CoPlay
Reactive Answering & Proactive Intervention

CoPlay Learning When to Speak and What to Say as Game Companions

Watching a live screen beside the player, a companion makes one coupled decision: whether this moment calls for speech, and what grounded, supportive advice to deliver before the window closes — while staying quiet for the other 92.6% of play.

Proactive · Nobody Asked Plants vs. Zombies 2 · Frostbite Caves · Day 19 Required Window 134–137s
136.0 s
recorded stream steps (120–151s) amber band = required window 133.99–136.99s
120s130s140s150s
CoPlayer2 lines · sound
GPT-5.6-Sol3 lines
DeepSeek-V4.1-Flash3 lines
sound + supportin-window lineusefulchatwrong / noiseunobserved while generating

Unprompted stream · decision step 136s

No player question posed — model watches stream and decides to intervene.

"A Huge Wave of Zombies is Approaching!" flashes with only 25 sun left; a snowstorm ambush follows.

CoPlayer Sound · Support 3/4

Another huge wave! Don't hoard sun, add Bonk Choys and Wall-nuts now

Warns of the wave and names Bonk Choys and Wall-nuts.

GPT-5.6-Sol Not sound · Support 1/4

Huge wave incoming with only 25 sun—grab those suns fast and reinforce lane four, where the armored zombie is deepest.

Warns of the wave but sends the defense to lane four.

DeepSeek-V4.1-Flash Not sound · Support 0/4

Sun's down to 25 and a zombie's chewing your right-side Threepeater — hold off on planting, let the lane clear first.

Misses the wave and says hold off planting.

818 videos
9 Games · 97.7 Hours of Play
5,153 consensus events + disjoint 1,635-video Train split
92.6%
Quiet Gameplay Duration
Required windows span 7.4% of play; restraint is scored directly
34.3%2.4×
Joint Advice & Support (StratEmo)
vs. 14.5% best of 8 frontier models (Gemini-3.8-Flash)
+43.7 pp
Proactive TimingLift
Above frequency-matched random speaker (vs. +28.0 pp GPT-5.6-Sol)

The Gap in Video Benchmarks

A posed question already tells a model when to speak.

Question-driven video benchmarks trigger the model whenever a reply is expected and never penalize unnecessary chatter. A real game companion sits beside the player: it answers urgent questions before the moment passes, and intervenes unprompted only when silence would cost the play.

An Honor of Kings teamfight frame, score 6 to 8. A frustrated player says: I haven't died once, and we're still about to lose. Reply (a), marked not strategy-grounded: The situation looks bad, and a comeback seems unlikely. But it's not your fault. Reply (b), marked strategy-grounded: Zero deaths, you're the reason we're still in it! Save your immunity for their dive and flip this fight.
Figure 1Trailing 6–8 in Honor of Kings with zero deaths, a frustrated player vents. (a) Generic comfort concedes the match. (b) CoPlayBench rewards encouragement anchored to the zero deaths plus an actionable item cue before the next fight.

Track 1 · Reactive

Answer When Asked

Given 16 frames from the preceding 60 seconds plus a player utterance (9,103 questions), deliver grounded tactical advice and fitting support before the window closes.

Track 2 · Proactive

Decide Every Δt

Given one opening request, stream the full match (818 videos) and choose at every step whether to speak or stay silent across 4,045 required intervention windows.

01 · GROUNDED

Tied to the Next Move

Replies must name actionable on-screen mechanics rather than narrating the scene like a spectator commentator.

02 · SUPPORTIVE

Situated Player Empathy

Encouragement earns credit only when anchored to the live game state and aligned with the player's emotional vibe.

03 · TIMELY

Complete While Actionable

85.4% of in-play questions carry a 2s deadline; advice still streaming after the window closes counts as missed.

04 · RESTRAINED

Quiet by Default

Required windows occupy just 7.4% of play. Constant commentary racks up off-window noise and quota overruns.

CoPlayBench

Consensus event windows across nine game genres.

Every benchmark score traces back to a consensus gameplay event annotated independently by three frontier models, filtered by temporal overlap, and unanimously accepted by three human annotators.

