acceptodds
Under review as a conference paper at ICLR 2027

Understanding Explicit and Implicit Affect through Neuro-Symbolic Reasoning

Abstract

Implicit affect recognition seeks to understand what people feel beyond what they literally say. This requires distinguishing surface affect from underlying attitudes and understanding how the two relate. However, recent LLM-based methods that train on final labels or overall rewards can overlook errors in either layer and fail to teach their relationship. We propose PAIR (Polarity-Aware Implicit Affect Recognition), a neuro-symbolic framework that represents both explicit and implicit layers and uses symbolic rules to connect them, making their individual correctness and joint consistency available for learning. Specifically, neural models interpret explicit expression and implicit attitude through two reasoning paths, while symbolic rules compose their polarity judgments into affective stance, preserving both agreement and divergence. During supervised fine-tuning, symbolic verification selects trajectories with correct judgments and valid composition; during reinforcement learning, local rewards refine individual judgments, while global rewards evaluate the composed stance alongside pragmatic intent and fine-grained emotion. On CueBench, PAIR achieves three-task average accuracies of 64.25% and 53.93% on Qwen3-4B and Qwen3-1.7B, outperforming DAPO by 5.88 and 3.37 percentage points, respectively. Further analyses show complementary gains from structured supervision and semantically aligned local feedback, including improved joint explicit–implicit correctness.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Reinforcement Learning, Neural-Symbolic Reasoning, Affective Reasoning, Emotion Recognition

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_X4JKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript addresses the difficulty of LLMs in distinguishing implicit emotional expressions from explicit emotion-related questions and proposes a neuro-symbolic reasoning framework named PAIR. The framework decomposes emotion understanding into five semantic regions: explicit emotion, implicit emotion, affective stance, pragmatic intent, and fine-grained emotion, and establishes their compositional relationships through predefined symbolic structures. By combining SFT and RL, PAIR enhances the model's ability to recognize these distinctions. Experimental results demonstrate the effectiveness of the proposed method on Qwen3-4B and Qwen3-1.7B.

Critical strengths
  1. The research problem addressed in the manuscript is important and well motivated.

  2. The proposed PAIR method is well designed and offers good interpretability.

  3. The experimental results provide strong evidence for the effectiveness of PAIR.

Critical weaknesses
  1. The manuscript appears to overstate the concept of Neuro-Symbolic Reasoning, as the proposed method does not actually learn a symbolic reasoning process.

  2. The manuscript does not report the actual training data size, class distribution, or Teacher accuracy before and after filtering. The performance gains of PAIR-SFT may therefore stem from data filtering rather than structured reasoning.

  3. The authors should further clarify why it is necessary to model deterministic label relationships as structures that need to be learned.

  4. The manuscript lacks essential ablation studies, such as ablations on the Global, Local, and Structural Rewards.

  5. The weights used to combine the Global, Local, and Structural Advantages should be explicitly reported in the manuscript.

Questions for authors

Please refer to weaknesses

6 comments
AuthorsOpenReview ↗

Response to Reviewer X4JK – [1/2]

Thank you for your valuable feedback and for recognizing the strengths of our work! We address the weaknesses and questions you raised below.

[W1 & W3] Why does PAIR use fixed symbolic rules in training?

Re: PAIR (1)links polarity and stance with a fixed rule and (2)uses that rule to check consistency and guide training, while (3)retaining the model’s independent stance judgment as the final prediction.

PAIR's neuro-symbolic formulation uses explicit structural relations to guide training.

(1) Positioning of Neuro-Symbolic Reasoning.

PAIR combines neural semantic reasoning with explicit symbolic constraints. The predefined symbolic rules remain fixed during training.

The model independently predicts Explicit Polarity, Implicit Polarity, and Affective Stance from natural language. These judgments rely on context and different semantic cues.

The predefined composition mapping Ω specifies how these judgments relate. The Structural advantage provides a training signal based on whether the predictions follow this mapping.

The explicit symbolic structure links the model's predictions and guides it as it learns to reason across semantic dimensions.

Reviewer ZESd also noted that this symbolic structure helps localize errors in multi-dimensional affective reasoning and makes compositional consistency an explicit training signal.

(2) Modeling deterministic label relationships.

