Research map / Large language models / Agents, code and math

VisInteract: Towards Dynamic Interactive Text-to-Visualization under Imperfect Queries

NeurIPS 2026 · Large language models › Agents, code and math · approach: LLM-guided program synthesis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces VisInteract-Bench, a dynamic interactive Text-to-Vis benchmark with imperfect queries and a leakage-controlled user agent, and Vis-MCTS, an MCTS method with progressive widening, cross-rollout memory, and dimension-aware reward decomposition.

Problem

Real-world visualization queries are ambiguous, incomplete, or factually incorrect, while existing Text-to-Vis systems assume well-specified inputs and single-pass generation; no benchmark or method supports multi-turn interactive intent recovery for charts.

Approach

Reframes Text-to-Vis as interaction-driven intent recovery, and adapts MCTS to it with progressive widening over unbounded tool arguments, shared QA/feedback memories across rollouts, and per-dimension reward routing (data, visual, intent) instead of uniform backpropagation.

Key results

• Vis-MCTS reaches Merge Task Score 51.46% (Qwen) and 57.19% (Gemini), +13.40% and +16.27% over the strongest interactive baseline Best-of-N (ReAct).
• Non-interactive Self-Correction LLM reaches only 7.38% (Qwen) and 10.66% (Gemini) Merge Task Score, more than 5x lower than Vis-MCTS.
• Ablation: removing visual feedback drops Merge Task Score by 27.22%; removing cross-rollout memory drops it by 17.40% on Qwen.

Surprising findings

• Best-of-N over ReAct gives inconsistent gains across backbones (+9.01% Qwen vs +1.30% Gemini).
• Vis-MCTS reaches 100% renderable rate on both backbones while nvAgent reaches only 66.67% and 81.06%.

Limitations

• Evaluation relies on LLM-based judges whose reliability is bounded by model capability.
• User Agent simulates idealized user behavior; real users may give noisier feedback.
• Higher computational cost than single-pass generation due to multiple rollouts.

Experimental setup

Datasets: BIRD Mini-Dev

Benchmarks: VisInteract-Bench

Metrics: KF Pass Rate Strict Success Task Score Renderable Rate Cohen's Kappa Fleiss' Kappa

Baselines: Self-Correction LLM nvAgent MultiVis-Agent ReAct Best-of-N (ReAct)

Builds on: Monte Carlo Tree Search Progressive Widening ReAct UCT BIRD Mini-Dev

Closest papers by meaning

• nvBench 2.0: Resolving Ambiguity in Text-to-Visualization through Stepwise Reasoning (NeurIPS 2025 · similarity 0.89)
• Multimodal DeepResearcher: Generating Text-Chart Interleaved Reports from Scratch with Agentic Framework (AAAI 2026 · similarity 0.86)
• Magnifying What Matters: Attention-Guided Adaptive Rendering for Visual Text Comprehension (NeurIPS 2026 · similarity 0.86)
• CoDA: Agentic Systems for Collaborative Data Visualization (ICLR 2026 · similarity 0.86)
• VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations (ICLR 2026 · similarity 0.85)
• JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence (ICLR 2026 · similarity 0.85)