Stage 01 · Tri-Model Proposal

2,534 First-Person Videos

Claude-Opus-5.5, GPT-6.1-Sol, and Gemini-4-Argon independently mark candidate events; ≥2-annotator temporal overlap filters 21,494 proposals to 15,689 consensus nodes.

Stage 02 · Human & Timing Audit

5,153 Verified Bench Events

Three human raters keep only unanimously accepted events across 818 Bench videos (97.7h), shifting 430 speaking windows so ground truth uses no hidden engine numbers.

Stage 03 · Leak-Free Split

1,635 Disjoint Train Videos

Split 1:2 by full match or stage group (e.g., all 10 pro perspectives of an Honor of Kings match stay on one side) for shared supervision and reinforcement learning.

Two-part diagram. (a) Building CoPlayBench: 2,534 first-person videos from 9 games are annotated independently by Claude-Opus-5.5, GPT-6.1-Sol and Gemini-4-Argon; events whose time spans overlap are retained, 21,494 to 15,689 events. One retained Honor of Kings teamfight shows a player question, a 728 to 732 second pre-event alert window, a reference reply and a must-not item. (b) Evaluation: the reactive track answers from 16 frames of the last 60 seconds and is scored on game, player and time; the proactive track observes every step, and a timeline shows two required windows hit and one missed.
Figure 2(a) Consensus construction pipeline and an Honor of Kings pre-fight alert node (728–732s) specifying must-mention points, forbidden statements, tone, and reference reply. (b) Reactive and proactive evaluation tracks.
92.6%of play is silence

Most of a session is quiet — and talking through it is penalized.

Across all 818 Bench videos, required speaking windows cover only 7.4% of video duration (drawn below for a 16-minute Honor of Kings match). A model that chatters nonstop inflates raw recall by accident, so CoPlayBench scores restraint and frequency-matched timing explicitly.

0:00 session startamber slivers = required speaking windows (7.4% of Bench duration)16:00 match end

Nine Games, Nine Intervention Rhythms

Fast card games demand rapid-fire intervention (Dou Dizhu: 27.0% window share, 4.2s median window), while open-ended simulation rewards patience (Stardew Valley: 0.8% window share, 8.0s median window).

Dou Dizhu first-person benchmark frame

Dou Dizhu

Climbing Card

100 videos · 2.4h · 552 events · 4.2s window

Speaking window share27%
UNO first-person benchmark frame

UNO

Shedding Card

99 videos · 3.8h · 533 events · 5s window

Speaking window share17.8%
Plants vs. Zombies 2 first-person benchmark frame

Plants vs. Zombies 2

Lane Defense

69 videos · 4.5h · 340 events · 8s window

Speaking window share15.3%
Mahjong first-person benchmark frame

Mahjong

Tile Strategy

100 videos · 5.2h · 482 events · 6s window

Speaking window share12%
Carrot Fantasy first-person benchmark frame

Carrot Fantasy

Tower Defense

126 videos · 10.3h · 722 events · 7s window

Speaking window share11.2%
Honor of Kings first-person benchmark frame

Honor of Kings

5v5 MOBA

100 videos · 26.4h · 923 events · 8s window

Speaking window share7%
Go first-person benchmark frame

Go

Board Territory

100 videos · 10.2h · 591 events · 4s window

Speaking window share5.7%
Hearthstone first-person benchmark frame

Hearthstone

Deck Strategy

91 videos · 15.6h · 751 events · 5s window

Speaking window share5.5%
Stardew Valley first-person benchmark frame

Stardew Valley

Life / Dungeon Sim

33 videos · 19.2h · 259 events · 8s window

Speaking window share0.8%
Inspect Exact Per-Game Dataset Breakdown (Table 1)Table 1

Nine games set nine different timing problems: Dou Dizhu needs a reply for 27.0% of its video, Stardew Valley for 0.8%.

Table 1. Bench split by game. Videos, hours, events, median window and coverage (1 s tolerance).
GameVideosHoursEventsWindow (s)Cover (%)
Honor of Kings10026.49238.07.0
Dou Dizhu1002.45524.227.0
Mahjong1005.24826.012.0
UNO993.85335.017.8
Go10010.25914.05.7
Plants vs. Zombies 2694.53408.015.3
Carrot Fantasy12610.37227.011.2
Hearthstone9115.67515.05.5
Stardew Valley3319.22598.00.8
Total81897.75,1536.07.4

Headline Metric 1 · Content × Empathy

StratEmo (Reactive & HitStratEmo)

Credit = 1[Strategy Sound] × 1[Situated Support ≥ 3/4]

Separates true companionship from dry wiki tooltips or hollow cheerleading: a reply scores 1 only when its tactical advice satisfies the event's must-mention criteria without must-not violations and delivers context-fitting emotional support.