We keep the deterministic composition rule fixed and use it as an explicit structural prior during training.

The model still has to infer coordinated semantic judgments from natural language, even when the label relationships are deterministic.

Final classification rewards alone provide limited information about conflicts among explicit polarity, implicit polarity, and overall stance. We use Ω to check consistency across these regions and feed the result back into training.

We compared two Qwen3-4B SFT variants with and without explicit modeling of the deterministic relationship between polarity and stance (Table R8).

Table R8. Direct stance prediction on CueBench.

SFT variantSupervised regionsCorrect / TotalAccuracy
Two regionsExplicit and implicit polarity2,160 / 5,16541.82%
Three regionsExplicit polarity, implicit polarity, and affective stance3,048 / 5,16559.01%

(3) Independent stance prediction.

We use the model's direct stance prediction as the final output. This encourages the model to make its own judgment about the overall affective stance.

Table R9 shows that each method correctly predicts some cases that the other gets wrong.

Table R9. Paired stance prediction outcomes on CueBench.

CueBenchΩ composition correctΩ composition incorrectTotal
Direct prediction correct1,3574381,795 (34.75%)
Direct prediction incorrect6192,7513,370
Total1,976 (38.26%)3,1895,165

We use the fixed composition rule to guide the model as it learns to make coordinated affective judgments across semantic dimensions.

AuthorsOpenReview ↗

Response to Reviewer X4JK – [2/2]

[W2] How was the PAIR-SFT training data filtered?

Re: To put the performance comparison in context, we report (1)the training set size and class distribution before and after filtering and (2)how many teacher responses passed the filter, then (3)compare PAIR-SFT with Structured SFT trained on unfiltered responses.

(1) Training set size and class distribution.

We started with 8,000 questions and retained 7,721 training examples. Appendix B.5 reports the class counts before and after filtering.

(2) Filtering results.

The filter required answers to match the reference labels, so 100% of the retained answers matched. The final training set accounts for 33.32% of all recorded trajectories (7,721/23,170).

(3) Structured SFT and PAIR-SFT.

Structured SFT uses unfiltered teacher responses. PAIR-SFT organizes rationales into distinct semantic regions and uses symbolic verification to check the judgments. Table R10 compares the two training setups.

Table R10. Qwen3-4B SFT comparison on CueBench.

MethodFormatVerificationAffective Stance Acc.Pragmatic Intent Acc.Fine-grained Emotion Acc.Acc Avg.
Structured SFT✓✗54.4636.9030.0140.46
PAIR-SFT (Ours)✓✓58.5352.9735.3748.96

[W4 & W5] How do the three advantage components contribute, and how are they weighted?

Re: Thank you for pointing this out. We (1)compare different combinations of the global, local, and structural advantages to assess each component’s contribution and (2)give the weights used to combine them in the full model.

(1) Ablation results. We ran ablations for the Global, Local, and Structural components and report the results in Table R11.

(2) Advantage weights. In Eq. (16), the weights for the Global, Local, and Structural Advantages are 1, 0.1, and 0.05, respectively.

We will include the ablation results and advantage weights in the revised manuscript.

Table R11. Ablation of the three advantage components.

MethodGlobalLocalStructuralAffective Stance Acc.Pragmatic Intent Acc.Fine-grained Emotion Acc.Acc Avg.
Global Only (GRPO)✓✗✗65.0360.9745.8757.29
Local Only✗✓✗64.2060.4046.4057.00
Structural Only✗✗✓60.5750.6633.2148.15
Global + Local✓✓✗67.0865.6052.0061.56
Global + Structural✓✗✓65.7759.6246.3357.24
Local + Structural✗✓✓66.1460.8049.0058.65
PAIR✓✓✓70.7167.0554.9964.25

The additional experiments and tables took some time to complete, and we appreciate your patience. We will incorporate your suggestions into the revised manuscript. Thank you again for your valuable feedback! If you have any further questions, we would be happy to answer them.

Dear Reviewer,

We hope our response has addressed the concerns raised in the review. As the discussion period is approaching its end, we wanted to check whether there are any remaining questions or points we could further clarify. We would be glad to provide additional detail if helpful.

