{"id":"0a70fae6-9ce6-4ed9-bb22-bb40ce8a660f","arxiv_id":"2607.05911","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":8,"one_line_summary":"PRPC reformulates CZSL as a five-step bidirectional reasoning process in an MLLM with GRPO-based RL post-training, achieving state-of-the-art on three benchmarks.","lead":"The paper applies structured chain-of-thought reasoning with reinforcement learning to compositional zero-shot learning, using an MLLM to iteratively predict and correct attribute-object pairs. It offers a new reasoning-centric paradigm for CZSL that could improve how vision-language models handle unseen concept combinations.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Table 6 ablation shows bidirectional correction steps (Steps 3–5) reduce HM and unseen accuracy on MIT-States, directly contradicting the central claim on the largest benchmark.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the specific concern should be redirected. The reader identified GPT-4o CoT quality as the weakest assumption and RL degradation as a key issue. While these are valid, the most load-bearing problem is that the paper's own ablation (Table 6) does not clearly support its central thesis: on MIT-States, removing the bidirectional correction steps yields better HM and unseen accuracy than the full model. The paper claims bidirectional correction 'provides a substantial performance boost' (§4.3), but the data shows this only holds on C-GQA, not MIT-States, and VAW-CZSL is unreported. This is not a missing experiment — it is an existing result that contradicts the claim. The verdict remains CONDITIONAL because: (1) Table 2/3 show primitive-level correction does help in step-wise analysis, (2) C-GQA shows some benefit, (3) the full pipeline with RL (Table 1) does achieve improvements. But the paper needs to either explain the MIT-States regression in Table 6 or soften its claim about bidirectional correction's universal effectiveness. The reader's other concerns (no code, no significance tests, missing recent CZSL baselines) remain valid supporting issues.","tokens_in":15881,"tokens_out":2983,"duration_ms":194834,"concrete_test":"Re-run the Table 6 ablation on all three datasets (including VAW-CZSL, which is currently missing), reporting Seen, Unseen, and HM for 'w/o Steps3-5' vs full 5-step PRPC under identical SFT-only conditions. If the correction steps consistently reduce or fail to improve Unseen accuracy and HM across all three datasets, the central claim that bidirectional correction improves compositional generalization is not supported by the paper's own evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that bidirectional corrective reasoning (Steps 3–5) outperforms unidirectional conditional modeling (Steps 1–2 only) for compositional generalization. Table 6 directly tests this by comparing 'PRPC w/o Steps3-5' (unidirectional) against full PRPC (bidirectional), both under SFT-only. On MIT-States — the largest and most standard benchmark — the unidirectional variant achieves HM=26.7 with Unseen=28.5, while the full bidirectional model achieves HM=26.4 with Unseen=25.0. The correction steps thus DECREASE harmonic mean by 0.3 points and unseen accuracy by 3.5 points on MIT-States. The bidirectional model only improves Seen accuracy (27.4 vs 25.2), suggesting it overfits to seen compositions rather than improving zero-shot generalization. On C-GQA, the improvement is modest (HM 20.8→21.9). Critically, VAW-CZSL results are omitted from Table 6 entirely, leaving only 1 of 2 reported datasets showing a benefit. This ablation is the most direct test of the paper's core contribution, and it does not clearly support the central claim on the primary benchmark. The reader focused on GPT-4o CoT quality and RL degradation (Table 4), which are real but secondary; the Table 6 result strikes at the heart of the bidirectional correction thesis itself.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes PRPC, a Progressive Reasoning framework with Primitive Correction for Compositional Zero-Shot Learning (CZSL). The core idea is to reformulate CZSL as a structured five-step Chain-of-Thought (CoT) reasoning process within a Multimodal Large Language Model (MLLM), where attribute and object predictions mutually correct each other bidirectionally, as opposed to independent prediction or unidirectional conditional modeling. The framework consists of two training stages: Stage I performs supervised fine-tuning (SFT) using GPT-4o-generated CoT traces, and Stage II applies Group Relative Policy Optimization (GRPO) with step-level exact-match rewards to further refine reasoning quality. Experiments are conducted on MIT-States, C-GQA, and VAW-CZSL under two evaluation settings (open-form generation and CLIP-similarity-based closed-set classification), with ablations on training stages, CoT designs, and reward weightings.","tokens_in":16191,"tokens_out":1619,"duration_ms":441619,"significance":"The paper introduces a novel reasoning-centric perspective to CZSL, a domain traditionally dominated by embedding-based and disentanglement-based approaches. The formulation of compositional recognition as bidirectional corrective reasoning in MLLMs is a conceptually interesting contribution. The two-stage training pipeline (SFT + GRPO with step-level rewards) is technically detailed, and the error injection experiments (Table 3) provide a falsifiable test of the correction mechanism's robustness. The approach is evaluated on three standard benchmarks with multiple MLLM baselines, and the qualitative examples (Table 5) illustrate the correction process concretely.","major_comments":[{"comment":"Table 6, MIT-States row: The central claim of the paper is that bidirectional corrective reasoning (Steps 3–5) outperforms unidirectional conditional modeling (Steps 1–2 only) for compositional generalization. However, the ablation directly contradicts this on the largest benchmark. 'PRPC w/o Steps3-5' achieves HM=26.7 and Unseen=28.5, while the full bidirectional model achieves HM=26.4 and Unseen=25.0. The correction steps thus decrease harmonic mean by 0.3 points and unseen accuracy by 3.5 points on MIT-States. The full model only improves Seen accuracy (27.4 vs. 25.2), suggesting overfitting to seen compositions rather than improving zero-shot generalization. This is the most direct test of the paper's core contribution, and it does not support the central claim on the primary benchmark. The text in §4.3 states 'The results clearly demonstrate the effectiveness of using CoT for Bidir[","section":null},{"comment":"Table 6: VAW-CZSL results are entirely omitted from the bidirectional correction ablation, leaving only C-GQA showing a benefit (HM 20.8→21.9, modest). Without VAW-CZSL, the ablation covers 2 of 3 benchmarks, and on 1 of those 2 the central claim is contradicted. The authors should include VAW-CZSL in Table 6 to allow readers to assess whether bidirectional correction generalizes.","section":null},{"comment":"Table 4, MIT-States and VAW-CZSL rows: Stage II (RL) alone sometimes degrades performance relative to Stage I (SFT). On MIT-States, AUC drops from 11.5 (Stage I) to 10.8 (Stage II), and on VAW-CZSL from 1.0 to 0.7. The paper's final PRPC results (combining both stages) show gains, but the claim that 'GRPO post-training enhances reliability and logical consistency' (§1, §3.6) is undermined when Stage II alone consistently underperforms Stage I. The authors attribute this to sparse rewards and malformed trajectories (§4.3), but this explanation is not rigorously tested. A more thorough analysis of why RL degrades performance on some benchmarks — and whether the gains in the full model come from SFT or RL — is needed.","section":null},{"comment":"§3.4: The entire SFT pipeline depends on GPT-4o-generated CoT traces, but no systematic quality evaluation of the generated data is provided. The paper states only that 'parsing functions verify the correctness.' Parsing functions can verify format and exact-match of the final label, but they cannot assess whether the intermediate reasoning steps are faithful, non-hallucinated, or logically consistent. If the generated traces contain systematic reasoning errors, the model learns to mimic flawed patterns. A human or automated quality assessment of a sample of generated CoT traces (e.g., step-level accuracy, logical consistency rate) would substantially strengthen the contribution.","section":null}],"minor_comments":[{"comment":"Table 2 and Table 3: The column headers are ambiguous. 'Obj-r' and 'Attr-r' appear twice with different values. It is unclear which columns correspond to which step. Clearer labeling (e.g., 'Obj-r (Step 1)', 'Obj-r (Step 3)') would improve readability.","section":null},{"comment":"§3.1: The term 'open-form compositional prediction' is introduced but not clearly contrasted with standard 'open-world CZSL.' The relationship between these settings should be clarified.","section":null},{"comment":"Table 1: The caption mentions 'Best accuracy values of primitives {Attr, Obj} are also reported,' but it is unclear whether these are from Setting 1 or Setting 2. This should be specified.","section":null},{"comment":"§4.2: The paper claims PRPC 'achieves state-of-the-art performance' but Table 1 shows CLIP outperforming PRPC on AUC on MIT-States (11.0 vs. 11.5 — a marginal difference) and VAW-CZSL (0.2 vs. 1.4). The SOTA claim should be qualified more carefully, especially given that PRPC uses the same CLIP text encoder for Setting 2.","section":null},{"comment":"Figure 2: The system prompt text is too small to read clearly. A higher-resolution or simplified version would help.","section":null},{"comment":"The paper uses 'Qwen3.0-VL-8B' in Table 1 but cites 'Qwen-VL [Bai et al., 2025]' whose title refers to 'Qwen2.5-VL.' The version should be consistently referenced.","section":null},{"comment":"§3.6, Eq. (5): The condition 'w_k > 0 only if both r_ans and r_fmt equal 1' is described in text but not reflected in the equation. This conditional gating should be made explicit in the formulation.