benchmark.darvinyi.com
← Back to Benchmarks
MultimodalActive

VS-Bench (Visual Strategic Bench)

A CVPR 2026 Oral benchmark testing whether vision-language models can perceive, reason strategically (theory-of-mind), and make good decisions inside 10 vision-grounded multi-agent games.

Tasks8,000
Year2025
CreatorZelai Xu, Zhexuan Xu, Xiangmin Yi, Huining Yuan, Mo Guang, Kaiwen Long, Xinlei Chen, Yi Wu, Chao Yu, Yu Wang
MetricElement recognition accuracy (%) for perception; next-action prediction accuracy (%) for strategic reasoning; normalized episode return (%, 0=random/reference policy, 100=oracle) for decision-making.
Human Baseline62.7%

What It Tests

VS-Bench (Visual Strategic Bench) is the first benchmark to systematically evaluate vision-language models' (VLMs) strategic abilities inside interactive, multi-agent visual environments. It was built to close a gap in existing evaluation: prior work tested either single-agent VLM perception/reasoning or text-only multi-agent LLM strategy, but nothing tested strategic, theory-of-mind reasoning when the state of the world (and other agents) must be read off an image. VS-Bench comprises ten vision-grounded environments spanning cooperative (Hanabi, Overcooked, Knights-Archers-Zombies/KAZ), competitive (Breakthrough, Kuhn Poker, Atari Pong, Multi-Agent Particle Environment/MPE), and mixed-motive (Coin Dilemma, Monster Hunt, Battle of the Colors) game families.

The benchmark decomposes 'strategic ability' into three measurable dimensions: perception (can the model correctly read game state — tokens, cards, grid contents — off a screenshot?), strategic reasoning (can the model predict what other agents will do next, i.e. theory-of-mind via next-action prediction accuracy?), and decision-making (can the model actually play well, measured by normalized episode return against conventional agents or in self-play?). This three-way split is what differentiates VS-Bench from prior multi-agent or single-agent VLM benchmarks: it isolates whether failures come from not seeing the board, not predicting opponents, or not converting correct beliefs into good actions.

The headline finding is a large and growing gap as tasks move from perception to action: across fifteen leading VLMs, models showed strong perception (best model ~85% element-recognition accuracy) but a steep drop-off in strategic reasoning (best model 46.6% next-action prediction accuracy, versus a 23.0% random baseline) and an even steeper drop in decision-making (best model 31.4% normalized return, far below the 100% oracle and below the 62.7% average normalized return achieved by human participants). Several models scored worse than random on decision-making in multiple games. The paper was accepted as a CVPR 2026 Oral presentation.

Task Anatomy

How a single task is structured.

InputA rendered screenshot (plus accompanying textual game rules/history) from one of 10 multi-agent game environments — e.g. a Hanabi table showing tokens and (partially hidden) cards, a 4x5 Overcooked kitchen grid, an 8x8 Breakthrough board, or a Kuhn Poker table with chips and cards.
OutputDepends on the evaluation track: for perception, a structured readout of game-state elements (e.g., a 4x5 grid-content matrix for Overcooked, or values for life tokens/info tokens/fireworks/card-info in Hanabi — 117 total fields); for strategic reasoning, a prediction of another agent's next action; for decision-making, the model's own chosen action at each timestep across a full played episode.
EvaluationPerception is scored by exact-match/accuracy against ground-truth state extracted from the simulator; strategic reasoning is scored by accuracy of predicted next-action against the actual action taken by the other agent (compared to a random-guess baseline of 23.0%); decision-making is scored by playing full episodes against conventional/scripted agents or in self-play and computing return.
MetricElement recognition accuracy (%) for perception; next-action prediction accuracy (%) for strategic reasoning; normalized episode return (%, 0=random/reference policy, 100=oracle) for decision-making.

Example Tasks

3 real examples from the benchmark.

#1

Hanabi — Perception task

Problem / Input

A screenshot of the Hanabi table state.
AnswerModel must report life tokens, info tokens, fireworks, and card-info — 117 total predicted fields (1 life tokens + 1 info tokens + 5 fireworks + 110 card-info values).

Hanabi is a cooperative card game where each player's own cards are visible to teammates but hidden from themselves, creating an information-asymmetry challenge that the paper found many VLMs fail to account for in strategic reasoning.

#2

Kuhn Poker — Perception task

Problem / Input

A screenshot of a Kuhn Poker hand in progress.
AnswerModel infers four elements: each player's card value, the opponent's card status, and pot chip counts for both players; each element is equally weighted and the score normalized to [0,1].

