acceptodds
Under review as a conference paper at ICLR 2027

PAIReID: Progressive Adaptation and Interactive Reflection for Person Re-Identification

16 views

Abstract

Interactive person re-identification (Re-ID) aims to retrieve a target pedestrian through multi-round interaction, where follow-up questions progressively enrich an initially incomplete description. Existing methods mainly improve the interaction process through supervised question learning, predefined questioning strategies, or LLM-based reasoning, while retrieval typically relies on a shared cross-modal alignment throughout the dialogue. We identify two overlooked challenges. First, as the query evolves from sparse attributes to increasingly fine-grained compositional cues, a fixed alignment may not be equally suitable for different query states. Second, a plausible follow-up question may still be redundant, visually unsupported, or uninformative at the current interaction stage. To address these issues, we propose PAIReID, which combines Progressive Adaptation with Interactive Reflection. On the retrieval side, PAIReID learns lightweight low-rank alignment anchors from progressive query states and dynamically composes them according to the current query semantics, without using explicit interaction-round identities. On the interaction side, a State-Aware Reflective Questioner tracks appearance coverage and selectively refines unsuitable questions without task-specific fine-tuning. Experiments on Interactive-PEDES show that PAIReID achieves 74.08% Rank-1 after five interaction rounds with a zero-shot Questioner and a BRI of 0.485, while evaluations on RSTPReid, CUHK-PEDES, and ICFG-PEDES further demonstrate its applicability to different Text-ReID Retrievers.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Interactive Person Re-identification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_o8EsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies interactive person re-identification, where follow-up questions progressively enrich an initially incomplete description. It proposes PAIReID with two components: a Progressive Query-Aligned Retriever, which learns low-rank text-encoder adaptations from progressive query states and mixes them through a query-conditioned router; and a State-Aware Reflective Questioner, which maintains structured attribute coverage, checks candidate-image visibility, and selectively redirects unsuitable questions without task-specific questioner fine-tuning. Experiments on Interactive-PEDES show promising interactive retrieval performance, while evaluations on three additional person-retrieval benchmarks and component ablations provide further evidence of applicability.

Critical strengths
  1. The framework connects query-dependent retrieval adaptation with explicit tracking of unresolved appearance information. These address complementary aspects of the interaction loop, making the overall design relevant to the task rather than a collection of disconnected modules.

  2. The retriever mixes complete low-rank updates within the text encoder while keeping the visual branch frozen. This permits query-dependent reconfiguration without recomputing gallery embeddings, and avoids the cross-expert terms that would arise from separately mixing the low-rank factors.

Critical weaknesses
  1. The experiments do not sufficiently establish the need for progressive, query-dependent alignment. Table 3 lacks a single adapter trained on pooled progressive queries, and uniform anchor mixing already accounts for much of the reported improvement. Moreover, round-indexed hard selection is compared against semantic soft mixing, changing both the conditioning signal and composition mechanism. Consequently, the results do not clearly distinguish evolving-query adaptation from additional training or adapter aggregation.

  2. The reflective pipeline introduces forced refinement, deterministic validity checks, regeneration, and a canonical fallback. Table 4 does not explain which safeguards remain active without reflection or compare against a rule-based rejection/regeneration baseline. Since candidate-conditioned redundancy filtering already exists in prior work, the current evidence does not establish how much benefit specifically requires the proposed Reflector.

Questions for authors
  1. How does the proposed retriever compare with a single LoRA adapter trained on pooled progressive queries and with a round-conditioned soft mixture using the same anchors? Are the questioning policy and dialogue trajectories held fixed in Table 3 so that retrieval adaptation can be assessed separately?

  2. When reflection is disabled in Table 4, which deterministic safeguards remain active? Can a rule-based validity filter followed by regeneration, using the same generator and comparable inference budget, recover the reported reflection benefit?

  3. Section 3.2 and Appendix A.2 allow partially specified factors to remain eligible for refinement, whereas Appendices A.4–A.5 redirect or prohibit previously queried factors. With an "asked" flag stored at the appearance-group level, it is unclear how an incomplete answer about a bag can be followed by a question about its unresolved properties. This ambiguity concerns a central capability motivating the method. Can an already queried appearance group remain eligible when its answer is only coarse-grained? Please specify the precedence between "asked" and "fact_type" and show a concrete state transition in which an incomplete bag description is subsequently refined.

4 comments
AuthorsOpenReview ↗

Response to the Reviewer o8Es(1/2)

Thank you for your insightful review and valuable feedback. We address your key concerns in detail below and provide additional clarification to further demonstrate the effectiveness of our proposed approach.


[W1, Q1]: Retriever Comparisons under Fixed Dialogue Trajectories

(1) Controlled evaluation. Yes, both the questioning policy and dialogue trajectories are held fixed in Table 3. All retriever variants are evaluated on identical initial descriptions and accumulated question–answer histories at each round, isolating retrieval adaptation from differences in information acquisition.

