q2)
1. Evidence on TSVD: Directional Interventions (
E2a), Answer-Span Probes (E2b), and Entropy Fusion (E2c) (W3 & Q2)We thank Reviewer
sUz3for the constructive evaluation. SVDoRA does not claim an inference-time verifier, nor claims that the LoRA module possesses a ready-made self-verification capability, nor claims that the SVDoRA training paradigm achieves full self-improvement (§1, §3.6, Limitations); However, we do hypothesize that the TSVD directions contain certain signals that can differentiate the model answer from correct or wrong. In paper's Table 6 (N=1319, positive rate 0.624), raw accuracy is compressed by class imbalance (weighted F1 is 0.639 vs. 0.479 majority), though single-token L21 AUROC (0.672) tied NullSpace (0.674). Following Q2, we ran inference-time directional interventions (E2a) on the aligned (E1-A)Qwen2.5-1.5B-Instructcheckpoints (r=16, k_g=k_v=4, uniform CE, seed 42) on MedQA (1,273 test / 1,272 dev) and StrategyQA (229 dev). NoteE2intervenes onE1-A(notE1-B/E1-C), testing reliance on the selected TSVD directions inE1-A. We also aligned the probe (E2b) with Phase-2 calibration via pre-hooks onq/k/v/o_projacross layers {7,14,21,27}, mean-pooling module inputs over the answer span and weighting by active channel gain (StandardScaler+ L2-LogisticRegression, C=0.05, 5-fold CV).Table R2a. Inference-time directional interventions (
E2a) on the unmodifiedE1-Acheckpoint (Qwen2.5-1.5B-Instruct, r=16, kg=kv=4, greedy decoding).Table R2b. Baseline correctness probing (
E2b) and incremental AUROC over output token entropy (E2c) across layers {7,14,21,27}.Key Takeaways (Q2): (i) MedQA: TSVD attains the highest multi-layer AUROC (0.6470, BAcc 0.6118), pure norm AUROC (0.6172, +3.62 pp over NullSpace), and 4-layer mean AUROC (0.5825). Zeroing (βj→0) or resetting (βj→σj) TSVD drops multi-layer AUROC by −0.1455 and −0.0736 (vs. ≤0.0373 for controls). In
E2c, fusing TSVD with token entropy yields positive AUROC increments across all readouts (+2.26 pp QDA, +0.48 pp 2D-logit, +6.85 pp GNB), whereas TSGD, NullSpace, and Random are negative. Zeroing TSVD also causes the largest MedQA test accuracy drop (−2.59 pp, vs. −1.96 pp for TSGD zeroing and −0.97 pp for frozen removal; βj→σj is −0.08 pp). (ii) StrategyQA (229 items): Results are mixed—TSVD leads on 4-layer mean AUROC (0.5206), multi-layer BAcc (0.5443), and βj→0 probe drop (−0.0682), and beats Random/TSGD onE2c2D-logit (+0.47 pp vs. −0.78 pp / −1.08 pp), whereas NullSpace has higher multi-layer AUROC (0.5776) andE2c2D-logit (+1.46 pp), and TSGD zeroing drops accuracy more (−3.49 pp vs. −2.18 pp).