acceptodds
Under review as a conference paper at ICLR 2027

SIGMA: Label-Efficient Training of Language Models to Follow Changing Structured Guidance

Abstract

Language models offer a flexible interface for task-specific decisions, yet accuracy under fixed instructions does not establish reliable behavior when those instructions change. We investigate how limited task annotations can teach models to follow updates that change the required decision while remaining stable under changes that preserve meaning. We introduce SIGMA, which pairs each labeled input with two versions of structured guidance and supervises both correct answers and the expected relationship between their predictions. By aligning the relative scores of all candidate outputs across guidance versions, SIGMA reuses existing annotations to learn how decisions should respond to guidance edits. On BANKING77, CLINC150, and HWU64 with Qwen2.5-7B-Instruct, SIGMA trained on four examples per intent exceeds the mean accuracy of supervised low-rank fine-tuning using sixteen examples per intent under unseen guidance combinations on every dataset, using 75% fewer training annotations with the same development set. The results support supervision across guidance versions as a way to improve annotation efficiency and handle unseen combinations of updates without further training.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Label Efficiency, Parameter-Efficient Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eZurOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SIGMA, a method featuring a new training objective termed "Equivalent Fine-tuning," designed for the label-efficient adaptation of Language Models to dynamically changing structured guidance. The method pairs labeled inputs under original and edited guidance, combining a standard supervised classification loss with a centered alignment loss. This multi-term loss is applied using a shared LoRA adapter while keeping the base pre-trained LLM weights frozen. Based on their experimental results, the authors claim that enforcing consistency and supervision across distinct guidance versions improves annotation efficiency and enables the model to handle unseen combinations of updates at inference time without requiring additional fine-tuning.

Critical strengths
  1. Zero-Shot Adaptation Focus: The method addresses an important failure mode in an LLM’s ability to follow complex instructions, leveraging parameter-efficient fine-tuning (LoRA) to successfully handle unseen combinations of instruction changes without test-time updates.

  2. Theoretical Foundation: The paper provides mathematically well-motivated objectives and theoretical analysis that formalizes the conditions under which consistency regularization can successfully improve learning dynamics.

Critical weaknesses
  1. Insufficient Motivation for the Core Method over its Variant: The paper does not provide convincing justification to prefer the primary SIGMA method over its proposed extension, SIGMA-F. Empirically, SIGMA-F appears less computationally expensive to train while achieving higher composition accuracy. While the authors note that SIGMA-F requires an explicit binding map and a semantic prediction interface, the experimental setups already provide structured bindings and fixed semantic labels. It is unclear what specific deployment constraints make SIGMA preferable, or under what practical scenarios semantic predictions would be unavailable. Additionally, the comparative performance of SIGMA-F across identical 1/2/4/8/16-shot label budgets remains unverified.

  2. Limited Methodological Novelty and Theoretical Scope: The core technical contribution offers incremental novelty, as it builds heavily on existing consistency regularization frameworks. While the proposed centered alignment loss improves performance over a transported KL baseline, the average empirical gains are relatively small. Furthermore, the introduced composition bound only controls the accumulated alignment error along a trajectory; it does not theoretically guarantee that these errors will remain bounded on completely unseen compositions.

  3. Narrow Evaluation and Generalization Scope: The empirical evaluation considers a highly restricted set of guidance modifications. While the authors test generalization on held-out compositions involving binding swaps and rule ordering, evidence for broader instruction reordering or fundamental changes to the guidance itself is missing. The evaluation fails to establish how the proposed approach scales or adapts when descriptions, label inventories, or underlying decision rules are modified.

Questions for authors

Please address the following methodological and empirical clarifications:

  1. SIGMA vs. SIGMA-F Trade-offs: Given that SIGMA-F demonstrates lower training overhead and superior composition accuracy within your current experimental setup, what are the concrete deployment constraints or scenarios that justify choosing the standard SIGMA method? Could you provide a clear performance comparison between SIGMA and SIGMA-F across the 1/2/4/8/16-shot label budgets?

  2. Robustness to Guidance Changes: How does the proposed alignment mechanism perform when subjected to structural alterations beyond binding swaps and rule ordering, such as changing the descriptive phrasing, modifying the label inventories, or introducing entirely new decision rules?

  3. Generalization of the Composition Bound: Can the theoretical analysis or the composition bound be extended to prove that accumulated alignment errors remain small on entirely unseen compositions, rather than just controlling the error trajectory over known spaces?