(2) Additional comparisons. Following the reviewer's suggestion, we evaluate (i) a single LoRA adapter trained on pooled progressive queries from Rounds 0–5 and (ii) a round-conditioned soft mixture using the same six alignment anchors, under the identical fixed-dialogue protocol.

Retriever variantR@1 (%) ↑mAP (%) ↑
Single LoRA on pooled queries72.8650.14
Round-conditioned soft mixture73.4050.23
Query-conditioned semantic mixture (ours)74.0850.85

Under zero-shot questioning, queries at the same round may differ in their semantic composition and granularity. Unlike pooled LoRA, which applies a shared adaptation, or round-conditioned mixtures, which assign shared mixture weights within each round, our router enables query-specific alignment through semantic anchor composition. Our method outperforms these two alternatives by 1.22/0.71 and 0.68/0.62 percentage points in R@1/mAP, respectively. These results support the benefit of conditioning retrieval adaptation on query semantics rather than interaction rounds alone.

[W2, Q2]: Effectiveness of Reflection Beyond Deterministic Safeguards

(1) Clarification of the original ablation. We acknowledge that Table 4 did not sufficiently distinguish reflection from its associated deterministic safeguards. Specifically, Coverage w/o Reflection retains structured memory construction, visual gating, and answer processing, but bypasses both the runtime question-validity guard and generator coverage fallback. Thus, the original ablation evaluates their joint removal rather than reflection alone. We will clarify this distinction in the revised manuscript.

(2) Additional controlled comparisons. Following the reviewer's suggestion, we evaluate two reflection-free alternatives using the same generator: (i) Rule-only, applying deterministic validity checks followed by template fallback; and (ii) Rule + Regeneration, allowing up to four generation attempts under the same validity checks and fallback mechanism.

MethodR@1 (%) ↑mAP (%) ↑BRI ↓
Rule-only68.2548.040.516
Rule + Regeneration68.6048.260.517
Full PAIReID74.0850.850.485

The regeneration baseline averages 3.05 generator calls per query-round, with a nominal maximum output allowance of 512 tokens, comparable in configuration to that of the full method.

Questions satisfying predefined constraints may still provide limited information for subsequent retrieval. Unlike deterministic filtering and regeneration, which enforce rule compliance, our reflection mechanism additionally evaluates contextual relevance, semantic redundancy, and visual support. Although regeneration reduces the fallback rate from 73.44% to 59.30%, it improves R@1 by only 0.35 percentage points, while our full method outperforms it by 5.48/2.59 percentage points in R@1/mAP. These results support the benefit of contextual semantic verification beyond deterministic validity checks and repeated generation.

AuthorsOpenReview ↗

Response to the Reviewer o8Es(2/2)

Q3: State-Guided Refinement and Query Eligibility

(1) Clarification of eligibility. We thank the reviewer for identifying this ambiguity. Our coverage state distinguishes information granularity (fact_type) from questioning history (asked). Specifically, asked=yes records a prior targeted inquiry rather than complete attribute knowledge. Under the current group-level policy, asked takes precedence over fact_type in determining eligibility for further targeted questioning.

(2) Illustrative state transition. An initial description of "carrying a bag" corresponds to (asked=no, fact_type=coarse), allowing a targeted question such as "What type of bag is the person carrying?". An answer such as "A black backpack" enriches the description and updates the state to (asked=yes, fact_type=fine). Even if the answer remains coarse, the group is no longer eligible for another targeted inquiry. Thus, our method supports refinement of initially coarse descriptions, but not repeated property-level refinement within an already queried group.

Under a limited interaction budget, our coverage state guides the refinement of initially incomplete descriptions while distributing questioning opportunities across appearance groups. This conservative strategy is intended to reduce repeated questioning and broaden appearance coverage. We will revise Section 3.2 and Appendices A.2–A.5 to clarify the eligibility rules and their scope. Repeated property-level refinement remains a possible extension of our current framework.

I thank the authors for for providing these additional experiments and explanations in the rebuttal. My questions about retrieval ablation and reflective questioner compared with rule‑based baselines are mostly solved by new experiment results. The authors also explain how the asked flag and fact_type work in state transition. But the proposed method still has one inherent limitation: it can not do further fine‑grained refinement for an appearance group that has already been queried. Besides, some misleading descriptions of ablation in the original paper need to be revised carefully. Considering these remaining problems, I will keep my original score.

AuthorsOpenReview ↗

Response to the Reviewer o8Es

We sincerely thank the reviewer for recognizing our additional experimental results and clarifications. We particularly appreciate this insightful observation, which raises an important question about the trade-off between within-group refinement and broader appearance coverage.

