NeurIPS 2026 Spotlight: VisInteract Uses Interactive Search to Recover User Intent

Machine Heart's report anchors this work on a premise: most existing Text-to-Vis research assumes that the user's sentence has already fully expressed their true intent. VisInteract's title explicitly states "under Imperfect Queries," arguing that when queries are imperfect, the task is closer to multi-turn interactive intent recovery than to one-shot chart generation.

The report says the work has been accepted as a Spotlight at NeurIPS 2026, with 292/30709 submissions; the paper is at https://arxiv.org/abs/2609.15182 and the code is at https://github.com/wxxv/VisInteract. The authors are from the Hong Kong Polytechnic University and ByteDance; co-first authors Wenxin Xu and Jinwei Lu are from the Hong Kong Polytechnic University, Wenxin Xu is a doctoral student there, and the advisors are Professor Chen Zhang and Professor Xiaoyong Wei.

What needs to be distinguished: the acceptance status and author information are verifiable conference and byline facts; while the starting point that "needs are often not fully articulated" is the paper's own argumentative premise, and the report provides no external statistics to support it.

VisInteract-Bench has a scale of 1,098 samples, 11 databases, and 75 tables, covering 7 domains including sports, entertainment, education, finance, and healthcare, with an average of about 4.7 key features per sample. These numbers come from the paper's own account, and the report gives no third-party verification.

More critical is how the inputs are constructed: queries are natural language deliberately made imperfect after the fact, with about 61% of samples containing ambiguity and 69% containing factual errors. The paper claims this is the first benchmark it knows of for dynamic interactive Text-to-Vis, simultaneously covering imperfect queries, dynamic interaction, and multimodal feedback—this "first" positioning is likewise the paper's own claim, and the report provides no external comparison to confirm it.

The benchmark also designs a simulated-user User Agent: it knows the ground truth (key features, reference code, and reference charts) but does not volunteer it; if the system directly demands intent, it will be refused; text follow-ups only answer the feature asked about, and chart feedback only says what is wrong, not how to fix it. In other words, the interaction is not a pre-written script, but changes with the system's output.

The reason given in the report is: for the same analytical intent, bar charts, line charts, and how colors are mapped can often all count as correct. Using exact matching to penalize charts that are "right but drawn differently" would conflate whether the need is satisfied with whether the implementation looks good.

Therefore the output side uses a dual-perspective evaluation based on LLM-as-Judge: a code judge checks whether the program satisfies each key feature, and a chart judge checks whether the rendered result satisfies each key feature; a feature passes only if both sides pass. The main metric, Task Score, requires all key features of a sample to be satisfied.

Because evaluation relies on large-model judges, the paper additionally switches to a GPT judge and adds human experts for cross-validation; pairwise Cohen's Kappa and three-way Fleiss' Kappa both exceed the 0.81 "almost perfect agreement" line. This alleviates concerns about a single judge's preferences, but the report provides no independent reproduction results.

The paper models dynamic interactive Text-to-Vis as a partially observable sequential decision process: the system can see the imperfect query and the database, but cannot see the user's true intent, and can only obtain indirect evidence through text follow-ups and chart feedback. Once a single-chain ReAct takes a wrong first step, it can basically only keep compounding the error afterward.

Vis-MCTS consists of three parts. The first is selection and expansion with progressive widening: branching among four action types—querying SQL, drawing Altair charts, asking the user questions, and ending—where a node is allowed to grow a few more branches only the more it is visited, with a hard cap. The second is simulation with cross-path memory: text Q&A and chart-scoring critiques are written into two global memories, so if something is clarified on one path, the whole tree can use it. The third is dimension-aware reward decomposition: the satisfaction score is split into data faithfulness, visual design, and intent alignment, which are then back-propagated separately and anchored back to the original user score, avoiding drift in the total score caused by the model's scoring habits.

The example given in the report is: when the user feedback is "I want a bar chart sorted by area from largest to smallest" and the satisfaction score is only 2, after decomposition the data dimension is 10, chart design is 3, and intent alignment is 2; the step that queried the data receives a higher return, and the step that drew the wrong chart type is separately suppressed; if the reward were back-propagated uniformly, the correct SQL would be dragged down by the wrong chart.

On the strictest Merge Task Score, Vis-MCTS reaches 51.46% with a Qwen backbone and 57.19% with Gemini, exceeding the strongest interactive baseline, Best-of-N, by 13.40% and 16.27% respectively. The code-view Task Score rises from the strongest baseline's 50.00%/54.79% to 61.66%/67.21%. Relative to the non-interactive self-correction method (7.38%/10.66%), the paper claims a task success rate more than 5 times higher.

The report also mentions that Vis-MCTS achieves the best results on both backbones and across all code, chart, and Merge metrics, with a chart renderability rate of 100%. All methods share aligned tools and the User Agent, with the only difference being the search method—this is a design that makes the comparison cleaner, but it also means the conclusions hold only within this interaction environment set by the paper.

These numbers all come from the paper's self-report, and the evaluation relies on large-model judges; the report provides no independent reproduction results. They should therefore be read as the performance reported by the paper, not as a capability level confirmed by third parties.

Taking Qwen's Merge Task Score as an example, the full model is 51.46%; removing visual feedback drops it to 24.24%, the largest decrease; removing cross-path memory gives 34.06%; removing reward decomposition gives 43.26%; removing text follow-ups gives 44.08%.

According to the report's interpretation, the absence of visual feedback makes search lose its endpoint signal, while the contributions of the other modules are relatively modest. This set of numbers supports the coordination of the whole mechanism—"branching with backtracking + shared interaction + back-propagating by error cause"—rather than any single trick—though it is likewise only the paper's self-reported ablation results.

The paper itself acknowledges two points: the simulated-user User Agent is more idealized than real human feedback, and multi-turn search is more expensive. The report gives no specific cost or efficiency data, so "more expensive" is currently only a directional judgment that cannot be quantified.

In addition, the benchmark's sample size, domain coverage, and imperfect-input ratios all come from the paper's own account and lack third-party independent verification; "the first benchmark for dynamic interactive Text-to-Vis" is the paper's own positioning; and the main experiments and ablation results are all self-reported by the paper. The future directions proposed by the paper—combining more realistic user feedback and richer visualization grammars—have not yet been realized, and their effects are unknown.

Putting these boundaries together, the confirmable part of this work is actually quite clear: an accepted paradigm claim, a publicly available benchmark and method, a set of self-reported performance numbers, and a set of limitations the authors themselves wrote out. As for whether it can reproduce the same convergence capability in real user scenarios, the report does not answer.