3 comments
AuthorsOpenReview ↗

Response to Reviewer eZur (1/3)

Thank you for your time and constructive feedback on the method and its evaluation. We address your questions about SIGMA-F, novelty, and generalization below.

W1 / Q1. Why choose SIGMA over SIGMA-F, especially at small label budgets?

Response.{\color{green}\mathbf{Response.}} SIGMA uses local guidance edits to teach both the correct answer and how candidate evidence should change. Paired-SFT and token-matched controls (Tables 2 and 11) measure this learning benefit. SIGMA-F first predicts the semantic class and then applies the binding map exactly. Both methods have the same descriptions and map in these experiments.

The Qwen2.5-7B-Instruct sweep uses matched LoRA capacity and five seeds. The methods share nested training subsets, development sets, and evaluation samples and weights. Entries are Comp. accuracy, mean ± SD (%). SIGMA values come from Table 5.

DatasetMethod1 shot2 shots4 shots8 shots16 shots
BANKING77SIGMA73.81 ± 2.1381.26 ± 1.7585.72 ± 1.1388.01 ± 0.9689.47 ± 0.76
BANKING77SIGMA-F75.80 ± 2.0082.10 ± 1.5586.30 ± 1.1889.10 ± 0.8691.10 ± 0.62
CLINC150SIGMA85.47 ± 1.9690.31 ± 1.7593.42 ± 1.1295.01 ± 0.9895.81 ± 0.74
CLINC150SIGMA-F86.90 ± 1.7590.90 ± 1.3093.80 ± 1.0095.85 ± 0.7096.70 ± 0.48
HWU64SIGMA72.65 ± 1.9780.41 ± 1.6085.28 ± 1.3088.37 ± 1.0389.91 ± 0.79
HWU64SIGMA-F75.10 ± 2.1082.00 ± 1.6586.10 ± 1.2890.05 ± 0.9291.35 ± 0.62

SIGMA-F has higher composition means at every budget. Its Base and Comp. accuracies are identical on matched samples and weights because of exact lookup. The following table reports measured GPU-hours for 4-shot training only.

DatasetSIGMA-4: Comp. (%) / GPU-hSIGMA-F-4: Comp. (%) / GPU-h
BANKING7785.72 / ≈5.4986.30 / ≈3.24
CLINC15093.42 / ≈20.8493.80 / ≈12.31
HWU6485.28 / ≈3.7986.10 / ≈2.24

SIGMA-F offers the better composition–cost trade-off here. SIGMA shows how explicit edit relations improve a language model's direct predictions under changing guidance. The 75% training-label saving refers specifically to SIGMA-4 versus SFT-16.

AuthorsOpenReview ↗

Response to Reviewer eZur (2/3)

W2(a). Is the contribution more than standard consistency regularization with a modest gain over KL?

Response.{\color{green}\mathbf{Response.}} The key insight is that a task annotation can also teach how predictions should respond to guidance. A known edit specifies how the full posterior should change. Even when the gold answer stays the same, the edit supervises relative evidence for other candidates. Equation (4) expresses this relation exactly through centered scores. The loss penalizes discrepancies in all transported pairwise log-odds (Appendix B.1).

Theorem 4.1 explains when this constraint reduces variance in a fixed-feature squared-loss model. The analysis accounts for the shared label noise across views and the bias caused by equivariance mismatch. It therefore connects the alignment objective to learning from reused annotations. Table 13 reports composition gains with paired pointwise 95% intervals:

DatasetSIGMA − paired SFT (pp)SIGMA − transported KL (pp)
BANKING77+3.19 [2.63, 3.76]+1.73 [1.19, 2.25]
CLINC150+1.95 [1.48, 2.43]+1.20 [0.79, 1.64]
HWU64+2.97 [2.38, 3.59]+2.14 [1.57, 2.72]

The gains are larger on examples that require a different answer: 8.80/6.15/9.65 pp over KL (Table 12). This directly measures improved responses to binding changes. The token-matched and opaque-output controls in Tables 10–11 also assess the roles of training compute and label-string semantics.

For comparison with PBM-LoRA [2], PBM-adapted + paired CE retains log-probability dispersion and entropy. It uses aligned full-sequence probabilities, two views, matched LoRA capacity, and at most 15 search trials. This adapts the original unsupervised multi-permutation protocol to our supervised setting. Entries are 10-shot accuracies (%). Changed denotes changed targets.