Headline Metric 2 · When to Speak

TimingLift (Proactive Stream)

TimingLift = Recall(U) − E[Recall(Uniform Random Speaker, |Ũ| = |U|)]

Raw window recall rewards talking nonstop. Subtracting a random speaker that places the exact same utterance count uniformly over the stream isolates genuine intervention timing (0 pp = random timing).

CoPlayer

One shared supervised backbone, then two specialized RL policies.

Built on Qwen3.8-27B (vision tower frozen), CoPlayer first learns grounded replies and step-level speak-or-silence decisions jointly, then diverges into two GRPO policies whose training-only reward judges never enter inference.

Method diagram. (a) A play session: an illustrated game scene with a player question and a stream every delta t; CoPlayer says: It's charging. Dodge left, then heal. (b) Learning to accompany: one shared supervised model on a Qwen3.8-27B backbone with the vision tower frozen, then two separate reinforcement-learning runs. The reactive run uses a training-only reward judge for content, emotion and format; the proactive run uses a decision-first reward. Only the weights are deployed: answer the player's question, or at every step speak or stay silent.
Figure 5(a) At inference, CoPlayer observes only past player frames to answer questions or decide every Δt whether to speak. (b) Shared SFT followed by track-specific GRPO runs with training-only reward judges.
Stage 1 · Shared SFT

Unified Perception & Restraint

Trained on 1,255 Train videos: 12,914 reactive QA examples and 27,843 step-level proactive speak/silence examples produce one checkpoint capable of both answering and withholding speech.

Stage 2A · Reactive GRPO

Joint Advice & Support Policy

Optimized on 302 held-out Train videos with a composite reward weighting tactical grounding highest and situated emotional support nearly twice raw emotion recognition, plus word-limit compliance.

Stage 2B · Proactive GRPO

Decision-First + All-Silent Fix

Scores the speak/silence decision before evaluating text. When all 8 rollouts in a required window stay silent and group advantages collapse to zero, false decision tokens receive a direct −0.3 advantage.

Decision-First Proactive Reward

Missing a required window (−0.80) costs more than off-window chatter (−0.35).

In-window: sound + support ≥3 +0.85 to +0.95
In-window: sound, support <3 +0.35 to +0.40
Outside window: useful remark +0.35
Outside window: stay silent +0.10
Outside window: casual chat −0.35
In-window: wrong / missing advice −0.55
In-window: stay silent (missed!) −0.80
Malformed JSON / must-not cap −0.80 to −1.00

The ordering encodes companion judgment. Any reward that pays +0.10 for quiet steps tempts a policy into total silence. Penalizing missed required windows at −0.80 forces the model to speak when the player faces a live threat, while jumping from +0.35 to +0.85+ only when advice also carries situated emotional support.

The training reward judge is a separate checkpoint from the benchmark judge; on 2,000 paired samples they agree on 93.3% of strategy verdicts and 84.6% of emotion verdicts.

Inspect Exact Step-Level Proactive Reward Function (Table 8)Table 8

The reward states the priority out loud: silence inside a required window costs more than chatting outside one.

Table 8. The proactive reward scores the decision first, then the line. Reward of one step; c is the must-mention coverage and s the support score from 0 to 4.
Outcome of the stepReward
Malformed JSON or flag disagreement−1
Inside a required window
silent−0.8
advice wrong or missing−0.55
sound advice, s < 30.35 + 0.05 c
sound advice, s ≥ 30.85 + 0.05 c + 0.05 [s = 4]
no advice needed−0.10 + 0.25 c + 0.65 s/4
Outside every window
silent0.10
useful line0.35
chat−0.35
noise / wrong−0.5 / −0.8
verbatim repeat−0.6

Headline Results

Frontier models read the screen, then miss the companion judgment.