Best regards, The Authors

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

The paper proposes PAIR, a neuro-symbolic post-training framework for Chinese implicit affect recognition. It organizes model outputs into five semantic regions: explicit polarity, implicit polarity, affective stance, pragmatic intent, and fine-grained emotion. Each polarity takes a positive, neutral, or negative value, and CueBench’s predefined mapping Ω composes the polarity pair into one of nine stance labels. PAIR-SFT filters teacher-generated trajectories using gold-label correctness and compositional consistency, while PAIR-RL combines global, local, and structural advantages through region-specific routing. The central contribution is to incorporate compositional label relationships into structured supervision and reinforcement-learning credit assignment.

Experiments on CueBench use Qwen3-4B and Qwen3-1.7B. PAIR-RL achieves average accuracy across the three tasks of 64.25% and 53.93%, respectively, exceeding GDPO by 4.63 and 0.63 percentage points. The paper also presents routing and reward-coupling analyses, with additional transfer results on CPED and MELD in Appendix B.

Critical strengths
  1. The task decomposition is well motivated by the distinction between literal expression and underlying attitude. The consistency constraint checks the compositional relationship between explicit polarity, implicit polarity, and stance, rather than forcing the two polarities to agree.

  2. The RL comparisons use a common PAIR-SFT initialization, making the evaluation more controlled and fair. Semantically aligned routing outperforms random and swapped routing, providing evidence that directing feedback to the relevant semantic regions matters.

Critical weaknesses

The central structural-feedback mechanism is insufficiently specified for reproducibility. Although the paper defines the structural-consistency reward and the local gating rules, it does not provide a complete, executable definition of G_struct in Eq. (16). In particular, it remains unclear how the structural advantage is routed to the explicit-polarity, implicit-polarity, and stance regions, and how the update direction is determined for each region.

These omissions concern the mechanism underlying the claimed reverse supervision from stance to intermediate polarity judgments, rather than peripheral implementation details. A reproducing researcher would therefore have to make substantive algorithmic choices that could affect the results. This limits my confidence in the reproducibility and interpretation of the central contribution. The paper should provide complete gating equations or a condition table, corresponding pseudocode, and the actual hyperparameter values.

Questions for authors
  1. How is R = 0.8 calculated for both responses in Figure 1? The displayed correctness indicators appear to imply global rewards of 0.3 and 0.7 under the reported weights. Please clarify whether these scores are illustrative or follow a different reward definition.

  2. Could you provide representative success and failure cases from actual model outputs, comparing PAIR with a baseline on the same inputs? Please include cases where structural feedback corrects errors and where predictions remain consistent but incorrect.

  3. Beyond GLM-5.1, could you include additional frontier models, such as GPT-5.6-sol and recent DeepSeek models, under matched evaluation settings, or clarify the intended scope of the cross-model claims?

  4. Does the 8,000-example count refer to the pre-filtering pool or the final training set? Please clarify candidate generation, resampling, retained counts and class distributions, and whether ordinary SFT used equivalent filtering. Were retained rationales manually audited? If so, please summarize the protocol and annotator agreement.

  5. Could you compare directly predicted stance with stance computed from the two predicted polarities using Ω? This would test whether the separate stance prediction and consistency reward outperform deterministic composition, which the coupled-versus-uncoupled comparison in Figure 4(b) does not establish.

  6. The Qwen3-1.7B improvement over GDPO is 0.63 percentage points (53.93 versus 53.30). Could you report results across multiple independent training seeds, including means, standard deviations, and uncertainty estimates for the difference?

8 comments
AuthorsOpenReview ↗

Response to Reviewer zfqZ – [1/2]

Thank you for your encouraging comments on PAIR’s design and evaluation! Our point-by-point responses follow.

[W1 & W2 & Q2] How does PAIR route structural feedback?

Re: To make PAIR’s structural update reproducible, we explain (1)when the gate applies a penalty, (2)which polarity or stance predictions receive it, and (3)how we implement the update in training, then illustrate the model’s behavior with (4)a success case and a failure case.

(1) Structural gate and update direction.

Equation (11) assigns a reward of 1 if the predicted stance matches the stance composed from the two predicted polarities. Otherwise, the reward is 0. Equation (12) normalizes this reward within each rollout group. The gate in Eq. (16) applies to explicit polarity (EE), implicit polarity (II), and stance (SS).

