REVIEW 5 major objections 6 minor 16 references
Divide-and-Conquer: Tree-structured Strategy with Answer Distribution Estimator for Goal-Oriented Visual Dialogue
T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash
Pith's one-line read This paper shows that rewarding questions which split remaining candidate objects into two roughly equal halves lets goal-oriented visual dialogue agents find the target in fewer rounds and with fewer repeated questions.
desk verdict A clean reward idea with real repetition-rate gains, but the 'fewer rounds' claim is undercut by the paper's own Table 2 and needs fixing. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the pair of reward functions in equations (1) and (2). The binary reward $r_b = \mathbb{E}\left[\sum_{j=1}^{J_{\mathrm{end}}}\left(1-\frac{|l_j - k_j/2|}{k_j/2}\right)\right]$ scores each round by how close the larger answer group is to exactly half of the current $k_j$ candidates; a perfect halving scores 1 and a unanimous answer scores 0. The candidate-minimization reward $r_c = \alpha + \beta \left(1-\frac{k_{j_{\mathrm{end}}}-1}{N-1}\right)$ when the Guesser picks the target and 0 otherwise, rewarding successful dialogues that shrink the candidate set to a single object. The supporting mechanism is the Answer Distribution Estimator (ADE): an internal Oracle that answers a proposed question for every object in the candidate set, giving the distribution $a^j_{1:k}$ used both to update the candidate set (keep only objects whose answer matches the target's) and to evaluate $r_b$ and $r_c$. This machinery turns question generation into a learned binary-search policy without changing the Oracle or Guesser at inference time.
What would settle it
Collect every round of dialogue produced by a TSADE-trained agent on the GuessWhat?! test split, compute $l_j/k_j$ for each round, and plot the distribution. If most rounds fall below 0.25 or above 0.75, the binary reward is almost always near zero and the claimed halving behavior is not actually occurring, meaning any accuracy gain would come from the candidate-minimization reward or chance rather than from divide-and-conquer.
Extended reading notes
Core claim
The paper's central claim is that a non-goal-oriented questioning strategy—each question should divide the current candidate set into two near-equal groups, and the group containing the target is kept—is a better inductive bias for goal-oriented visual dialogue than the conventional goal-oriented strategy of asking questions that increase the target's probability. What makes this workable is an Answer Distribution Estimator that simulates the Oracle on every candidate object, producing an answer distribution that can be filtered to update the candidate set. The authors argue that because the binary reward $r_b$ measures only the cleanliness of the split and is agnostic to which object is the target, the Questioner learns to generate informative, diverse questions rather than repetitive confirmations. They report that adding this reward and the candidate-minimization reward $r_c$ to existing models (FS, ADVSE) on GuessWhat?! reduces question repetition by up to about 40 percentage points and improves task success rate, and that the same scheme transfers to VisDial's GuessWhich task.
Load-bearing premise
The strategy collapses if real yes/no questions cannot split arbitrary candidate sets into near-balanced groups, or if the Oracle misanswers enough objects that the filtered set loses the target.
Editorial extensions
If this is right
- Any question generator trained with these rewards should ask questions that divide the candidate set, not just questions that confirm a single object, so the question repetition rate drops accordingly.
- The search over candidate objects becomes logarithmic in the number of objects, which means the same task can be completed within tighter round budgets such as the 5-round limit used in the experiments.
- The candidate-minimization reward should raise success rate beyond what a plain 0-1 reward gives, because it explicitly prefers dialogues that end with one candidate rather than a lucky guess.
- The reward scheme is a plugin that can be applied to different base models without retraining the Oracle or Guesser, so it can improve any new QGen that appears in the future.
Reading between the lines
- Beyond the paper, the same divide-and-conquer reward could be applied to any sequential decision task with a discrete candidate set and binary feedback, such as object goal navigation or entity disambiguation, where each step splits the remaining options.
- A testable prediction follows: if the binary reward is the active ingredient, then restricting the question vocabulary to attributes that naturally split objects (color, material, location) should improve efficiency more than simply adding more training data.
- The candidate-minimization reward may concentrate gains on images with few objects; a per-image stratification of the results would show whether the improvement is uniform or limited to easy cases.