(1) Practical relevance of within-group refinement. In interactive person Re-ID, we observe that even when models are allowed to revisit previously queried appearance groups, scenarios requiring such revisits appear to be relatively rare. Specifically, we conduct a rule-based analysis of 7,373 five-round dialogues (36,865 questions) from our reproduced LLaVA-ReID baseline, where within-group refinement is permitted. A same-group revisit is defined as a question targeting an appearance group that has already been queried within the same dialogue.

Table R1. Within-dialogue question revisit statistics in LLaVA-ReID.

Metric per QuestionRate (%)
Same-group revisits3.11
1st same-group revisit question3.07
2ed same-group revisit question0.04
Exact or paraphrased repeats0.11

Even without hard group-level exclusion, only 3.11% of questions revisit previously queried appearance groups, with second revisits accounting for merely 0.04% of all questions. These observations support our decision to restrict within-group revisits and prioritize broader appearance coverage, with the aim of improving retrieval accuracy across a wider range of queries under a limited interaction budget.

(2) Empirical effectiveness of our design. Furthermore, despite this restriction, the ablation studies in Table 4 and our additional comparisons with rule-based baselines demonstrate the overall effectiveness of our state-aware reflective questioning mechanism. In particular, our full method outperforms Rule + Regeneration by 5.48/2.59 percentage points in R@1/mAP. Together, these results support the practical value of our current design, while leaving selective within-group refinement as a potential direction for further improvement.

We will incorporate this additional analysis and explicitly discuss the limitation in the revised manuscript. We sincerely hope that these empirical findings and clarifications address the reviewer's remaining concern and further strengthen their confidence in the rationale and effectiveness of our approach.

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

This paper proposes PAIReID, a framework for interactive person re-identification (Re-ID) that aims to improve retrieval performance through progressive query adaptation and state-aware question generation.

The proposed framework consists of two main components:

  1. Progressive Query-Aligned Retriever: The authors introduce six LoRA-based alignment anchors trained on textual queries corresponding to different interaction stages. A lightweight semantic router dynamically combines these anchors based on the current query representation, enabling query-dependent cross-modal alignment without explicitly using interaction-round information
  2. State-Aware Reflective Questioner: The authors introduce a structured coverage state to track acquired and unresolved appearance information. A zero-shot multimodal language model generates follow-up questions, while a separate reflection mechanism evaluates their relevance, redundancy, and visual support, selectively refining unsuitable questions.

Experiments on Interactive-PEDES demonstrate that PAIReID achieves 74.08% Rank-1 accuracy after five interaction rounds, outperforming existing interactive retrieval methods in Rank-1 accuracy and achieving a favorable BRI score. Additional experiments on RSTPReid, CUHK-PEDES, and ICFG-PEDES investigate the applicability of the framework to different text-based person Re-ID retrievers. Ablation studies further examine the effectiveness of semantic routing and reflective questioning.

Critical strengths
  • The paper identifies two relevant challenges in interactive person Re-ID: adapting cross-modal alignment to progressively enriched textual queries and selecting informative follow-up questions based on the current interaction state. These challenges are well motivated, and addressing both retrieval and question generation within a unified framework is a meaningful direction.
  • The Progressive Query-Aligned Retriever provides a practical approach to adapting retrieval representations through multiple LoRA anchors and semantic routing. The design preserves the frozen backbone and shared visual embedding space while allowing query-dependent adaptation. The ablation study in Table 3 shows that semantic routing outperforms both uniform anchor combination and round-indexed selection.
  • The State-Aware Reflective Questioner explicitly distinguishes resolved, partially specified, and unexplored appearance attributes. Combining structured coverage tracking with reflective verification is a reasonable approach to reducing redundant or visually unsupported questions. The use of deterministic safeguards to manage interaction states is also a useful design choice.
  • PAIReID achieves competitive results on Interactive-PEDES, particularly in Rank-1 accuracy and BRI, without task-specific fine-tuning of the Questioner. The experiments on three additional text-based person Re-ID benchmarks and the ablation studies provide further evidence of the framework's effectiveness.
Critical weaknesses
  • The proposed framework mainly combines established techniques, including LoRA-based adaptation, mixture-of-experts-style routing, structured attribute tracking, and LLM-based question reflection. Although their integration into interactive person Re-ID is meaningful, the paper does not sufficiently clarify what is fundamentally novel about the proposed mechanisms compared with existing adaptive retrieval and reflective question-generation approaches. In particular, the semantic router resembles conventional expert-mixture mechanisms, while the reflective questioning component relies largely on prompting and deterministic rules.
  • The main comparison in Table 1 involves methods with different Questioner architectures and training strategies. For example, PAIReID uses Qwen-2.5-VL as its zero-shot Questioner, whereas LLaVA-ReID employs a different model and questioning pipeline. Consequently, the observed improvements may partly reflect differences in the underlying multimodal language models rather than the proposed methodological contributions.
  • Although Table 3 demonstrates that semantic routing outperforms uniform combination and round-indexed selection, the improvement over uniform routing is relatively modest (74.08% versus 73.36% Rank-1). Furthermore, the routing analysis in Appendix A.8 mainly reports average routing weights across interaction rounds. These results do not sufficiently demonstrate that the router responds to differences in semantic composition rather than simply learning a proxy for interaction progress, such as query length or the number of accumulated attributes.
  • Table 4 reports substantial improvements from reflective verification, with Rank-1 accuracy increasing from 63.52% to 74.08%. However, the paper provides limited direct analysis of the reflection decisions themselves. For example, the frequency of KEEP and REFINE decisions, the proportion of genuinely unsuitable questions corrected by reflection, and the quality of regenerated questions are not quantitatively evaluated. Without such analyses, it is difficult to determine whether the improvements result from accurate reflection or simply from introducing additional question-generation steps and constraints.
  • The proposed framework introduces six separately trained LoRA anchors, an additional semantic routing pass, and multiple language-model-based operations for question generation, state extraction, visibility assessment, and reflection. Although the appendix reports training costs for the retrieval components, the end-to-end inference latency and computational overhead of the complete interactive system are not sufficiently analyzed.
Questions for authors