= (1-R_i^{\mathrm{struct}})(1-R_{i,r}^{\mathrm{local}}), \qquad r\in\\{E,I,S\\}.$$ - **Activation.** The local reward is 1 when the region's prediction matches its gold label. The gate opens only for **incorrect regions in structurally inconsistent responses**. - **Direction.** Selected regions receive a **negative structural term** in groups with both consistent and inconsistent responses. If the reward is constant within the group, the structural advantage is zero. **(2) Structural feedback to polarity judgments.** When the predicted and composed stances disagree, gold labels identify the incorrect regions. - **Stance is correct.** Feedback goes to the incorrect polarity regions. - **Both polarities are correct.** Feedback goes to the incorrect stance region. - **All three predictions are wrong.** All three regions receive feedback. **(3) Reproducibility and training setup.** We will **open-source the experiment code and configurations on GitHub as soon as the paper is accepted**. **Table 4, in the main paper** (Appendix B.3, RL Training Setup), lists the rollout and optimization settings, including 8 rollouts per input. The global reward weights for stance, intent, and emotion are **0.4 / 0.3 / 0.3**. The global, local, and structural advantage coefficients are **1**, **0.1**, and **0.05**, respectively. ```text Initialize the policy from PAIR-SFT. For each training step, sample 8 responses per input. For each group, compute rewards with Eq. (20) and normalize with Eq. (21). For each token, start with λ_global × A_global. In any semantic region r, add λ_local × (1 − R_local[r]) × A_local[r]. In E, I, or S, add 0.05 × G_struct[r] × A_struct using the structural gate above. Update the policy with the group-relative objective and these token advantages. ``` **(4) Case studies.** **Success case** GRPO predicts Neutral / Neutral / Veiled Negative. Its implicit polarity and stance are incorrect, and the stance conflicts with its two predicted polarities. PAIR predicts Neutral / Positive / Understated Positive. All three labels are consistent and match the gold labels. **Failure case** PAIR predicts Negative / Positive / Affiliative Positive. The three labels are structurally consistent, and all differ from the gold labels (Neutral / Negative / Veiled Negative). See the **Appendix C.3** for the full inputs and model outputs for both cases. --- > [Q1] Clarification of Figure 1 **Re:** The equal rewards in Figure 1 were illustrative, and we will revise the example to reflect the actual reward weights. We apologize for the oversight. We intended the value of 0.8 in Figure 1 to show that responses with different path-level correctness patterns can receive the same global reward. That value does not match the actual reward weights. We will address this oversight in the revised PDF. --- > [Q3] How do additional frontier models perform on CueBench? **Re:** We evaluated DeepSeek-V4.1-Flash and GPT-5.6-Sol on CueBench under the same settings as GLM-5.1. Thank you for your suggestion. We will provide additional evaluation results for DeepSeek-V4.1-Flash and GPT-5.6-Sol under the same evaluation settings, as shown in **Table R4**. **Table R4. Evaluation of additional frontier models on CueBench.** | Model | Stance Acc. (%) | Intent Acc. (%) | Emotion Acc. (%) | Acc Avg. (%) | |---|---:|---:|---:|---:| | DeepSeek-V4.1-Flash | 51.55 | 44.19 | 34.11 | 43.28 | | GPT-5.6-Sol | 49.81 | 47.09 | 30.62 | 42.51 |
AuthorsOpenReview ↗

Response to Reviewer zfqZ – [2/2]

[Q4] How was the PAIR-SFT training set built?

Re: To clarify how we built the PAIR-SFT training set, we (1)distinguish the initial question pool from the final set, (2)explain how we generated and filtered candidates, and (3)report the resulting class coverage, then clarify (4)which filters Ordinary SFT shared and (5)whether the retained trajectories underwent human review.

(1) Training set size. We gave the teacher 8,000 questions before filtering. The final PAIR-SFT training set contains 7,721 examples.

(2) Candidate generation and filtering. Qwen3-30B generated all three candidate trajectories for each question at once, yielding 24,000 candidates. We removed 830 candidates that failed the format check, leaving 23,170. For each question, we kept the first candidate in generation order whose answer labels all matched the reference labels. 7,777 questions had a qualifying candidate. During rationale cleaning, we removed 56 trajectories with problematic chains of thought. This left 7,721 training examples.