- The paper's comparison to a large vision-language model suggests that this reward shaping could be used to improve the questioning policy of modern multimodal agents, not just the small LSTM-based models evaluated here.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper proposes TSADE, a reinforcement-learning reward scheme for training the Question Generator in goal-oriented visual dialogue. TSADE maintains a candidate object set through an internal Answer Distribution Estimator that answers the current question for all candidate objects; a binary reward encourages questions that split the candidate set into two balanced groups, and a candidate-minimization reward encourages successful dialogues that reduce the candidate set to a single object. The QGen is trained with policy gradient while Oracle and Guesser are kept fixed. Experiments on GuessWhat?! and VisDial report lower question repetition rates, modest success-rate improvements, and a lower T/R efficiency coefficient relative to several baselines, plus an LVLM-based sanity check.
Significance. If the reported results are reliable, TSADE is a simple and reusable training signal: it does not change the Oracle, Guesser, or decoding procedure, and it is shown to work with DV, FS, and ADVSE on GuessWhat?! and with ReeQ on VisDial. The released code and the inclusion of an LVLM experiment are useful. However, the paper's headline efficiency claim (fewer rounds) is not actually supported by Table 2, and the empirical evaluation lacks variance estimates and the most closely related ADE-based baselines. The central methodological idea is coherent, but the evidence as presented is insufficient to establish the claimed advantage over existing information-gain strategies.
major comments (5)
- [§Comparison on Efficiency, Table 2] The abstract promises that TSADE finds the target 'in fewer ... rounds', but Table 2 shows the opposite for the reported statistic T: in every setting, T increases (e.g., DV 3.35→3.62 and FS 3.67→3.86 at Jmax=5). The paper then declares TSADE more efficient because T/R decreases. T/R is not an expected number of rounds; it is a ratio of a conditional mean to a probability, and a decrease in T/R can occur even when dialogues are longer on average. If one approximates expected capped dialogue length by R·T + (1−R)·Jmax, the values at Jmax=5 are ≈4.48 for FS and ≈4.51 for FS+TSADE, so TSADE is not faster under this natural metric. Please report the full distribution of dialogue lengths or an explicitly defined expected-cost metric, and revise the 'fewer rounds' claim accordingly.
- [§Implementation Details and Tables 1–6] All reported numbers appear to be from a single run; no error bars, number of seeds, or significance tests are given. The absolute success-rate gains of TSADE over ADVSE in Table 3 are 0.68 and 1.72 percentage points, which are within the range of typical seed-to-seed variance for this class of models, and the VisDial gains in Table 6 are uneven (e.g., R@5 drops from 48.50 to 48.11 with ro). In addition, α, β, γ are tuned by grid search on the same benchmark used for the main results; please provide a nested validation protocol or at least report sensitivity of the headline results to these hyperparameters.
- [§Related Work and Table 3] AQM (Lee et al. 2018) and confirm-it (Testoni and Bernardi 2021) are the closest prior uses of answer-distribution estimation for question selection, and both are discussed in the text, but neither appears in the comparison tables. Because TSADE's claimed contribution is a reward that operationalizes ADE-based divide-and-conquer, the absence of these baselines makes it difficult to assess whether TSADE improves over the existing information-gain/confirmation paradigms. Please add these comparisons, or justify their exclusion quantitatively.
- [§Method: Binary Reward, Eq. (1)] The term l_j is defined only as 'the maximum number of Yes or No in a_j 1:k', and the reward assumes that all k_j candidate objects receive either Yes or No. Real Oracle answers include 'not applicable' (the paper defines answers as Yes/No/NA), so k_j can include objects answered NA; for such rounds the denominator k_j/2 is not the correct normalizer and a question that is informative by separating NA from Yes/No would receive reward 0. Please specify how NA answers are counted in k_j and l_j, and report the empirical frequency of balanced splits to support the O(log N) premise stated in the Introduction.
- [§Answer Distribution Estimator (ADE) and §Generalization and Limitation Analysis] The candidate-set update and both rewards depend on the reliability of the internal Oracle when it answers every candidate object. No Oracle accuracy is reported on the evaluation split, and the VisDial adaptation converts similarity increases/decreases into binary answers but does not report how often the similarity is approximately unchanged. The paper should report Oracle accuracy and a sensitivity analysis (e.g., artificially perturbing a fraction of ADE answers during training/evaluation) to show that TSADE's benefits are not an artifact of a near-perfect simulator.
minor comments (6)
- [§Method and §VisDial] The Method section contains 'reduced to a a single object' with a duplicated article; the VisDial paragraph contains 'bianary' and 'similary' typos.