Please see the weaknesses section.

3 comments
AuthorsOpenReview ↗

Response to the Reviewer CEEL(1/3)

Thank you for your insightful review and valuable feedback. We address your key concerns in detail below and provide additional clarification to further demonstrate the effectiveness of our proposed approach.


W1: Novelty and Differences from Existing Methods.

Thanks for raising this concern. We would like to clarify that PAIReID is not a trivial combination of existing techniques. While recent interactive person Re-ID methods have improved how appearance information is acquired and accumulated through dialogue, two problems remain insufficiently addressed: how to adapt cross-modal alignment to progressively evolving queries and how to determine what appearance information should be acquired next. Compared with existing works, the differences of our method are summarized as follows.

(1) Progressive Query-Aligned Retriever. Conventional LoRA learns a fixed adaptation, while existing LoRA-based MoE methods focus on expert composition. Neither explicitly addresses the evolving cross-modal alignment requirements in interactive person Re-ID. We therefore learn alignment anchors from progressively enriched queries and combine them according to the current query representation rather than the interaction round, accounting for variations in query information across samples. Our comparison with learned round-conditioned soft routing further supports this design.

(2) State-Aware Reflective Questioner. Existing interactive retrieval methods mainly focus on question generation, dialogue aggregation, or question filtering, leaving the information state of individual appearance factors less explicitly modeled. We instead track appearance coverage and jointly evaluate each proposed question against the current information state and retrieved visual context, enabling targeted information acquisition without task-specific Questioner fine-tuning.

We believe that these designs provide useful insights into cross-modal alignment and information acquisition under evolving queries.

W2: Comparison under Identical Questioner Backbones.

To disentangle the influence of Questioner backbones from the effectiveness of our framework, we conducted additional comparisons between PAIReID and LLaVA-ReID baseline using two alternative Questioner backbones: DeepSeek Flash and Qwen3.5-Flash. Within each setting, both methods use the same Questioner backbone without task-specific fine-tuning, the same Answerer configuration, and five interaction rounds on the full Interactive-PEDES test set.

QuestionerMethodR@1 (R5)mAP (R5)
DeepSeek FlashLLaVA-ReID46.8234.20
DeepSeek FlashPAIReID74.4650.85
Qwen3.5-FlashLLaVA-ReID42.1731.51
Qwen3.5-FlashPAIReID73.7450.31

Under identical Questioner backbones, PAIReID outperforms the corresponding baselines by 27.64 and 31.57 percentage points in Rank-1 accuracy, respectively. Furthermore, PAIReID achieves consistently larger improvements through interaction (+21.35/+20.63 points) than the corresponding baselines (+8.71/+4.06 points).

These results demonstrate that the system-level advantage of PAIReID persists when controlling for the Questioner backbone and zero-shot task setting, rather than relying on differences between underlying Questioner models. We will incorporate these matched-backbone comparisons and clarify the experimental settings in the revised manuscript.

AuthorsOpenReview ↗

Response to the Reviewer CEEL(2/3)

W3: Effectiveness and Query Dependency of Semantic Routing.

To further investigate the contribution of query-conditioned routing, we provide two additional analyses.

(1) Comparison with learned round-conditioned routing. Beyond uniform and round-indexed routing, we evaluate a learnable round-conditioned soft mixture (Round-soft), which learns six anchor weights for each interaction round. All routing strategies are evaluated on identical dialogue trajectories with the same base retriever, alignment anchors, and gallery.