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Table 6 is the most serious issue: the central ablation does not support the central claim on MIT-States. Combined with the RL degradation in Table 4 and the missing VAW-CZSL ablation, the paper's core contribution is not adequately validated. The paper is conceptually interesting and the writing is competent, but the experimental evidence does not currently support the claims made. I would recommend the authors address these issues with additional experiments and a more nuanced discussion of when bidirectional correction helps vs. hurts. The GPT-4o CoT quality concern is also important but is a more standard 'missing analysis' issue that could be addressed in revision."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful and constructive review. The referee raises four major points: (1) the bidirectional correction ablation in Table 6 appears to contradict the central claim on MIT-States; (2) VAW-CZSL is missing from that ablation; (3) Stage II (RL) alone degrades performance on some benchmarks, undermining the claim about GRPO; and (4) no systematic quality evaluation of GPT-4o-generated CoT traces is provided. We address each point below.","responses":[{"response":"The referee is correct that on MIT-States under the Setting 1 (open-form generation) evaluation, the variant without Steps 3–5 achieves slightly higher HM (26.7 vs. 26.4) and Unseen (28.5 vs. 25.0) than the full five-step model. We acknowledge that this is a genuine weakness in the current manuscript and that the text in §4.3 overstates the evidence on this benchmark. However, we respectfully disagree that this invalidates the core contribution, for three reasons. First, Table 6 evaluates only Stage I (SFT) variants — none of the rows include Stage II (RL). The full PRPC model (SFT + RL) in Table 4 achieves HM=29.2 and Unseen=30.7 on MIT-States, substantially outperforming both Stage I ablation variants. The benefit of bidirectional correction is realized most strongly when combined with RL post-training, because the step-level rewards in Stage II specifically target the correction steps (Steps 3–4) and the reward weighting ablation in Table 7 shows that higher rewards on later steps yield the best performance. Second, the error injection experiment (Table 3) provides a controlled, falsifiable test: when an incorrect object is injected at Step 1, the correction steps recover accuracy (e.g., on MIT-States, Pair-r improves from 1.0 to 12.7). This directly demonstrates that the correction mechanism functions as intended. Third, on C-GQA the ablation does show a consistent benefit (HM 20.8→21.9, Unseen 19.3→20.6). We concede that the MIT-States Setting 1 result is mixed and that the current text is too strong. In the revision we will (a) revise the claim in §4.3 to accurately reflect that bidirectional correction shows clear benefits on C-GQA and in the controlled error-injection setting, but yields mixed results on MIT-States under Setting 1 alone; (b) add a discussion of当","revision_made":"no","referee_comment":"Table 6, MIT-States row: The central claim of the paper is that bidirectional corrective reasoning (Steps 3–5) outperforms unidirectional conditional modeling (Steps 1–2 only) for compositional generalization. However, the ablation directly contradicts this on the largest benchmark. 'PRPC w/o Steps3-5' achieves HM=26.7 and Unseen=28.5, while the full bidirectional model achieves HM=26.4 and Unseen=25.0. The correction steps thus decrease harmonic mean by 0.3 points and unseen accuracy by 3.5 points on MIT-States. The full model only improves Seen accuracy (27.4 vs. 25.2), suggesting overfitting to seen compositions rather than improving zero-shot generalization. This is the most direct test of the paper's core contribution, and it does not support the central claim on the primary benchmark. The text in §4.3 states 'The results clearly demonstrate the effectiveness of using CoT for Bidir["}],"tokens_in":15989,"tokens_out":769,"duration_ms":235404,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper reformulates CZSL as open-form generative reasoning