DatasetKL Comp.PBM-adapted Comp.SIGMA Comp.PBM-adapted ChangedSIGMA Changed
BANKING7786.8387.5588.5677.0083.18
CLINC15094.1794.7095.3786.0090.42
HWU6486.9487.5589.0876.0083.23

The Symbol Tuning-style adaptation [1] uses opaque outputs and global random binding augmentation. Schemas provide the mappings supplied by in-context examples in the original method. Both variants use the same augmentation and paired supervision, making the alignment benefit directly measurable. Entries are Comp. accuracy (%).

DatasetGlobal binding augmentation + paired CESame augmentation + SIGMA alignmentGain (pp)
BANKING7786.8588.10+1.25
CLINC15093.9594.95+1.00
HWU6486.5088.20+1.70

Together, these comparisons show that supervising the response to an edit improves learning over paired supervision, global mapping augmentation, and the PBM-adapted objective. The benefit is especially clear when the guidance requires the model to change its answer.

[1] Symbol tuning improves in-context learning in language models.

[2] Quantifying and Mitigating Selection Bias in LLMs: A Transferable LoRA Fine-Tuning and Efficient Majority Voting Approach.

AuthorsOpenReview ↗

Response to Reviewer eZur (3/3)

W2(b) / Q3. Does the composition bound guarantee small errors on entirely unseen compositions?

Response.{\color{green}\mathbf{Response.}} Equation (7) already applies to unseen valid paths. For εj=∣uj−Pjuj−1∣2\varepsilon_j=\\|\mathbf u_j-P_j\mathbf u_{j-1}\\|_2, permutation orthogonality gives

\leq \sum_{j=1}^{k}\varepsilon_j.$$ With positive initial gold margin $\gamma_0$, Corollary B.4 certifies the transported answer whenever $$R_{\mathrm{mr}}=\frac{\sqrt{2}\sum_{j=1}^{k}\varepsilon_j}{\gamma_0}<1.$$ A uniform extension follows if $\varepsilon_j\leq\delta$ holds along every path under consideration. The total error is then at most $k\delta$, and $\sqrt{2}k\delta<\gamma_0$ certifies the transported answer. This assumption concerns the residuals on the unseen paths themselves. The experiments below measure those residuals directly. We evaluate frozen models on held-out binding paths using all 77/150/64 candidates. Results at the two largest depths report **median path-average local residual / all-path Comp. accuracy (%)**. Depth 16 is reserved for the final stress test. | Dataset | Depth | Paired SFT | Transported KL | SIGMA | | --- | --- | --- | --- | --- | | BANKING77 | 8 | 0.44 / 84.67 | 0.35 / 86.23 | 0.25 / 88.09 | | BANKING77 | 16 | 0.49 / 82.40 | 0.39 / 84.00 | 0.29 / 85.70 | | CLINC150 | 8 | 0.38 / 92.85 | 0.30 / 93.67 | 0.21 / 95.00 | | CLINC150 | 16 | 0.42 / 90.50 | 0.34 / 91.70 | 0.24 / 93.50 | | HWU64 | 8 | 0.50 / 85.55 | 0.41 / 86.44 | 0.30 / 88.68 | | HWU64 | 16 | 0.55 / 82.50 | 0.46 / 84.00 | 0.34 / 86.00 | SIGMA has lower residuals and higher accuracy at both depths on every dataset. This provides direct empirical support for transfer along the tested unseen paths. The certificate remains conditional on each path's residuals and initial margin. Accuracy covers the full candidate inventory and all sampled paths, including initially incorrect predictions. --- > #### **W3 / Q2. What happens under changed descriptions, broader ordering changes, new inventories, or new decision rules?** ${\color{green}\mathbf{Response.}}$ To address your description and structural-edit questions, we test frozen 10-shot models on equivalent paraphrases, randomized child orders, and cross-group binding swaps. Paraphrases are verified independently of test outcomes. SIGMA-F receives the same updated descriptions and map. Each cell reports **Comp. / changed-target accuracy (%)**. | Dataset | Frozen test condition | Paired SFT | Transported KL | SIGMA | SIGMA-F | | --- | --- | --- | --- | --- | --- | | BANKING77 | Paraphrased descriptions + compositions | 83.90 / 67.00 | 85.45 / 72.20 | 87.20 / 80.80 | 89.20 / 88.90 | | BANKING77 | Random serialization + compositions | 84.60 / 68.30 | 85.95 / 73.20 | 87.85 / 82.00 | 89.91 / 89.54 | | BANKING77 | Cross-group binding compositions | 77.80 / 43.50 | 79.50 / 48.00 | 82.60 / 58.50 | 89.91 / 89.30 | | CLINC150 | Paraphrased descriptions + compositions | 91.85 / 78.40 | 92.75 / 82.00 | 94.65 / 88.80 | 95.65 / 95.40 | | CLINC150 | Random serialization + compositions | 92.40 / 79.80 | 93.20 / 83.50 | 94.90 / 89.50 | 96.25 / 96.13 | | CLINC150 | Cross-group binding compositions | 88.50 / 55.00 | 89.50 / 60.00 | 91.50 / 70.00 | 96.25 / 95.80 | | HWU64 | Paraphrased descriptions + compositions | 84.55 / 66.80 | 85.45 / 71.50 | 87.65 / 80.50 | 89.80 / 89.50 | | HWU64 | Random serialization + compositions | 84.80 / 67.80 | 86.10 / 72.90 | 88.15 / 81.80 | 90.49 / 90.60 | | HWU64 | Cross-group binding compositions | 76.80 / 42.00 | 78.60 / 47.50 | 81.90 / 57.00 | 90.49 / 90.40 | SIGMA exceeds both controls in every condition. On cross-group swaps, it achieves 58.50/70.00/57.00% changed-target accuracy, improving over transported KL by 10.5/10.0/9.5 pp. These results establish transfer across descriptions, serializations, and bindings within a fixed semantic task and candidate inventory. Adding or removing candidates changes the space of $P_e$. New decision rules change the semantic posterior in Eq. (2). Both changes call for a corresponding supervision relation. Table 10 addresses label-string semantics separately by evaluating models retrained with opaque output IDs.
Reviewer_nT5oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

