- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes ROOMER, a multi-turn vision-language agent for room re-identification. Instead of a fixed coarse-to-fine pipeline such as AirRoom, a learned Router reads a shortlist from a frozen global retriever and decides at each turn whether to commit to a candidate, run a geometric verifier, or semantically inspect a candidate with an Inspector. The Inspector is trained with Sight-OPSD, an on-policy self-distillation scheme in which a teacher sees training-only regional and geometric views while the student sees only the raw image pair. The Router is a LoRA adapter on the frozen Inspector backbone, trained with grouped multi-turn RL that combines trajectory-level Outcome Credit with localized Commitment and Probe Credits, which supervise when to stop and which candidate to inspect. The authors also introduce AirRoom-Contested, a diagnostic subset of ambiguous queries. ROOMER improves over strong fixed pipelines on both the contested subset and the natural distribution while using less runtime, and the distilled 4B Inspector outperforms a much larger general-purpose model on pair verification.
Critical strengths
- The method improves over strong baselines, including AirRoom+SLiM, on both the known benchmarks and on the ambiguous queries subset. A distilled 4B Inspector outperforming a general-purpose 397B model is an impressive result.
- The available ablations (Tables 3 and 4) are informative and appreciated, and the routing behavior table (Table 5) gives useful insight into how the Router allocates evidence.
Critical weaknesses
- Efficiency claims are not carefully supported.
- The motivating claim that later noisy evidence overturns correct decisions in fixed pipelines is not quantified (see Question 3).
- Table 5 does not state the hardware or the inference engine (optimized or plain PyTorch). The compared systems also spend their time on different kinds of computation (autoregressive VLM generation versus feature extraction and matching), so total latency is difficult to interpret without a per-component breakdown (direct, geometry, inspection) for ROOMER and for AirRoom with and without SLiM. The "38.8% less tool budget" is expressed in hand-set abstract cost units (Appendix E) that are not defined in the main text and are not wall-clock time.
- Probe Credit relies on hand-designed components that are not ablated. Eq. 7 introduces a lexicographic ordering with a verdict tier, and Eq. 8 defines displacement as the sum of a rank change and a tier change. Under this ordering a positive verdict changes both quantities, so the signal is partly double counted. The paper gives one sentence of justification. Table 4 ablates the credits as a whole but not these internal choices.
Questions for authors
- Why is the Router built on the Stage-I Inspector backbone rather than a separate model? Does the verification-oriented Stage-I training help or hurt routing, for example compared with a Router on the unadapted Qwen3.5-4B?
- Was RL training stopped at only 120 steps because performance had converged, or does it saturate or degrade with longer training?
- The paper motivates adaptive routing by noting that later, noisier evidence can overturn correct earlier decisions in fixed pipelines. Could the authors quantify this, for example the fraction of queries where AirRoom's Local or Fine-Grained stage changes a correct earlier top-1 into an incorrect one, and the corresponding rate for ROOMER (final answer versus the retrieval and geometric leaders)?
- In Table 1, the natural-distribution AnyLoc row matches the values reported in the original AirRoom paper (its Accuracy column) on all four datasets, whereas other baselines (e.g., AirRoom Full, DINOv2) differ. It is also unclear why AirRoom Global differs substantially from AnyLoc on the natural distribution (e.g., Gibson 93.45 vs. 85.85) while the two nearly coincide on AirRoom-Contested. Could the authors clarify whether all baselines were re-evaluated?
- The abstract uses several terms before defining them (Router, Sight-OPSD, evidence displacement, "afraid-to-commit collapse"). Could the authors briefly define or describe each term on first use in the abstract?
- The abstract reports gains over AirRoom Full and over a 397B general-purpose model. Could the authors also report the gains over the strongest task-specific baselines in Table 1 (AirRoom+SLiM, SAGE)?
- Could the authors define T_i, K_i and β in Eqs. 4 to 6?