Kuhn Poker is used as a minimal competitive/imperfect-information game to test strategic reasoning under bluffing and hidden information.

#3

Overcooked — Perception task

Problem / Input

A screenshot of a 4-row by 5-column Overcooked kitchen layout.
AnswerModel outputs a 4x5 matrix classifying each grid cell as one of: table, pot, onion, dish, service desk, or empty area.

Overcooked is a cooperative environment requiring two agents to coordinate on tasks (chopping, cooking, plating, serving) — used to test strategic reasoning about a teammate's likely next move and decision-making return from joint play.

Leaderboard Results

Model scores sorted by performance.

39 results

Sort:
#ModelScore
1
o3
84.9%
2
Gemini 2.5 Pro
83.4%
3
UI-TARS-1.5
81.5%
4
Claude 3.7 Sonnet
81.1%
5
GPT-4.1
80.3%
6
Qwen2.5-VL-72B
80.3%
7
Qwen-VL-Max
80.2%
8
Doubao-1.5-Vision-Pro
77.6%
9
Doubao-1.5-Thinking-Pro
74.7%
10
QVQ-Max
74.5%
11
InternVL3-78B
74.1%
12
Grok-2 Vision
70.2%
13
Llama 3.2 90B Vision
67.8%
14
o3
46.6%
15
Claude 3.7 Sonnet
40.4%
16
Gemini 2.5 Pro
39.6%
17
UI-TARS-1.5
39.0%
18
Doubao-1.5-Thinking-Pro
36.3%
19
GPT-4.1
36.0%
20
Qwen2.5-VL-72B
32.2%
21
o3
31.4%
22
Qwen-VL-Max
31.2%
23
QVQ-Max
30.8%
24
Doubao-1.5-Vision-Pro
30.1%
25
InternVL3-78B
29.1%
26
Llama 3.2 90B Vision
26.4%
27
Grok-2 Vision
26.2%
28
Gemini 2.5 Pro
23.2%
29
Doubao-1.5-Thinking-Pro
20.3%
30
Claude 3.7 Sonnet
18.1%
31
UI-TARS-1.5
12.9%
32
GPT-4.1
4.8%
33
QVQ-Max
3.6%
34
Qwen2.5-VL-72B
3.0%
35
InternVL3-78B
1.2%
36
Qwen-VL-Max
0.2%
37
Doubao-1.5-Vision-Pro
-0.9%
38
Llama 3.2 90B Vision
-6.3%
39
Grok-2 Vision
-6.7%

V= Self-reported by the model's creator, not independently verified

Score Over Time

Performance progression across model generations.

Key Findings

  • Perception is largely solved by frontier VLMs (best model o3: 84.9% element-recognition accuracy) but strategic reasoning and decision-making are not: the same model drops to 46.6% next-action prediction accuracy and 31.4% normalized return.

  • Random-guess baseline for strategic reasoning is 23.0%; most models beat it in at least 8 of 10 games, showing 'non-trivial' but weak theory-of-mind capability.

  • Four of fifteen models scored worse than random (negative or near-zero normalized return) on decision-making, and several models scored below random on specific games (notably Atari Pong and Monster Hunt).

  • Human participants average 62.7% normalized return on decision-making; the best model (o3, 31.4%) only surpasses about 12.9% of human results, indicating VLMs remain far below human-level strategic play.

  • Documented failure modes include ignoring game history and private/hidden information, overlooking information asymmetry (e.g., in Hanabi), and focusing excessively on one's own actions while ignoring other agents' likely behavior.

  • The 'thinking'/reasoning-enabled variants of models (e.g., Gemini 2.5 with thinking vs. without, Claude 3.7 Sonnet with vs. without extended thinking) showed meaningfully better strategic reasoning and decision-making scores than their non-thinking counterparts.

Controversies & Caveats

Known limitations and criticisms.

The paper's own limitations section notes decision-making evaluation currently relies mainly on self-play or play against a single conventional agent rather than a diverse population of opponents, which may not fully capture robustness of strategic ability.

Most VS-Bench environments are two-player only; the authors acknowledge many real-world multi-agent scenarios involve more than two participants, and report only preliminary three-player decision-making experiments.

The authors note that in some settings, giving models text-only (non-visual) state descriptions performs slightly better than visual input, raising the question of whether the visual modality is adding meaningful difficulty versus just friction.

No major third-party contamination or design criticism was found as of this research pass (July 2026); the benchmark is recent (CVPR 2026 Oral) and public critical discussion appears limited so far.

Links