SIGMA uses the expected relationship between decisions before and after a guidance change as a training signal. For the same labeled input, it learns when the prescribed output should change and when predictions should be preserved, jointly constraining the relative evidence for all candidate outputs. Each annotation thus supervises both task correctness and guidance use; the adapted model then scores outputs under current guidance without further training. Across the evaluated datasets and backbones, gains on unseen edit combinations indicate that local training interventions can support transferable guidance use, with improved training-label efficiency on the three primary intent benchmarks.

Critical strengths
  1. This work explores an interesting question: Can a small set of task annotations teach a language model to follow changing guidance and generalize this behavior to unseen combinations of updates? The authors propose SIGMA, which trains language models to respond appropriately to structured-guidance edits. Training combines supervision on both views with a penalty that aligns their centered candidate scores. At inference, the adapted model scores outputs under the current guidance with frozen parameters.

  2. SIGMA improves generalization to unseen guidance compositions across three backbones. Controls using identical paired views and token-matched finetuning assess gains beyond additional views or computation.

Critical weaknesses
  1. The practical advantage over the factorized alternative remains unclear. With fixed semantic labels and descriptions and an explicit binding map, SIGMA-F guarantees binding compliance, achieves higher composition accuracy on all three primary datasets, and uses 8.11 versus 13.73 GPU-hours in the reported BANKING77 configuration (Tables 1 and 8). Although SIGMA has higher Base accuracy, the paper does not identify deployment constraints favoring direct output scoring over semantic classification with exact binding lookup. SIGMA-F is also absent from the label-efficiency sweep, leaving SIGMA’s annotation-efficiency advantage over this alternative unestablished.

  2. The evaluated changes cover a narrow family of structured-guidance updates. Training and transfer manipulate output bindings and serialization while preserving the semantic classification task, candidate inventory, hierarchy, and descriptions. Held-out compositions are meaningful within this family, but do not test changed decision boundaries, new rules, cross-group binding edits, or altered descriptions. The theoretical composition result depends on local errors along the evaluated path; it does not establish small errors on new semantic rule changes. This limits how broadly the findings support the title's changing-guidance motivation, even though the paper generally states its mathematical assumptions carefully.

  3. Some figures would benefit from improved visual clarity and readability. In particular, Figure 1 is difficult to read at normal page size, with small text and crowded annotations. Increasing the font size, simplifying the layout, and ensuring sharp rendering would make the comparison between SFT and SIGMA easier to follow.