- In Eq. 8, tier is written as a function of c* while the rank terms are written as ρ. Could the authors clarify and unify this notation?
- Could the authors state in the main text that each candidate is inspected at most once and geometry is called at most once, and clarify what "multi-turn" means under this constraint?
- Could the authors use a consistent capitalization of ROOMER across sections?
- Could the authors show qualitative examples of how the teacher's token-level predictions on the student's responses change once the regional or geometric view is provided?
3 comments
Q5, Q6 & Q10: Abstract Clarity, Baseline Comparisons, and Naming Consistency
We have revised the abstract to define Sight-OPSD, Inspector, and Router at first mention (Q5). We also explicitly report ROOMER's improvements over the strongest task-specific baselines. Finally, we have standardized the method name as ROOMER throughout the paper (Q10). Please see the revised Abstract (p. 1).
Q7: Definitions of , , and
We have clarified these notations in Sec. 3.3 (Eqs. 4–6). denotes the terminal decision turn, specifies the turns receiving Commitment Credit (terminal commitment or unnecessary subsequent tool calls), and controls its strength. These clarifications do not change the underlying method.
Q8: Notation Consistency in Eq. (8)
We have unified the notation in Sec. 3.3 (Eq. 8) by explicitly writing . Both rank and tier displacement are evaluated for the same ground-truth candidate . This is a notational clarification without changing the computation.
Weakness-2: Probe Credit Design and Ablation
We agree that rank and tier displacement are coupled, as tier participates in the lexicographic ranking. However, they provide complementary supervision: rank captures relative-order improvements and penalizes harmful promotions of incorrect candidates, while tier rewards direct positive confirmation even when the ground-truth rank remains unchanged. Their combination is an intentional reward-shaping design, rather than two independent information gains.
To assess the additional value of tier, we added a Rank-only Probe Credit ablation, retaining the tier-aware ranking in Eq. (7) while removing the explicit tier-displacement term in Eq. (8). This isolates the incremental value of the explicit confirmation bonus. Removing tier from both equations would yield zero inspection displacement, since semantic inspection changes neither geometric evidence nor retrieval similarity, leaving candidate ordering unchanged. Such a setting would therefore provide no meaningful inspection supervision.
| Method | R@1 ↑ | Budget ↓ | Inspect ↓ |
|---|---|---|---|
| Outcome-only GRPO | 87.20 | 4.10 | 42.0% |
| + Rank-only Probe Credit | 87.16 | 3.52 | 31.4% |
| + Probe Credit (Rank+Tier) | 87.43 | 3.70 | 34.0% |
Rank-only delivers substantial efficiency gains. Adding tier further improves R@1 from 87.16% to 87.43%, with a modest increase in budget (3.52→3.70) and inspection rate (31.4%→34.0%). This is consistent with tier encouraging additional semantic confirmation beyond immediate ranking gains, yielding an accuracy–cost trade-off.
We have clarified the design in Sec. 3.3 (Eqs. 7–9) and added the ablation and discussion in Sec. 4.4 (Table 4).
Q9: Clarification of Multi-Turn Routing
Multi-turn refers to sequential, state-dependent decisions, not repeated verification of the same candidate. After each tool result, the Router updates its evidence state and chooses whether to inspect another candidate or commit. Appendix G.2 (Table 16) shows a four-turn trajectory and we have clarified this in Sec. 3.3, with tool-call constraints detailed in Appendix C.3.
Q11: Example for Token-Level Distillation
We have added a controlled fixed-prefix diagnostic in Appendix B.4 (Figure 4). Using identical frozen 4B weights, prompts, and a student-generated prefix, privileged regional and geometric contexts reduce the probability of the student's erroneous next token (different) from 20.62% to 10.23% and 3.01%, respectively. Across 269 response positions, explanation-content tokens account for 70.9% and 67.7% of the measured KL divergence. This demonstrates that privileged visual context substantially reshapes token-level predictions, supporting our motivation for token-level distillation.