Routing StrategyR@1 (R3)R@1 (R5)mAP (R5)
Uniform64.4973.3650.26
Round-indexed61.7471.6748.88
Learned Round-soft65.2573.4050.23
Semantic Routing66.9774.0950.85

Compared with uniform routing, semantic routing provides a larger improvement at intermediate interactions (+2.48 points at Round 3). It also consistently outperforms the learned Round-soft alternative at the reported stages, improving Round-3/5 Rank-1 by 1.72/0.69 points and final mAP by 0.62 points. These results support the empirical value of conditioning anchor mixtures on individual queries rather than solely on interaction indices.

(2) Within-round routing variability. To complement the round-wise averages in Appendix A.8, we report the mean within-round variance of the six expert weights across all 7,373 queries.

Routing Variance (×10⁻³)R0R1R2R3R4R5
Learned Round-soft000000
Semantic Routing0.4671.3882.6192.9232.1301.207

Round-soft has zero within-round variance by construction, as all queries at the same round share identical expert weights. In contrast, semantic routing exhibits non-zero variability within every round. Together, these analyses support the effectiveness of query-conditioned soft routing beyond the tested round-only alternatives.

W4: Decision Analysis and Effectiveness of Reflective Verification.

To better understand the contribution of reflective verification beyond downstream retrieval performance, we conducted a comprehensive analysis of all 36,865 query–round records from the main experiment.

(1) Reflection decisions. The original Reflector outputs are KEEP in 31.21% of cases and REFINE in 68.79%. After deterministic schema and coverage corrections, the executed proportions become 23.68% and 76.32%, respectively.

(2) Question refinement quality. We evaluated initial and regenerated questions against the same pre-refinement coverage state using consistent format and coverage criteria.

Metric (%)Before REFINERaw RegenerationFinal Question
Validity check pass rate ↑40.4167.1584.35
Exact historical repetition ↓26.672.550.00

Among 16,765 initial questions with detectable violations, raw regeneration resolves all checked violations in 65.33% of cases, increasing to 79.02% after the complete verification pipeline.

(3) Beyond additional generation. To investigate whether the observed gains simply arise from additional generation steps and constraints, we further compare reflective verification with rule-based filtering and regeneration.

Verification StrategyR@1mAP
Rule-based Filtering68.2548.04
Rule-based Filtering + Regeneration68.6048.26
Full PAIReID74.0850.85

Additional rule-based regeneration yields only a 0.35-point Rank-1 improvement, whereas the full framework achieves substantially higher performance. Although differences in prompting and deterministic processing preclude a strictly isolated causal comparison, the tested regeneration control does not reproduce the gains of our reflective verification pipeline.

These analyses provide complementary question-level and retrieval-level evidence supporting the effectiveness of reflective verification. We will incorporate these quantitative analyses and clarify their scope in the revised manuscript.

AuthorsOpenReview ↗

Response to the Reviewer CEEL(3/3)

W5: Computational Overhead and Inference Efficiency.

Thank you for highlighting this important practical concern. To quantify the computational overhead of PAIReID, we analyzed the complete main experiment on Interactive-PEDES, covering 7,373 queries and five interaction rounds.

(1) Retriever overhead. The six frozen LoRA anchors introduce 2.95M parameters, while the semantic router adds 0.134M parameters, corresponding to approximately 11.76 MiB of additional FP32 weights. Semantic routing requires a base text-encoding pass followed by a routed pass involving all six anchors. Gallery features are extracted once and reused throughout the interaction.

(2) End-to-end computational analysis. The complete evaluation takes 25.16 hours with a nominal batch size of 40, corresponding to approximately 293 completed five-round queries per hour. We further analyze the component-wise latency across 925 batch–round records.

ComponentMean (s/batch-round)P95 (s/batch-round)
Memory Compression22.3731.31
Visual Feedback33.4135.45
Question Generation7.258.71
Reflector23.4126.34
Conditional Regeneration7.589.16
Answerer3.364.26
Overall Policy Pipeline97.44106.64

The policy pipeline accounts for 99.49% of the total evaluation time, with reflection and conditional regeneration accounting for 31.80% of policy time. The six instrumented modules require an average of 5.76 logical model calls per query–round, excluding conditional summarization. These results identify LLM/MLLM-based operations, rather than retrieval alone, as the primary computational bottleneck.

(3) Performance–efficiency trade-off.

Interaction RoundCumulative Policy Time (s/batch)R@1mAP
0—53.1138.91
195.5657.0941.30
2185.4862.2944.27
3283.0666.9347.04
4383.9871.0849.38
5487.2174.0850.85

Despite the additional computational cost, progressive interaction consistently improves retrieval performance, yielding a 20.97-point Rank-1 improvement after five rounds. The intermediate results illustrate the potential for balancing retrieval accuracy against different interaction budgets.