(3) Class coverage. The final dataset covers all 9 affective stance classes, 8 pragmatic intent classes, and 25 fine-grained emotion classes. Per-class counts range from 243 to 1,295 for stance, 427 to 2,207 for intent, and 12 to 680 for emotion. We will report the per-class counts in the appendix of the revised PDF.

(4) Ordinary SFT filtering. Ordinary SFT used the same format and answer-label filtering rules as PAIR-SFT.

(5) Human review. We used automatic filtering to select the reasoning trajectories, with no additional human review.


[Q5] How does direct stance prediction compare with deterministic composition?

Re: We (1)compare direct stance prediction with deterministic composition on the same inputs and use the cases each gets right to explain (2)why PAIR retains an independent stance judgment.

(1) Direct prediction and deterministic composition. Table R5 shows that deterministic composition achieves higher overall stance accuracy than direct prediction in this experiment. In Table R6, each method correctly predicts some cases that the other gets wrong.

We attribute this difference to the difficulty of predicting stance directly. Stance prediction must integrate multiple levels and perspectives. Explicit polarity focuses on the target utterance, and implicit polarity focuses on the context. These narrower judgments are comparatively easier.

(2) Why we retain independent stance prediction. Our method encourages the model to interpret the input through explicit polarity, implicit polarity, and overall stance. We therefore retain independent stance prediction alongside deterministic composition. The results in Table R6 suggest that the two judgments can inform each other.

The structural advantage applies a consistency constraint across regions and provides feedback for multi-region reasoning. This design preserves an independent stance judgment. We make no claim of higher standalone stance accuracy for direct prediction.

Table R5. Qwen3-4B different prediction on CueBench.

MethodCorrect / TotalAffective Stance Acc. (%)
Deterministic composition1,976 / 5,16538.26
Direct prediction1,795 / 5,16534.75

Table R6. Paired stance prediction outcomes on CueBench.

CueBenchΩ composition correctΩ composition incorrectTotal
Direct prediction correct1,3574381,795 (34.75%)
Direct prediction incorrect6192,7513,370
Total1,976 (38.26%)3,1895,165

[Q6] How stable are Qwen3-1.7B’s results across training seeds?

Re: We report Qwen3-1.7B results across independent training seeds.

We trained Qwen3-1.7B with multiple independent random seeds and kept the experimental settings fixed. The experimental results are shown in Table R7.

The revised appendix also reports Qwen3-4B results for different gating strategies across random seeds.

Table R7. Qwen3-1.7B results across random seeds.

MethodSeedAffective StancePragmatic IntentFine-grained EmotionAcc Avg.
PAIR 1.7B184726395160.5055.8043.5053.27
620583174961.4056.6044.8054.27
937164820562.9058.0046.4055.77
Mean ± SD61.60 ± 1.2156.80 ± 1.1144.90 ± 1.4554.43 ± 1.26

We appreciate your patience while we ran the additional experiments and put together the tables. Thank you again for your helpful comments! We’ll incorporate your feedback into the revised manuscript. We are happy to take any further questions you may have.

Dear Reviewer,

We hope our response has addressed the concerns raised in the review. As the discussion period is approaching its end, we wanted to check whether there are any remaining questions or points we could further clarify. We would be glad to provide additional detail if helpful.

Best regards, The Authors

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

The paper studies implicit affect recognition as a structured reasoning problem and proposes PAIR, a neuro-symbolic two-stage framework. PAIR represents a response as five semantic regions (explicit polarity, implicit polarity, affective stance, pragmatic intent, fine-grained emotion), uses a symbolic composition rule to relate explicit/implicit polarity to affective stance, and trains first with verifier-filtered structured SFT and then with region-routed RL that combines global, local, and structural advantages. On CueBench with Qwen3-4B and Qwen3-1.7B, PAIR-RL reports the best three-task average accuracy, with additional transfer results on CPED and MELD.

Critical strengths
  1. The central formulation is useful and well motivated: treating explicit polarity, implicit polarity, and affective stance as a symbolically checkable structure gives a concrete way to localize errors in multidimensional affective reasoning. This is consequential because it moves beyond response-level rewards and makes compositional consistency an explicit learning signal.