in an MLLM with a five-step bidirectional correction loop, and that framing is genuinely new. But the ablation that most directly tests the core thesis — that bidirectional correction beats unidirectional conditioning — does not clearly support that claim on MIT-States, the largest and most standard benchmark. That problem is real and is the main reason I'm hesitant rather than positive. What's new and good: casting CZSL as structured CoT reasoning rather than closed-set classification is a meaningful direction. The five-step template (predict object, predict attribute, re-correct object from attribute, re-correct attribute from object, conclude) is a clean design. The error injection experiment (Table 3) shows the model can recover from injected wrong objects to some degree, which is encouraging. The step-wise reward weighting ablation (Table 7) is a useful finding — later-step rewards matter more than early-step ones. And the full PRPC (SFT+RL) does improve over SFT-only across all three datasets (Table 4), so RL adds value when built on top of SFT. Now the soft spots. The stress-test concern about Table 6 is largely correct. On MIT-States under SFT-only, the full bidirectional model gets HM=26.4 and Unseen=25.0, while the unidirectional variant (Steps 1-2 only) gets HM=26.7 and Unseen=28.5. So the correction steps slightly hurt harmonic mean and notably hurt unseen accuracy on the primary benchmark. The bidirectional model does better on Seen (27.4 vs 25.2) and AUC (11.5 vs 10.9), which suggests it may be fitting seen compositions more tightly rather than improving zero-shot generalization. On C-GQA the improvement is modest (HM 20.8 to 21.9). VAW-CZSL is absent from Table 6 entirely. The paper claims Steps 3-5 provide a 'substantial performance boost,' which is overstated given these numbers. Second, the Table 1 comparison excludes specialized CZSL methods (Troika, prompting-based approaches, etc.), so the 'state-of-the-art' claim is not well-supported — the comparison is against generic MLLMs and a CLIP baseline. Third, no code or hyperparameters are released, and there are no error bars. The GPT-4o CoT quality concern raised by the reader is valid but secondary — the bigger issue is the ablation. This paper is for CZSL researchers and people interested in reasoning-based visual recognition. The idea is worth pursuing and the direction is right, but the authors need to confront the Table 6 result honestly, add VAW-CZSL to that ablation, and compare against actual CZSL specialists. I'd recommend a serious referee — the contribution is real enough to warrant careful review, but the central empirical claim needs to be either softened or better supported.","headline":"Bidirectional correction idea is novel for CZSL, but the key ablation (Table 6) does not cleanly support the central claim on the largest benchmark.","tokens_in":16681,"tokens_out":1935,"would_cite":false,"duration_ms":132676,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Bidirectional reasoning corrects its own mistakes in zero-shot recognition","keywords":["compositional zero-shot learning","chain-of-thought reasoning","bidirectional correction","multimodal large language model","reinforcement learning post-training","GRPO","attribute-object recognition","step-level rewards"],"falsifier":"If the bidirectional correction steps (Steps 3–4) were removed and performance remained unchanged or improved, or if the error-injection experiment showed no recovery from deliberately incorrect initial predictions, the central claim that mutual correction reduces error propagation would be undermined.","tokens_in":16087,"feed_emoji":"🔄","tokens_out":974,"duration_ms":175892,"temperature":0.7,"pith_summary":"The paper proposes PRPC, a framework that recasts compositional zero-shot learning — the task of recognizing unseen attribute-object pairs from known primitives — as a structured five-step chain-of-thought reasoning process inside a multimodal large language model. The central idea is that attribute and object predictions should not be made independently or in a single unidirectional pass, because an early error in one primitive propagates and corrupts the other. Instead, PRPC has the model first predict an object, then an attribute conditioned on that object, then re-correct the object using the attribute as a disambiguating cue, then re-confirm the attribute given the corrected object, and finally emit the composition. This bidirectional correction loop is the paper's core mechanism. The model is trained in two stages: supervised fine-tuning on GPT-4o-generated chain-of-thought traces to learn the five-step format, followed by reinforcement learning post-training using GRPO with step-level exact-match rewards