Reviewer_mRmfOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the PAIReID framework for the problem of interactive pedestrian re-identification. The problem is formulated as follows: through multi-round conversational interaction, gradually acquire appearance attribute information about the target pedestrian, thereby retrieving the target individual from a pedestrian database. Existing methods primarily focus on improving the interaction process (e.g., via supervised problem learning, predefined query strategies, or reasoning based on large language models); however, at the retrieval stage, they typically employ a shared cross-modal alignment function throughout the entire conversation. The authors identify two overlooked challenges: (1) as queries evolve from sparse attributes to increasingly granular composite clues, fixed cross-modal alignment may no longer be suitable for different query states; and (2) a seemingly reasonable subsequent query may remain redundant, visually unsupported, or information-deficient given the current interaction state. To address these issues, PAIReID comprises two core components: (a) a progressive query alignment retriever, which learns multiple lightweight LoRA alignment anchors and dynamically combines these anchors based on the current query semantics via a semantic router, without relying on an explicit interaction-round index; and (b) a state-aware reflective questioner, which maintains an attribute-level coverage state to track resolved and unresolved attribute information, and utilizes a Reflector to make KEEP/REFINE decisions for generated queries without requiring task-specific fine-tuning. In terms of experimental evaluation, the authors conducted primary assessments on the Interactive-PEDES benchmark, while also validating the method's transferability across three traditional text-based pedestrian re-identification benchmarks: RSTPReid, CUHK-PEDES, and ICFG-PEDES. On Interactive-PEDES, PAIReID achieved Round 3 R@1 = 66.93% (surpassing the fine-tuned version of LLaVA-ReID, which scored 63.96%) and Round 5 R@1 = 74.08% using zero-shot Questioner, while also attaining the best BRI = 0.485. Ablation experiments demonstrated the effectiveness of the semantic routing strategy and the reflection-based validation component.

Critical strengths
  1. The authors accurately identify two key yet overlooked challenges in interactive pedestrian re-identification: the issue of alignment mismatch caused by query semantic evolution, and the lack of problem utility evaluation during the interaction process. These two challenges are introduced through a logical chain that naturally arises from the limitations of existing methods, providing a sound and persuasive motivation. (2) The design of the incremental query alignment retriever is particularly outstanding—by dynamically combining alignment anchors via a semantic router rather than a round-based index, the model can adaptively adjust its alignment strategy based on the semantic content of the query, rather than simply switching based on the interaction round. This design avoids the generalization limitations associated with round-based indexing and offers strong theoretical elegance. Furthermore, the method only adjusts LoRA parameters on the text encoder side while keeping the backbone network frozen, resulting in high parameter efficiency. (3) The architectural design that separates semantic reasoning (LLM) from deterministic state management is well-founded: the LLM handles semantic understanding and visual fact extraction, while the programmatically maintained state variables (query state, fact granularity, and retrieval priority) ensure the controllability and auditability of the interaction strategy. The two-stage verification mechanism of KEEP/REFINE effectively reduces the generation of invalid queries. (4) The authors conducted comprehensive comparative experiments on the Interactive-PEDES main benchmark, covering baseline methods with different queryer types (heuristic, few-shot, zero-shot, and fine-tuned). Additionally, the method's transferability was validated across three traditional Text-ReID benchmarks. The ablation study covers key design choices for routing strategies (uniform combination / round-indexed routing / semantic routing) and questioner components (covering modeling / reflective validation). (5) The paper is clearly written. Its structure is well organized, the methods are described in detail, and the definitions of formulas are clear. The appendix provides comprehensive implementation details (hyperparameters, training configurations, prompt templates, etc.), facilitating reproducibility. The disclosure of the ethics statement and AI usage statement also complies with conference guidelines.
Critical weaknesses
  1. The paper claims to use Qwen-2.5-VL as a "zero-shot Questioner"; however, in practice, this approach relies heavily on carefully designed prompt engineering, which includes multiple prompt templates covering state extraction, visual feedback evaluation, question generation, and reflective verification. The term "zero-shot" can mislead readers into believing that the method functions without any prompt engineering; it is therefore recommended that the approach be more accurately described as "prompt-based" or "template-based" questioner, or that the complexity of the prompt design and its contribution to performance be explicitly stated in the paper. (2) The comparison with existing methods on Interactive-PEDES covers only four approaches: SimIRV, ChatIR, PlugIR, and LLaVA-ReID. Given the rapid development of the interactive image retrieval field in recent years, the paper lacks comparisons with the following works: (a) other LLM-based interactive retrieval methods; (b) the latest methods for multimodal dialogue retrieval; and (c) related work on active query strategies. Furthermore, while comparing the fine-tuned version against LLaVA-ReID is meaningful, the paper lacks comparisons with other methods that may similarly employ LLMs but adopt different strategies. (3) The paper provides no analysis of retrieval failure cases. Error analysis can help readers understand the scenarios in which the method performs poorly (e.g., for specific attribute types, specific interaction rounds, or specific query complexities), which is highly valuable for assessing the method's practical limitations and guiding future improvements. It is recommended that case studies or discussions of failure modes be added. (4) In the paper, the Answerer uses pre-labeled dialogue annotations to simulate witness feedback; this is essentially an offline evaluation setting. In real-world scenarios, human user responses may contain noise, be incomplete, or be biased. The paper does not discuss the gap between this simulated setting and actual deployment, nor does it analyze the impact of the Answerer's simulation errors on overall system performance. It is recommended that the paper include a discussion or experiments on the Answerer's noise robustness. (5) The scope of the evaluation dataset is limited. Experiments were primarily conducted on Interactive-PEDES, which is derived from CUHK-PEDES and ICFG-PEDES. Although the authors performed transfer experiments on three traditional Text-ReID benchmarks, they lack validation on other independent interactive retrieval datasets (such as COCO Interactive and Flickr30k Interactive); therefore, the method's generalization capability requires further verification.
Questions for authors