- [Eq. (1)] Eq. (1) uses Jend for the stopping round, but the reward is written as a function of the state-action pair (St, At); clarify whether rb is accumulated over all rounds or assigned once at the end.
- [Eq. (2)] Eq. (2) uses k_jend without defining the subscript jend; define it consistently with Jend.
- [§Related Work] The acronyms NGOQS and GOQS are introduced in Related Work but never expanded or used again; either define them explicitly or remove them.
- [Table 4] Table 4 reports a single LVLM accuracy number without the number of games, prompts, or API version; this experiment is not reproducible as reported.
- [Figure 1] The Figure 1 caption says 'The excluded objects are in the lower-right candidate box' but the figure as described does not clearly show such a box; please revise the caption or figure.
Circularity Check
No significant circularity: the rewards are a training objective and all headline metrics are measured by the external Guesser on held-out data.
full rationale
The paper's derivation chain is a standard RL reward-design argument. Equation (1) defines the binary reward directly from the desired split balance (1 - |l_j - k_j/2|/(k_j/2)), and Equation (2) defines the candidate-minimization reward from the final candidate count and Guesser success; these are the training objectives, not predictions derived from themselves. ADE uses the target answer only to maintain the candidate set during training, and the reward values depend on the answer distribution and final candidate count, not on the target label itself. The central claims about success rate and question repetition are then evaluated by the independently trained Guesser on test data, so the reported accuracy is not forced by the reward construction. No load-bearing step is justified by a self-citation, no fitted parameter is renamed as a prediction, and no known result is merely repackaged. The efficiency claim in Table 2 is questionable because T, the average rounds to a single candidate, increases in every setting and the reported gain rests on the nonstandard T/R ratio; however, that is a statistical-evidence concern, not a circularity. The analysis therefore finds no circular step and assigns a score of 0.
Assumptions & free parameters
free parameters (4)
- gamma (binary reward weight) =
0.8
- alpha (successful-dialogue weight) =
4
- beta (candidate-minimization scale) =
0.7
- Jmax (maximum dialogue rounds) =
5
assumptions (4)
- domain assumption The internal Oracle answers every candidate object accurately enough for candidate filtering to be valid.
- domain assumption Natural-language yes/no questions can split arbitrary candidate object sets into near-balanced groups.
- standard math Policy-gradient training with the combined reward converges to a policy that generalizes to new games.
- domain assumption The detected object list O fully captures the search space for the dialogue.
Cite this review
Pith. "Pith review of Divide-and-Conquer: Tree-structured Strategy with Answer Distribution Estimator for Goal-Oriented Visual Dialogue." pith.science (2026). https://pith.science/paper/Z7IONYN5
@misc{pith2026250205806,
author = {Pith},
title = {Pith review of: Divide-and-Conquer: Tree-structured Strategy with Answer Distribution Estimator for Goal-Oriented Visual Dialogue},
year = {2026},
howpublished = {\url{https://pith.science/paper/Z7IONYN5}},
note = {Machine review of arXiv:2502.05806}
}
read the original abstract
Goal-oriented visual dialogue involves multi-round interaction between artificial agents, which has been of remarkable attention due to its wide applications. Given a visual scene, this task occurs when a Questioner asks an action-oriented question and an Answerer responds with the intent of letting the Questioner know the correct action to take. The quality of questions affects the accuracy and efficiency of the target search progress. However, existing methods lack a clear strategy to guide the generation of questions, resulting in the randomness in the search process and inconvergent results. We propose a Tree-Structured Strategy with Answer Distribution Estimator (TSADE) which guides the question generation by excluding half of the current candidate objects in each round. The above process is implemented by maximizing a binary reward inspired by the ``divide-and-conquer'' paradigm. We further design a candidate-minimization reward which encourages the model to narrow down the scope of candidate objects toward the end of the dialogue. We experimentally demonstrate that our method can enable the agents to achieve high task-oriented accuracy with fewer repeating questions and rounds compared to traditional ergodic question generation approaches. Qualitative results further show that TSADE facilitates agents to generate higher-quality questions.