Questions for authors
  1. What concrete deployment constraint makes direct candidate scoring preferable to semantic classification followed by explicit binding lookup? In the evaluated setup, does SIGMA-F require information or access unavailable to SIGMA?
  2. How does SIGMA-F perform over the same 1/2/4/8/16-shot sweep? Does SIGMA retain any annotation-efficiency advantage when this matched-capacity factorized alternative is included?
  3. Should the main claim be interpreted specifically as robustness to binding permutations and equivalent serialization? If a broader claim about changing decision requirements is intended, what current evidence supports transfer beyond that edit family? I would not treat an extensive new benchmark suite as a prerequisite if the claim is explicitly narrowed.
3 comments
AuthorsOpenReview ↗

Response to Reviewer nT5o (1/2)

We appreciate the time and care you put into your review, especially your comments on SIGMA-F and the scope of the claims. Below, we clarify these points and present the matched comparisons.

W1 / Q1 / Q2. What favors SIGMA over SIGMA-F, and how do they compare across label budgets?

Response.{\color{green}\mathbf{Response.}} SIGMA learns how to respond to guidance edits. Each task annotation supervises the correct answer and how candidate evidence should change. The resulting model handles unseen compositions with frozen parameters. SIGMA-F predicts the semantic class, then applies an exact lookup to obtain the current output (Appendix A.3). Both methods receive the same information here.

Paired SFT makes the learning benefit measurable. With the same inputs, edits, and labels, SIGMA improves composition accuracy by 3.19/1.95/2.97 pp (Table 13). Your comparison with SIGMA-F also shows how this learned behavior compares with exact binding lookup.

The sweep uses Qwen2.5-7B-Instruct with matched LoRA capacity and five seeds. Both methods share nested training subsets, development sets, and evaluation samples and weights. Entries are Comp. accuracy, mean ± SD (%). SIGMA values come from Table 5.

DatasetMethod1 shot2 shots4 shots8 shots16 shots
BANKING77SIGMA73.81 ± 2.1381.26 ± 1.7585.72 ± 1.1388.01 ± 0.9689.47 ± 0.76
BANKING77SIGMA-F75.80 ± 2.0082.10 ± 1.5586.30 ± 1.1889.10 ± 0.8691.10 ± 0.62
CLINC150SIGMA85.47 ± 1.9690.31 ± 1.7593.42 ± 1.1295.01 ± 0.9895.81 ± 0.74
CLINC150SIGMA-F86.90 ± 1.7590.90 ± 1.3093.80 ± 1.0095.85 ± 0.7096.70 ± 0.48
HWU64SIGMA72.65 ± 1.9780.41 ± 1.6085.28 ± 1.3088.37 ± 1.0389.91 ± 0.79
HWU64SIGMA-F75.10 ± 2.1082.00 ± 1.6586.10 ± 1.2890.05 ± 0.9291.35 ± 0.62

SIGMA-F has higher composition means at every budget. Exact lookup also makes its Base and Comp. results identical on matched samples and weights. The following table reports measured GPU-hours for 4-shot training only.

DatasetSIGMA-4: Comp. (%) / GPU-hSIGMA-F-4: Comp. (%) / GPU-h
BANKING7785.72 / ≈5.4986.30 / ≈3.24
CLINC15093.42 / ≈20.8493.80 / ≈12.31
HWU6485.28 / ≈3.7986.10 / ≈2.24

These results favor SIGMA-F for deployment in this interface. SIGMA's contribution is the demonstrated improvement in the model's own use of guidance under controlled comparisons. Its 75% training-label saving refers specifically to SIGMA-4 versus SFT-16.

AuthorsOpenReview ↗

Response to Reviewer nT5o (2/2)

W2 / Q3. Should the claim be restricted to binding permutations and equivalent serialization?

Response.{\color{green}\mathbf{Response.}} Your interpretation matches Eq. (2). The claim is compositional transfer for a fixed semantic task and candidate inventory, with changing bindings and equivalent representations. The held-out guidance documents are distinct from the single-edit training views and are reserved for testing.

We test the frozen 10-shot models on equivalent paraphrases, random child orders beyond the training reversal, and cross-group swaps beyond sibling edits. SIGMA-F receives the same updated descriptions and map. Each cell reports Comp. / changed-target accuracy (%).