Q1: The training of the semantic router relies entirely on the retrieval loss function (LSDM + ID + MLM), without introducing any router-specific auxiliary objectives (such as router entropy regularization or expert load balancing). Has the author attempted to incorporate these auxiliary objectives? If not, could you discuss why retrieval supervision signals are sufficient for effectively training the router? How does the router's initialization strategy (zero-initialized output layer) affect the convergence speed? Q2: The paper uses six alignment anchors to cover Rounds 0 through Round 5. In practical applications, the number of interaction rounds may exceed five. How does the method perform over more rounds (e.g., Rounds 6–10)? Should additional anchors be considered, or should a continuous anchor parameterization approach be adopted? Q3: The performance of the "Zero-shot Questioner" depends heavily on the capabilities of Qwen-2.5-VL. Has the author experimented with different base models (such as GPT-4V or Gemini)? To what extent does the choice of model affect the final performance? Additionally, has the design of the prompt template undergone sensitivity analysis? Q4: In the reflective verification mechanism, the Reflector employs a binary classification decision between KEEP and REFINE. Could the re-generation after REFINE potentially fall into a cycle (i.e., repeatedly applying REFINE without generating a satisfactory question)? Does the paper define any termination conditions, such as a maximum number of re-generation attempts? Q5: In the actual deployment of interactive retrieval, user responses may contain noise or be inaccurate. How robust is Coverge State's update mechanism with respect to Answerer noise? Has noise robustness analysis been conducted?

2 comments
AuthorsOpenReview ↗

Response to the Reviewer mRmf(1/3)

Thank you for your insightful review and valuable feedback. We address your key concerns in detail below and provide additional clarification to further demonstrate the effectiveness of our proposed approach.


Q1(a): Necessity of Router-specific Auxiliary Objectives.

We conducted additional controlled experiments with entropy regularization and soft load balancing (λ = 0.01), keeping the alignment anchors, training configurations, and dialogue trajectories identical. Results are reported as mean ± std over three random seeds.

Router ObjectiveR@1mAPNorm. EntropyImbalance
Retrieval-only74.28 ± 0.0950.91 ± 0.010.5860.275
+ Entropy74.27 ± 0.0950.90 ± 0.010.6750.204
+ Load Balancing74.28 ± 0.0950.90 ± 0.010.6340.232

Although both auxiliary objectives encourage more uniform routing distributions as intended, neither yields additional retrieval gains. This observation is consistent with two aspects of our design: (1) the pre-trained alignment anchors provide complementary alignment capabilities, allowing retrieval losses to directly supervise their query-dependent combination; (2) unlike sparse MoE, our dense soft routing does not require balanced expert utilization for computational efficiency, and heterogeneous query semantics may naturally favor unequal anchor contributions.

These results empirically support the sufficiency of retrieval-only supervision under the evaluated setting. We will include this ablation and further clarify our design rationale in the revised manuscript.

Q1(b): Router Initialization and Convergence.

To investigate the effect of zero initialization, we conducted a controlled comparison with Xavier initialization of the router's output layer, using identical training protocols and three paired random seeds.

InitializationInitial EntropyProbe R@1 AUCFinal R@1Final mAP
Zero1.000091.380974.28 ± 0.0850.91 ± 0.01
Xavier0.944691.435874.23 ± 0.0450.88 ± 0.02

Zero initialization is intended to provide a uniform initial mixture of alignment anchors rather than explicitly accelerate convergence. Although it initially blocks task gradients to the preceding hidden layer, our gradient analysis confirms that the output layer receives non-zero gradients at the first backward pass, and hidden-layer task gradients become non-zero at the second step. The convergence trajectories on a fixed training probe vary across seeds, with Xavier achieving a slightly higher step-averaged probe R@1, while zero initialization yields comparable final test performance with marginally higher mean retrieval accuracy. We will include these results and clarify the rationale for our initialization strategy in the revised manuscript.