that specifically reinforce correctness at each intermediate reasoning step. Experiments on three CZSL benchmarks (MIT-States, C-GQA, VAW-CZSL) show state-of-the-art results, and controlled error-injection experiments demonstrate that the correction loop can recover correct primitives even when the initial object prediction is deliberately wrong.","feed_headline":"Bidirectional reasoning corrects its own mistakes in zero-shot recognition","feed_subtitle":"An MLLM that re-checks object and attribute predictions against each other outperforms one-pass methods on unseen concept pairs.","key_machinery":"Five-step bidirectional corrective reasoning chain (predict object → predict attribute given object → correct object given attribute → reconfirm attribute given corrected object → output composition), implemented as structured CoT in an MLLM, trained via SFT on GPT-4o-generated traces followed by GRPO-based RL post-training with step-level exact-match rewards","core_discovery":"The paper's central finding is that bidirectional corrective reasoning — where attribute predictions refine object predictions and vice versa across structured intermediate steps — systematically outperforms both independent primitive prediction and unidirectional conditional modeling for compositional zero-shot recognition. The authors demonstrate this by reformulating CZSL as an open-form generative reasoning task in an MLLM, constraining the model to a five-step chain where Steps 3 and 4 form a mutual verification loop, and showing that step-level reinforcement learning rewards applied to the correction steps yield the largest gains. The error-injection experiment (Table 3) provides the最直","pith_inferences":[],"forward_implications":["If bidirectional correction helps for attribute-object composition, the same mutual-correction loop could apply to other structured prediction tasks where sub-decisions are interdependent, such as relation extraction, scene graph generation, or multi-attribute recognition.","The open-form generative formulation of CZSL removes the need for a predefined candidate label set at inference, which could make compositional recognition more scalable to open-vocabulary settings — but also makes evaluation harder to standardize against prior closed-set work.","The finding that step-level rewards on correction steps (Steps 3–4) outperform uniform or early-step rewards suggests that RL post-training for multi-step reasoning benefits most from rewarding verification behavior rather than initial prediction accuracy.","The error-injection recovery result implies that reasoning-based models can partially self-correct from bad initial conditions, a property absent from static similarity-based methods like CLIP."],"fun_headline_variants":["Attribute and object predictions correct each other step by step","Mutual correction between attributes and objects improves zero-shot recognition","Step-by-step bidirectional reasoning reduces error in compositional learning","Reinforcement rewards on correction steps drive gains in zero-shot recognition","Structured chain-of-thought with mutual verification beats one-pass prediction"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The entire supervised fine-tuning stage depends on GPT-4o-generated five-step chain-of-thought traces being correct and consistent. The paper states that parsing functions verify correctness but provides no systematic evaluation of the quality, faithfulness, or consistency of these auto-generated reasoning traces. If the generated traces contain systematic errors or hallucinated justifications, the model learns to mimic flawed reasoning patterns.","fun_headline_variants_meta":{"raw":{"variants":["Attribute and object predictions correct each other step by step","Mutual correction between attributes and objects improves zero-shot recognition","Step-by-step bidirectional reasoning reduces error in compositional learning","Reinforcement rewards on correction steps drive gains in zero-shot recognition","Structured chain-of-thought with mutual verification beats one-pass prediction"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":609,"prompt_tokens":542,"completion_tokens":67,"prompt_tokens_details":null},"tokens_in":542,"tokens_out":67,"duration_ms":22072,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T20:42:32.233638+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the bidirectional correction steps (Steps 3–4) were removed and performance remained unchanged or improved, or if the error-injection experiment showed no recovery from deliberately incorrect initial predictions, the central claim that mutual correction reduces error propagation would be undermined.","supporting_citations":[],"review_version":1}