I thank the authors for the detailed response and the additions, which address most of my concerns. In particular, the transition analysis (Table 9) quantifies how often fixed pipelines overturn correct decisions, the AnyLoc entry in Table 1 has been corrected, and the abstract and notation have been clarified. A few concerns are only partially addressed. The runtime comparison still lacks a per-component breakdown (direct, geometry and inspection for ROOMER, and the stages of AirRoom with and without SLiM), and the abstract still reports the 38.8% budget reduction in hand-set cost units defined only in Appendix E. For Probe Credit, the new rank-only ablation shows that the tier term adds only 0.27 points of R@1 while increasing the budget (87.43% vs. 87.16%, 3.70 vs. 3.52), so the benefit of this design choice remains small. I appreciate that the authors report this trade-off openly. Finally, the Router trained from the unadapted Qwen3.5-4B is described as comparable but is not reported quantitatively.
Overall, I keep my positive assessment and maintain the score of 3.
We sincerely thank the reviewer for the thoughtful and constructive feedback. Your comments have been very helpful in improving the clarity, rigor, and completeness of our manuscript. We have carefully addressed the concerns and highlighted the corresponding revisions in blue in the updated manuscript for ease of review.
Q1: Shared Inspector Backbone for Routing
Our primary motivation is parameter efficiency, not an assumed transfer benefit from verification training. We reuse the frozen 4B Inspector backbone and train only a lightweight LoRA adapter for routing; removing the adapter preserves the original verification capability. We still conducted a new experiment: training the router starting from a completely new initialization of qwen3.5-4B + LoRA, and observed comparable performance, suggesting that Stage-I adaptation is not essential for initial routing competence.
Q2: Justification for the 120-Step Checkpoint
We selected the step-120 checkpoint based on held-out development performance. Dev R@1 remains stable around 80% through step 120 but drops to approximately 76.5% at step 140. Meanwhile, declining policy entropy and group-advantage magnitude indicate reduced exploration and weaker optimization signals. Thus, step 120 balances training stability and generalization while avoiding late-stage performance degradation.
Q3: Over-Processing and Runtime Efficiency
We have added a paired transition analysis in Sec. 4.5 and Appendix D.3 (Table 9). AirRoom's Global-to-Local and Local-to-Fine-Grained stages overturn 294 (7.74%) and 380 (9.40%) previously correct predictions, respectively. ROOMER correctly resolves 279/294 (94.90%) and 278/380 (73.16%) of these cases through adaptive stopping and selective verification. On the full AirRoom-Contested evaluation set, ROOMER achieves a lower average wall-clock latency of 0.543 s/query, compared with 0.676 s/query for AirRoom Full, measured on NVIDIA A100 80GB GPUs. Despite the autoregressive overhead of Router generation, ROOMER invokes semantic inspection for only 18.8% of contested queries, while 12.0% are resolved directly and 69.1% stop after geometric verification. This demonstrates that adaptive evidence acquisition and timely stopping can offset the generation overhead. We will release the wall-clock benchmarking code upon acceptance to facilitate reproducibility.
Q4: Baseline Evaluation Consistency
We thank the reviewer for pointing out this issue. Upon re-examination, we found that the AnyLoc results on the natural distribution were inadvertently retained from the original AirRoom paper, rather than reflecting a separate evaluation in our submission. In our implementation, AirRoom Global and AnyLoc use the same global retrieval method. We have therefore removed the redundant AnyLoc entry and merged the two into a single AirRoom Global (AnyLoc) entry in the revised Table 1, reporting the results obtained under our evaluation protocol.
We clarify that all reported methods were evaluated under the same scene-disjoint protocol, with identical test queries and galleries within each setting, as detailed in Appendix A. We have also carefully rechecked the remaining baseline results and confirmed their correctness and consistency with our evaluation protocol. Upon acceptance, we will release the model weights and complete evaluation protocol to facilitate independent reproduction and verification by the research community.