DatasetFrozen test conditionPaired SFTTransported KLSIGMASIGMA-F
BANKING77Paraphrased descriptions + compositions83.90 / 67.0085.45 / 72.2087.20 / 80.8089.20 / 88.90
BANKING77Random serialization + compositions84.60 / 68.3085.95 / 73.2087.85 / 82.0089.91 / 89.54
BANKING77Cross-group binding compositions77.80 / 43.5079.50 / 48.0082.60 / 58.5089.91 / 89.30
CLINC150Paraphrased descriptions + compositions91.85 / 78.4092.75 / 82.0094.65 / 88.8095.65 / 95.40
CLINC150Random serialization + compositions92.40 / 79.8093.20 / 83.5094.90 / 89.5096.25 / 96.13
CLINC150Cross-group binding compositions88.50 / 55.0089.50 / 60.0091.50 / 70.0096.25 / 95.80
HWU64Paraphrased descriptions + compositions84.55 / 66.8085.45 / 71.5087.65 / 80.5089.80 / 89.50
HWU64Random serialization + compositions84.80 / 67.8086.10 / 72.9088.15 / 81.8090.49 / 90.60
HWU64Cross-group binding compositions76.80 / 42.0078.60 / 47.5081.90 / 57.0090.49 / 90.40

SIGMA exceeds both controls across all three edit types. Cross-group swaps are particularly useful here because they test bindings beyond the sibling groups used in training. The results support transfer across a broader range of guidance while preserving the semantic task. Equation (7) bounds error along valid unseen paths through their local residuals. New classes or decision rules require a corresponding change to the supervision relation.


W3. Figure 1 is difficult to read at normal page size.

Response.{\color{green}\mathbf{Response.}} Thank you for pointing this out. The table below gives a readable summary of the two schematic cases in Figure 1.

Guidance edit for the same inputRequired responseFailure illustrated for single-view SFTConstraint learned by SIGMA
Change the output bindingReturn the newly bound output πe(y)\pi_e(y)Retain the outdated outputTransport candidate evidence: ug′≈Peug\mathbf u_{g'}\approx P_e\mathbf u_g
Reorder equivalent rulesPreserve the output yyChange the answer unnecessarilyPreserve candidate evidence: ug′≈ug\mathbf u_{g'}\approx\mathbf u_g

Equation (5) combines paired target supervision with these relative-evidence constraints. Figure 3 shows the training flow. All model inputs and outputs are text.


Thank you for the detailed rebuttal and the additional experiments.

The matched shot sweep answers my label-efficiency question. SIGMA-F achieves higher mean composition accuracy at every tested label budget on all three datasets and requires less training compute in the reported 4-shot comparisons. I appreciate the gains over paired SFT and transported KL.

My main remaining concern is when direct scoring would be preferable to exact lookup in practice. The application scope is not clear.

The new cross-group swaps, equivalent paraphrases, and randomized ordering largely address my scope concern.

Thanks.

Reviewer_9jd3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to teach LLMs to following changing guidance with limited task annotations and generalize this behavior to unseen combinations of guidance changes. It introduces SIGMA (Structured Interventions for Guidance-aware Model Adaptation), which pairs the same input with two different versions of guidance, and perform supervised learning based on the two correct answers and expected relationship between the two predictions.

Preliminary study shows that, with the same data where the model can predict correctly under the original guidance, simply swapping output mapping or reordering guidance can cause the model to lost its accuracy a lot (for example, exact top-10 rank retention is only 27.3%), indicating the necessity of the study.

The main design of the method can be explained by equation (5), in section 4. It calculates a paired loss that contain 3 components: (1) the cross-entropy loss of predictions under original guidance, (2) the CE loss of predictions under the updated guidance, and (3) the alignment loss of predictions under two versions of guidance. The paper further analyzed why adding the alignment loss can make model learn better under fewer annotations.

As a result, across the three datasets (BANKING77, CLINC150, and HWU64), SIGMA achieves the highest base accuracy (accuracy under original guidance) and a smaller drop with composed guidance edits compared to baselines. Ablation studies against several control settings, label efficiency analysis, and margin-residual diagnostics are also performed.

Critical strengths

1. Preliminary study makes the motivation clear and compelling.

Section 3 - preliminary study - serves as a great illustration of why the work is necessary. The findings that exact top-10 rank retention is only about 27% (at budget 8) demonstrates that even with simple output binding swap or rule reordering, the model performance can be very unstable in a lot of cases, thus justifying the motivation of the work.

2. The method looks solid, and also lead to improved performance.

Section 4 clearly gives the settings, definitions, annotations and proofs. Table 1 show that across 3 datasets, the proposed methods achieve overall highest base accuracy and accuracy under composed guidance edits. Table 2 (ablations) show that removing alignment loss (or replace it with transported KL) or either of the two edits would cause performance drop. Figure 4 (backbone comparison) shows that SGIMA always has the highest accuracy given the same backbone and same number of shots across the experimented settings. The claim that less annotation is used is also well supported by Figure 5.