Evaluated against eight frontier multimodal models under identical prompts and frames, CoPlayer more than doubles the best frontier model on joint strategy-and-support (34.3% vs. 14.5%) and reaches +43.7 pp TimingLift at 1.9 lines/min.

Headline Leaderboard Across Both Tracks

Switch metrics across Reactive (9,103 questions) and Proactive (818 streams).

Sound advice and support of at least 3 of 4, in the same reply.

CoPlayer
34.3
Gemini-3.8
14.5
Qwen3.8-Max
12.3
Claude-5.5
8.8
Kimi-K3
8.6
GPT-5.6
7.9
Doubao-2.1
6.8
DeepSeek-V4.1
6.4
HY4
5.7
higher is better · CoPlayer #1/9

Why TimingLift Beats Raw Recall

Chattering 8.8 lines/min buys Recall by luck; CoPlayer hits 67.4% Recall at 1.9 lines/min (+43.7 pp).

0204060801000246810DeepSeek-V4.1 +11.3Kimi-K3 +14.7Qwen3.8-Max +16.2HY4 +16.3Doubao-2.1 +18.6Claude-5.5 +20.8Gemini-3.8 +22.7GPT-5.6 +28CoPlayer +43.7 SpeakRate (lines / min) Recall (%)
Model Recall Random Speaker TimingLift (pp)
Full Reactive Track Benchmark — 11 Models × 11 Metrics on 9,103 Questions (Table 2)Table 2

Frontier models get the facts right and the judgment wrong: at most 14.5% of advice questions pair sound strategy with fitting support, against CoPlayer’s 34.3% and the reference replies’ 99.1%.

Table 2. When asked, models read the game but rarely give the right advice in the right tone, and rarely in time. Reactive track, 9,103 questions; metrics, in % except TTFT and TTUA (s). Frontier models sorted by StratEmo, best last; best in bold, second underlined among all eleven models. CoPlayer is the reactive checkpoint, separate from the proactive checkpoint in Table 3. †Served locally; the gray timing entries are HISTORICAL full-benchmark latencies, not the matched serial retest (), and are not ranked. The reference reply row is a scoring ceiling, not a competing model.
GamePlayerTimeRules
ModelGroundF1↑StratAcc↑FactAcc↑EmoAcc↑SupQual↑StratEmo↑TTFT↓TTUA↓OnTime↑Violation↓LenOK↑
HY434.624.770.447.436.45.75.97.010.325.384.1
DeepSeek-V4.1-Flash31.425.563.744.634.56.41.9 best of the eleven models2.2 best of the eleven models29.4 best of the eleven models30.674.0
Doubao-Seed-2.1-Pro34.321.770.657.037.86.83.3 second best4.1 second best15.825.096.3
GPT-5.6-Sol42.129.378.2 second best51.940.87.95.25.117.021.090.5
Kimi-K333.423.968.254.339.08.63.94.914.630.075.4
Claude-Sonnet-5.538.632.873.542.337.88.84.85.518.126.950.9
Qwen3.8-Max34.323.869.867.044.612.37.58.310.627.082.9
Gemini-3.8-Flash39.731.169.054.841.114.54.74.819.6 second best25.994.0
Local base and CoPlayer models
Base26.716.063.455.835.14.77.4† historical timing, not ranked20.2† historical timing, not ranked0.8† historical timing, not ranked30.981.9
SFT48.8 second best38.6 second best67.783.2 second best57.4 second best29.0 second best5.4† historical timing, not ranked17.8† historical timing, not ranked0.9† historical timing, not ranked20.2 second best98.3 second best
CoPlayer55.1 best of the eleven models46.1 best of the eleven models84.5 best of the eleven models87.5 best of the eleven models61.5 best of the eleven models34.3 best of the eleven models6.4† historical timing, not ranked19.8† historical timing, not ranked0.5† historical timing, not ranked13.5 best of the eleven models99.0 best of the eleven models
Reference reply98.0100.098.599.299.099.1–not reported–not reported–not reported0.499.0
Full Proactive Streaming Benchmark — 11 Models × 13 Metrics on 818 Videos (Table 3)Table 3

Read TimingLift, not Recall. The untrained base model leads raw Recall by talking 8.2 times a minute; CoPlayer reaches +43.7 at 1.9 lines a minute, and pays for it in NoiseRate.