  2. The two-stage design is coherent and empirically supported: PAIR-SFT provides large gains over ordinary SFT, and PAIR-RL further improves over GRPO/DAPO/GDPO under the same initialization. The region-level and credit-direction analyses also help explain where the gains come from, which strengthens the paper’s main claim.

Critical weaknesses
  1. The comparison between ordinary SFT and PAIR-SFT may conflate several factors: structured output format, teacher-generated trajectories, verifier filtering, and symbolic composition. A structured SFT baseline without the symbolic verifier/composition check would be needed to isolate how much of the gain comes from the proposed neuro-symbolic structure rather than from richer supervision and output formatting. This matters because the paper’s main claim is about symbolic structure, not just structured prompting.

  2. The method depends on fine-grained annotations for all five semantic regions and on a fixed benchmark-specific composition mapping. This makes the approach potentially difficult to extend beyond CueBench-like data, and the paper does not sufficiently analyze robustness when some region labels are missing or when the composition rule is noisy. This matters for the claimed generality of the framework.

Questions for authors
  1. Can you add a structured SFT control that uses the same five-region output format and teacher trajectories but removes the symbolic verifier and composition check? This would clarify how much of PAIR-SFT’s gain is due to symbolic structure versus structured supervision.

  2. How sensitive are the results to the fixed composition mapping Ω and to missing or noisy region-level annotations? Can you report a stress test or a variant with partial labels?

  3. Can you report variance or significance across seeds for the main CueBench results and the credit-direction comparison, especially since the error-only gate is claimed to be the best final setting?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ZESd

Thank you for highlighting the value of PAIR’s symbolic structure! We address your comments and questions below.

[W1 & Q1] What do symbolic verification and composition checks add?

Re: Symbolic verification and composition checks improve accuracy across all three CueBench tasks.

We added a Structured SFT control. It used the same five-region output format and teacher-generated candidate trajectories as PAIR-SFT. We disabled symbolic verification and composition checks. Both runs used the same remaining training and evaluation settings. Results are reported in Table R1.

Table R1. Qwen3-4B SFT comparison

MethodFormatVerificationAffective Stance Acc.Pragmatic Intent Acc.Fine-grained Emotion Acc.Acc Avg.
Ordinary SFT✗✗32.8334.9026.0931.27
Structured SFT✓✗54.4636.9030.0140.46
PAIR-SFT (Ours)✓✓58.5352.9735.3748.96

[W2 & Q2] How does PAIR-SFT perform in the stress test?

Re: We test (1)how PAIR-SFT responds to changes in the CueBench training data and (2)how well it transfers beyond CueBench to CPED and MELD.

Thank you for this helpful suggestion. We are happy to share the results of our stress test.

(1) CueBench stress test. We stress-tested PAIR-SFT with three changes to the training data. We introduced errors in the composition mapping (Ω), randomly masked region annotations, and corrupted region labels.

Average accuracy was lower than for PAIR-SFT trained on clean data. It remained above the Ordinary SFT baseline. Results are reported in Table R2.

(2) Transfer to CPED and MELD. We also evaluated transfer on CPED and MELD. See Appendix B.4 for the results.

Table R2. Qwen3-4B SFT stress test on CueBench

MethodAffective Stance Acc.Pragmatic Intent Acc.Fine-grained Emotion Acc.Acc Avg.
PAIR-SFT (stress test)53.01%35.00%28.85%38.95%

[Q3] How do the gates compare across training seeds?

Re: We report results across multiple random training seeds.

We compared three credit assignment strategies on Qwen3-4B. Each used the same three random seeds and training settings. Table R3 reports the results for each seed, along with the mean and standard deviation.