Critical weaknesses

1. The scope of application is limited.

The work studies two types of guidance changes: output binding swap, and rule reordering. They might carry some real-world benefits, but as far as I understand, there are a lot of more complex, and even more common guidance changes in real scenarios. For example, a specific guidance change on the refund criteria can change the answer of a given refund request input from "refundable" to "non-refundable". Compared to it, the guidance changes studied by this work are rather simply and straightforward. The work covers generalization study on unseen composed edits, but it is unclear whether it can also generalize to this kind of more complex real-world guidance changes.

2. The computational trade-off is under clarified.

One of the main claimed contributions is the training-label effiency. But it is unclear whether this can compensate for the higher training overhead. The paper reports that comparing 4-shot SIGMA with 16-shot SFT, SIGMA achieves higher accuracy with fewer annotations. And it also reports that comparing 10-shot SIGMA with 10-shot SFT, SIGMA achieves higher accuracy with same amount of annotations. In this case, however, the training GPU cost of SIGMA is significantly higer (5.89 additional GPU-hours, about 75.1% more). Hence, the work's trade-off between annotation efficiency and training cost is unclear.

Questions for authors

Section 2 listed several previous work about "context following and permutation consistency". However, it did not sufficiently explain how the proposed method differs from the cited previous work. For example, the paper states that symbol tuning teaches arbitrary label mappings and PBM-based LoRA explicitly trains consistency across option permutations. Both methods seem to have some design overlap with the paper. Can the author explicitly compare the proposed method with them, and justify why they were not chosen to be part of the baseline methods?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9jd3 (1/2)

Thank you for taking the time to read our paper carefully and for your thoughtful comments. We respond below to your questions about scope, computational cost, and related work.

W1. Do binding and ordering edits support more general changes, such as revised refund criteria?

Response.{\color{green}\mathbf{Response.}} Your refund example changes the semantic posterior r(z∣x)r(z\mid x). SIGMA studies changes to output bindings y=ag(z)y=a_g(z) that preserve this posterior (Eq. (2)). It also studies equivalent serializations that preserve rule priority. The motivation is straightforward: a model can recognize the intent correctly and still return an outdated output.

We evaluate the frozen 10-shot models on verified equivalent paraphrases, randomized child orders, and cross-group binding swaps. Cross-group swaps extend beyond the sibling swaps used in training. SIGMA-F receives the same updated descriptions and map. Each cell reports Comp. / changed-target accuracy (%).

DatasetFrozen test conditionPaired SFTTransported KLSIGMASIGMA-F
BANKING77Paraphrased descriptions + compositions83.90 / 67.0085.45 / 72.2087.20 / 80.8089.20 / 88.90
BANKING77Random serialization + compositions84.60 / 68.3085.95 / 73.2087.85 / 82.0089.91 / 89.54
BANKING77Cross-group binding compositions77.80 / 43.5079.50 / 48.0082.60 / 58.5089.91 / 89.30
CLINC150Paraphrased descriptions + compositions91.85 / 78.4092.75 / 82.0094.65 / 88.8095.65 / 95.40
CLINC150Random serialization + compositions92.40 / 79.8093.20 / 83.5094.90 / 89.5096.25 / 96.13
CLINC150Cross-group binding compositions88.50 / 55.0089.50 / 60.0091.50 / 70.0096.25 / 95.80
HWU64Paraphrased descriptions + compositions84.55 / 66.8085.45 / 71.5087.65 / 80.5089.80 / 89.50
HWU64Random serialization + compositions84.80 / 67.8086.10 / 72.9088.15 / 81.8090.49 / 90.60
HWU64Cross-group binding compositions76.80 / 42.0078.60 / 47.5081.90 / 57.0090.49 / 90.40

SIGMA exceeds both direct-scoring controls in every condition. On cross-group edits, changed-target accuracy improves over transported KL by 10.5/10.0/9.5 pp. These results show transfer to equivalent descriptions and broader bindings within the same semantic task and candidate inventory.


W2. Does annotation efficiency compensate for the additional training cost?

Response.{\color{green}\mathbf{Response.}} At 10 shots, SIGMA costs 75.1% more than single-view SFT. The extra cost over paired SFT using the same views is 4.2% (Table 8). The comparison below combines Table 5 accuracy with measured training times under the same hardware and accounting settings.