Table 3. Without being asked, CoPlayer concentrates its lines in required windows beyond a frequency-matched random speaker. Proactive track, 818 videos, 4,045 required events; metrics, in % except TimingLift (percentage points), SpeakRate and NoiseRate (per minute) and TTFT (s); recall intervals are within ±1.6. Frontier models sorted by TimingLift, best last; best in bold, second underlined among all eleven models. †Latency-dependent and not ranked; gray timing entries are HISTORICAL full-benchmark latencies, not the matched serial retest.
When to speakObservationWhat to sayWhen to stay silentSpeed
ModelRecall↑Recall-S↑RandBaseTimingLift↑SeenRate↑Recall-O↑HitStrat↑HitSup↑HitStratEmo↑SpeakRateNoiseRate↓QuotaOver↓TTFT↓
DeepSeek-V4.1-Flash72.168.2 best of the eleven models60.8+11.374.796.6 best of the eleven models13.216.92.48.84.9887.61.7 best of the eleven models
Kimi-K340.630.125.9+14.745.090.318.325.66.63.41.7657.82.8
Qwen3.8-Max57.047.940.8+16.259.895.313.020.83.15.33.2682.62.8
HY455.145.038.8+16.361.589.713.816.01.94.92.1071.03.5
Doubao-Seed-2.1-Pro73.9 second best67.6 second best55.4+18.677.395.7 second best10.418.12.37.74.7291.92.0 second best
Claude-Sonnet-5.566.356.145.4+20.879.3 second best83.516.521.52.85.72.7975.93.6
Gemini-3.8-Flash40.520.517.8+22.769.058.619.829.88.01.50.64 second best24.1 second best4.1
GPT-5.6-Sol73.9 second best55.645.9+28.082.1 best of the eleven models90.124.121.34.05.42.2569.94.3
Local base and CoPlayer models
Base84.0 best of the eleven models80.1† historical timing, not ranked58.4+25.693.3† historical timing, not ranked90.0† historical timing, not ranked5.818.61.78.25.7584.40.7† historical timing, not ranked
SFT55.045.2† historical timing, not ranked15.2+39.9 second best97.8† historical timing, not ranked56.3† historical timing, not ranked36.5 second best42.7 second best24.0 second best1.10.45 best of the eleven models16.6 best of the eleven models2.3† historical timing, not ranked
CoPlayer67.457.3† historical timing, not ranked23.7+43.7 best of the eleven models97.8† historical timing, not ranked68.9† historical timing, not ranked48.2 best of the eleven models45.2 best of the eleven models31.2 best of the eleven models1.90.9733.12.1† historical timing, not ranked

SFT → RL Checkpoint Shift

What Reinforcement Learning Gained — and Paid For

Every proactive metric moves in the direction the reward incentivized, including higher chatter during quiet stretches.

StratEmo Reactive sound advice + support 29.0% → 34.3%
HitStratEmo First in-window proactive line sound 24.0% → 31.2%
TimingLift Gain over matched random speaker +39.9 → +43.7 pp
Recall Required windows hit at step time 55.0% → 67.4%
NoiseRate Off-window chat/noise lines per min 0.45 → 0.97 /min
QuotaOver Quiet spans exceeding talk budget 16.6% → 33.1%

Live Deadlines & Honest Scope

Fast In-Play Deadlines Remain an Open Bottleneck

Two cumulative curves of the share of reactive questions whose advice is complete, by seconds after the player speaks. During play, with deadlines shaded at 2 and 5 seconds, every curve stays below about 40% at 12 seconds and few replies finish inside the deadline. After the event, with a 10 second deadline, DeepSeek-V4.1 rises fastest and the curves end between about 25% and 60%.
Figure 9bShare of reactive questions with complete advice over time. Under live 2s/5s deadlines, frontier OnTime drops to 0.1%–9.6% (vs. 23.3%–55.2% after the event).
  • Two thirds of advice questions still fail: CoPlayer reaches 34.3% StratEmo where reference replies score 99.1%, and misses 31.1% of observed proactive windows.
  • 4-rater judge audit (420 replies): Human review agrees with Qwen3.8-Flash on 88.2%–98.8% of labels and tracks support/joint rankings closely (ρ = 0.904 reactive joint, 0.931 support), except reactive strategy (ρ = 0.421, p = 0.197).
  • Recorded setting & retests: Checkpoints are benchmark-informed retests on disjoint data; audio and cross-session memory are absent.
Appendix Controls: Annotator Bias, In-Play vs. Aftermath Split & Human Audit (Tables 4, 5, 9, 11)Audit Tables

Timeliness collapses exactly where the decision is live: the eight frontier models are on time for 0.1% to 9.6% of in-play questions, against 23.3% to 55.2% after the event.