Q2: Extensibility beyond Five Interaction Rounds.

We extended all 7,373 test conversations on Interactive-PEDES to ten interaction rounds, retaining the original questioning policy and six alignment anchors, without any additional training.

Evaluation StageR@1R@5R@10mAP
Round 574.0891.4895.3950.85
R6–10 (Avg.)75.6292.2696.0451.77
R6–10 (Best)75.7592.3996.1851.88
Round 1075.5792.1695.9951.71

Despite using only six anchors trained on queries from Rounds 0–5, PAIReID maintains improvements across all extended rounds, achieving up to 1.67 and 1.03 percentage points of additional gains in Rank-1 and mAP, respectively. Importantly, the number of alignment anchors does not determine the maximum interaction horizon. Rather than assigning a dedicated anchor to each round, our semantic router constructs query-dependent alignment functions through continuous combinations of existing anchors. This design enables retrieval adaptation beyond the interaction depths covered during training. The results empirically demonstrate that additional anchors are not required to extend PAIReID to ten rounds under the evaluated setting. Increasing the number of anchors or learning a continuous anchor parameterization could further enrich alignment flexibility, which we leave for future investigation.

AuthorsOpenReview ↗

Response to the Reviewer mRmf(2/3)

Q3(a): Generalization across Questioner Backbones.

To investigate the dependence of PAIReID on Qwen2.5-VL-7B, we conducted additional experiments with DeepSeek Flash, Qwen3-8B, and Qwen3.5-Flash on Interactive-PEDES.

QuestionerR@1 (R1)R@1 (R2)R@1 (R3)R@1 (R4)R@1 (R5)mAP (R5)
Qwen2.5-VL-7B57.0962.2966.9371.0874.0850.85
DeepSeek Flash59.4564.0068.5672.6474.4650.85
Qwen3-8B55.5960.7465.9670.8173.6749.97
Qwen3.5-Flash57.5162.3867.4570.9973.7450.31

All four configurations consistently improve retrieval performance across interaction rounds, achieving 20.56–21.35 percentage points of Rank-1 improvement over the same initial query baseline. Despite variations in intermediate performance, their final Rank-1 accuracies differ by only 0.79 percentage points. Notably, DeepSeek Flash consistently outperforms Qwen2.5-VL-7B throughout all five interaction rounds.

This robustness is consistent with our design, where structured coverage tracking and reflective verification guide question generation without requiring backbone-specific fine-tuning. These results demonstrate that PAIReID can effectively accommodate diverse Questioner backbones, supporting the generalizability of our framework beyond Qwen2.5-VL-7B.

Q3(b): Sensitivity to Prompt Templates.

We conducted a prompt sensitivity analysis on 1,000 randomly sampled Interactive-PEDES test queries with the full gallery. Using Qwen3.5-Flash as the Questioner, we compared three template configurations: (1) Original, (2) Paraphrased instructions, and (3) Reformatted and reordered instructions. The templates for generation, reflection, and refinement were jointly modified while preserving the task constraints, output schemas, and all other system components.

Prompt TemplateR@1 (R5)ΔR@1 (R5−R0)mAP (R5)
Original73.90+23.3048.92
Paraphrased71.80+21.2048.69
Reformatted72.90+22.3048.54

All three templates maintain consistent progressive retrieval improvements, achieving Rank-1 gains of 21.20–23.30 percentage points after five interactions. While the final mAP varies by only 0.38 percentage points, paraphrasing introduces a 2.10-point decrease in R@1 (paired bootstrap 95% CI: [−4.00, −0.39]), indicating some sensitivity in top-1 retrieval.

These results suggest that although prompt formulation affects question generation and retrieval performance, the progressive retrieval benefits of PAIReID are preserved across the tested template variations and are not restricted to the original prompt formulation.

Q4: Termination of Reflective Verification.

We clarify that our reflective verification adopts a bounded, single-step refinement strategy rather than an iterative refinement loop. Specifically, a KEEP decision retains the initial question, whereas a REFINE decision triggers at most one re-generation guided by the Reflector's feedback. The revised question is not submitted to the Reflector again.

If the revised question fails subsequent deterministic validity checks (e.g., repeated inquiries or attribute-group inconsistency), a rule-based fallback question is used instead of further re-generation. Therefore, termination is guaranteed by design, with at most two question-generation attempts per interaction round (one initial generation and one optional refinement), independent of whether the revised question would receive a KEEP decision.

Comments on the paper

1 comment
Reviewer_mRmfOpenReview ↗

Respond to author

I thank the authors for their detailed response and for revising the manuscript. Several of my concerns have been addressed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.