Table R3. Qwen3-4B results across three seeds

GateTraining SeedAffective StancePragmatic IntentFine-grained EmotionAcc Avg.
Error-only184726395169.8767.6554.0063.84
620583174970.7566.4755.2064.14
937164820571.6567.3556.1265.04
Mean ± SD70.76 ± 0.8967.16 ± 0.6155.11 ± 1.0664.34 ± 0.62
Correct-only184726395163.8361.8245.9257.19
620583174965.3060.4147.3057.67
937164820565.0061.5548.1158.22
Mean ± SD64.71 ± 0.7861.26 ± 0.7547.11 ± 1.1157.69 ± 0.52
Bidirectional184726395166.7265.0049.6060.44
620583174968.1063.7150.8060.87
937164820568.0064.6552.0061.55
Mean ± SD67.61 ± 0.7764.45 ± 0.6750.80 ± 1.2060.95 ± 0.56

We appreciate your patience while we ran the additional experiments and put together the tables. Thank you again for your helpful comments! We will incorporate your suggestions into the revised manuscript. We appreciate the opportunity to improve the paper and are happy to address any further questions.

Dear Reviewer,

We hope our response has addressed the concerns raised in the review. As the discussion period is approaching its end, we wanted to check whether there are any remaining questions or points we could further clarify. We would be glad to provide additional detail if helpful.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Chairs and Reviewers,

We sincerely thank the chairs and reviewers for their time and effort in handling the large volume of ICLR 2027 submissions. We are deeply grateful for your feedback on our paper. So far, we have received ratings of 3, 2, and 2.

We introduce PAIR, a verifiable neuro-symbolic learning framework for coordinated implicit affect reasoning. All reviewers responded positively to our core contributions.

  1. Reviewer ZESd (Rating 3) recognized (1) the value of our symbolic coupling design. It helps localize errors in multidimensional affective reasoning and makes compositional consistency an explicit learning signal. The reviewer also acknowledged (2) the coherence and performance gains of the two-stage learning framework and (3) the value of the empirical analyses.

  2. Reviewer zfqZ (Rating 2) found (1) the task decomposition well motivated and recognized (2) the effectiveness of semantically aligned credit assignment. The reviewer confirmed that our responses clarified (3) concerns about the central structural-feedback mechanism and the fairness of the training comparison. The reviewer also noted (4) that PAIR enabled smaller models to outperform the evaluated frontier models on CueBench.

  3. Reviewer X4JK (Rating 2) found (1) our research problem important and well motivated. The reviewer appreciated (2) PAIR's design and interpretability and (3) the strong empirical evidence for its effectiveness. The reviewer also confirmed (4) that four of the five concerns had been adequately addressed or clarified.

Reviewer zfqZ (Rating 2): Misunderstanding about the ablations and symbolic coupling. We have fully addressed all of Reviewer zfqZ's concerns. In the reviewer's last response, only one remaining question was raised:

Do independent stance prediction and structural feedback contribute to model training?

  • SFT ablation (Table R8): We previously reported results with and without stance supervision. We then reported the deterministic composition results. Stance supervision improves the accuracy of both direct stance prediction and deterministic composition.
  • RL ablation (Table R11): We previously reported the structural-feedback ablation. Adding structural feedback to global and local feedback further improves performance across all three tasks.

These experiments show how independent stance prediction and structural feedback contribute to coordinated learning. They directly address the reviewer's concern.

Reviewer X4JK (Rating 2): Misunderstanding about the SFT controls and the source of performance gains. Our rebuttal fully addresses every concern raised by Reviewer X4JK. In the latest exchange, only one remaining question was raised:

Are the performance gains of PAIR-SFT primarily attributable to filtering correctly labeled samples?

  • Matched-filter SFT control: The original paper compares Ordinary SFT and PAIR-SFT under matched conditions. Both use the same teacher model, answer-label filtering criteria, and number of training examples. PAIR-SFT achieves substantial performance gains in this comparison.
  • Structured SFT control: We also compared PAIR-SFT with Structured SFT using the same output format. PAIR-SFT also performs better in this comparison.

These results show that PAIR-SFT's gains extend beyond the shared answer-label filtering.

We have directly addressed the remaining question from each reviewer, zfqZ and X4JK. We understand that the reviewers may not yet have had time to consider our follow-up responses or revisit their ratings.

We submit this work with great passion. It reflects our careful design and bold rethinking of how to combine neuro-symbolic reasoning with RL for implicit affect recognition. We also devoted considerable effort during the rebuttal to clarifying and strengthening our case. We hope our rebuttal helps inform your final assessment. We sincerely thank you for taking the time to consider our work.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.