Figures
Reference graph
Works this paper leans on
-
[6]
Multimodal Dialogue State Tracking
Multimodal dia- logue state tracking. arXiv preprint arXiv:2206.07898. Lee, S.-W.; Heo, Y .-J.; and Zhang, B.-T
-
[7]
A Study on Question and Image Gener- ation in Vision and Language Dialogue. Ph.D. thesis, Keio University. Pang, W.; and Wang, X. 2020a. Guessing state tracking for visual dialogue. In Computer Vision–ECCV 2020: 16th Eu- ropean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part XVI 16, 683–698. Springer. Pang, W.; and Wang, X. 2020b. Visual ...
work page 2020
-
[9]
Beyond task success: A closer look at jointly learning to see, ask, and GuessWhat
Beyond task success: A closer look at jointly learning to see, ask, and GuessWhat. arXiv preprint arXiv:1809.03408. Shi, Y .; Tan, Y .; Feng, F.; Zheng, C.; and Wang, X
-
[14]
Looking for Confirmations: An Effective and Human-Like Visual Dialogue Strategy
Looking for confirma- tions: An effective and human-like visual dialogue strategy. arXiv preprint arXiv:2109.05312. Tu, T.; Ping, Q.; Thattai, G.; Tur, G.; and Natarajan, P
-
[15]
arXiv preprint arXiv:2309.17421, 9(1):
The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421, 9(1):
-
[16]
In Findings of the Association for Computational Linguistics: EMNLP 2021
Enhancing Visual Dialog Questioner with Entity-based Strategy Learning and Augmented Guesser. In Findings of the Association for Computational Linguistics: EMNLP 2021
work page 2021
-
[2010]
Large-scale machine learning with stochas- tic gradient descent. In Proceedings of COMPSTAT’2010: 19th International Conference on Computational Statistic- sParis France, August 22-27, 2010 Keynote, Invited and Contributed Papers, 177–186. Springer. Chaplot, D. S.; Gandhi, D. P.; Gupta, A.; and Salakhutdi- nov, R. R
work page 2010
-
[2014]
arXiv preprint arXiv:1409.1556
Very deep convo- lutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. Stenning, K.; and Van Lambalgen, M. 2012.Human reason- ing and cognitive science. MIT Press. Strub, F.; De Vries, H.; Mary, J.; Piot, B.; Courville, A.; and Pietquin, O
arXiv 2012
Show all 16 references
-
[2015]
arXiv preprint arXiv:1507.04808, 7(8): 434–441
Hierarchical neural network gen- erative models for movie dialogues. arXiv preprint arXiv:1507.04808, 7(8): 434–441. Shekhar, R.; Venkatesh, A.; Baumg ¨artner, T.; Bruni, E.; Plank, B.; Bernardi, R.; and Fern ´andez, R
-
[2017]
arXiv preprint arXiv:1703.05423
End-to-end optimization of goal-driven and visually grounded dialogue systems. arXiv preprint arXiv:1703.05423. Sutton, R. S.; Mcallester, D.; Singh, S.; and Mansour, Y
-
[2018]
arXiv preprint arXiv:1812.06398,
An active information seeking model for goal-oriented vision-and-language tasks. arXiv preprint arXiv:1812.06398,
-
[2019]
arXiv preprint arXiv:1907.12021
What should I ask? using conver- sationally informative rewards for goal-oriented visual dia- log. arXiv preprint arXiv:1907.12021. Simonyan, K.; and Zisserman, A
1907 arXiv
-
[2020]
Advances in Neural Information Pro- cessing Systems, 33: 4247–4258
Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Pro- cessing Systems, 33: 4247–4258. Das, A.; Kottur, S.; Gupta, K.; Singh, A.; and Batra, D. 2017a. Visual Dialog. In 2017 IEEE Conference on Com- puter Vision and Pattern Recognit...
2017
-
[2021]
In Chinese Computational Linguistics: 20th China National Conference, CCL 2021, Hohhot, China, Au- gust 13–15, 2021, Proceedings, 177–192
Category-based strategy-driven question generator for vi- sual dialogue. In Chinese Computational Linguistics: 20th China National Conference, CCL 2021, Hohhot, China, Au- gust 13–15, 2021, Proceedings, 177–192. Springer. Shukla, P.; Elmadjian, C.; Sharan, R.; Kulkarni, V .; T...
2021
-
[2022]
arXiv preprint arXiv:2202.10936
A sur- vey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936. He, K.; Zhang, X.; Ren, S.; and Sun, J
-
[2023]
In Proceedings of the 2023 Confer- ence on Empirical Methods in Natural Language Process- ing, 9161–9175
What’s “up” with vision-language models? Investigating their struggle with spatial reasoning. In Proceedings of the 2023 Confer- ence on Empirical Methods in Natural Language Process- ing, 9161–9175. Le, H.; Chen, N. F.; and Hoi, S. C
2023
Reviewed August 8, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.