DatasetTraining labels: 4 / 16 shotsTrain + dev labels: 4 / 16 shotsSIGMA-4: Comp. (%) / GPU-hSFT-16: Comp. (%) / GPU-h
BANKING77308 / 1,2321,848 / 2,77285.72 / ≈5.4982.61 / ≈12.54
CLINC150600 / 2,4003,600 / 5,40093.42 / ≈20.8491.54 / ≈47.60
HWU64256 / 1,0241,332 / 2,10085.28 / ≈3.7983.14 / ≈8.67

SIGMA-4 achieves higher composition accuracy than SFT-16 with 75% fewer training labels and approximately 56% less training time. Including development labels, savings are 33.3%, 33.3%, and 36.6%. GPU-hours cover training only.

We also compare the methods at 10 shots on BANKING77 under common training-time caps. Results average five seeds. Each run stops when training finishes or the cap is reached. The SFT row at the higher cap uses the token-matched control from Table 11.

Method7.84 GPU-h cap: Comp. (%)13.73 GPU-h cap: actual GPU-h / Comp. (%)
LoRA-SFT80.9613.51 / 80.43
Paired SFT82.1013.18 / 85.37
Transported KL83.4513.46 / 86.83
SIGMA85.9013.73 / 88.56
SIGMA-F89.508.11 / 89.91

SIGMA retains its advantage over the direct-scoring baselines at the lower cap. This supports the value of its training objective under the same time limit.

AuthorsOpenReview ↗

Response to Reviewer 9jd3 (2/2)

Q1. How does SIGMA differ from Symbol Tuning and PBM-LoRA, and how does it compare experimentally?

Response.{\color{green}\mathbf{Response.}} SIGMA uses known guidance edits to teach how the full prediction should change (Eqs. (2)–(5)). Even when a swap leaves the gold answer unchanged, it constrains the evidence for other candidates. Each annotation therefore teaches both task correctness and guidance use.

Symbol Tuning [1] learns mappings from in-context examples. PBM-LoRA [2] reduces option-permutation bias through log-probability dispersion and entropy. Their original settings are multitask ICL and unsupervised MCQ answering. To answer your comparison directly, we evaluate controlled adaptations alongside the paired-SFT and KL controls in Tables 1–2.

PBM-adapted + paired CE retains the dispersion and entropy terms. We adapt its multi-permutation LoRA-1 protocol to two guidance views, full-sequence candidate probabilities, and matched LoRA capacity, with at most 15 search trials. SIGMA's centered loss removes the common score offset and penalizes discrepancies in transported pairwise log-odds exactly (Appendix B.1). Entries are 10-shot accuracies (%). Changed denotes changed targets.

DatasetKL Comp.PBM-adapted Comp.SIGMA Comp.PBM-adapted ChangedSIGMA Changed
BANKING7786.8387.5588.5677.0083.18
CLINC15094.1794.7095.3786.0090.42
HWU6486.9487.5589.0876.0083.23

Symbol-style augmentation uses opaque IDs and global random bindings sampled independently of the gold label. The schema supplies the mapping that in-context examples provide in Symbol Tuning. Both variants use the same augmentation and paired supervision, so their difference measures the benefit of alignment. Entries are Comp. accuracy (%).

DatasetGlobal binding augmentation + paired CESame augmentation + SIGMA alignmentGain (pp)
BANKING7786.8588.10+1.25
CLINC15093.9594.95+1.00
HWU6486.5088.20+1.70

Alignment improves performance with global binding augmentation, and SIGMA also exceeds the PBM-adapted control. These comparisons show that the benefit persists with stronger mapping and consistency baselines.

[1] Symbol tuning improves in-context learning in language models.

[2] Quantifying and Mitigating Selection Bias in LLMs: A Transferable LoRA Fine-Tuning and Efficient Majority Voting Approach.


Thank you for the response and additional experiments.

On scope: the new experiments remain within the same family of edits of the paper. This is confirming my concern that the application scope is limited, since many real-world guidance changes do alter the semantics of the guidance.

On training cost: my concern is resolved.

On related work: The additional comparisons are helpful, and SIGMA consistently outperforms the adapted controls. I encourage the authors to state the methodological distinction explicitly in the paper, as it is currently left for the reader to infer.

Thanks!

Comments on the paper

1 comment
Area_Chair_vaVTOpenReview ↗

Author Reviewer Discussion

Dear reviewers,

Thanks again for agreeing to review for ICLR.

As we are approaching the end of the discussion period, I would appreciate your acknowledgment of the authors' rebuttal and any feedback you would like to share with the authors.

This will help to improve the submissions, streamline decision making, and keep ICLR's original idea of having such a discussion period valuable to all participants in the review process.

Best, Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.