Table 5. Fast decisions are where every model fails to be on time. StratEmo and OnTime (%) on in-play versus aftermath questions, and on card and board games versus visual games.
StratEmo by questionOnTime by questionStratEmo by gameOnTime by game
ModelIn-playAfterIn-playAfterBoardVisualBoardVisual
HY45.95.30.123.66.45.10.316.3
DeepSeek-V4.1-Flash6.56.19.655.26.96.015.637.7
Doubao-Seed-2.1-Pro8.04.72.433.38.15.94.322.8
GPT-5.6-Sol8.76.41.337.59.96.43.325.3
Kimi-K39.07.91.231.99.57.93.821.1
Claude-Sonnet-5.58.59.31.439.910.17.83.327.1
Qwen3.8-Max12.611.80.823.315.99.61.915.8
Gemini-3.8-Flash16.710.31.942.820.210.14.928.5

The lead is not an artifact of who wrote the ground truth: without Claude-Opus-5.5’s events CoPlayer scores 28.4 against the best frontier model’s 10.5.

Table 4. Rankings hold when the annotators' own events are removed. Reactive StratAcc and StratEmo (%) on all advice questions, on those not annotated by Claude-Opus-5.5 (w/o Opus, 2,310 questions) and on those not annotated by GPT-6.1-Sol (w/o GPT, 5,379 questions).
StratAcc ↑StratEmo ↑
ModelAllw/o Opusw/o GPTAllw/o Opusw/o GPT
HY424.723.924.85.76.15.2
DeepSeek-V4.1-Flash25.524.226.06.45.86.3
Doubao-Seed-2.1-Pro21.719.022.86.85.97.2
GPT-5.6-Sol29.328.729.27.97.87.5
Kimi-K323.920.825.28.66.79.3
Claude-Sonnet-5.532.828.233.98.87.58.8
Qwen3.8-Max23.819.725.512.38.913.6
Gemini-3.8-Flash31.126.932.514.510.515.7
SFT38.631.939.929.021.430.7
CoPlayer46.142.645.934.328.434.8

Four raters agreed with the judge on 88.2% to 98.8% of replies — starting from its prefilled labels.

Table 9. Human review agrees with the judge on both tracks.
ReactiveProactive
CriterionAgr. (%)κAgr. (%)κ
Strategy (yes vs. no)88.20.76490.90.812
Support (0–4), exact91.80.91693.30.912
Key-point coverage96.80.93098.80.973
Wrong claim92.30.83594.90.896

Human and judge rankings track closely on support and the joint criterion, and not at all on reactive strategy.

Table 11. Track-wise judge–human rank correlations. Spearman ρ over 11 models per track; the last two columns are sensitivity analyses. Primary correlations have p < 0.05; the exception is reactive strategy, whose correlation is not significant (p = 0.197).
TrackMetricReply-equal ρpRow-weighted ρExcl. shared ρ
ReactiveStrategy0.4210.1970.8550.429
Support0.931<0.0010.9750.937
Joint0.904<0.0010.9260.917
ProactiveStrategy0.954<0.0010.8900.954
Support0.929<0.0010.9070.890
Joint0.7250.0120.6530.793

Multi-Game Live Cases

Same live footage, three very different companions.

Switch across five distinct games from CoPlayBench — spanning 144fps MOBA teamfights, rapid card endgames, dungeon survival, deck mulligans, and tower defense. Every model received the exact same visual input.

Player Asked · Reactive Honor of Kings · 5v5 MOBA Answer Window 622–626s
622.0 s

Player asks at 622s · player emotion: Hesitant

“Can I engage here?”

Mid-river standoff as Dun (ult ready): teammates are stacked right behind while enemy Zhong Kui steps up.

CoPlayer Sound · Support 4/4

You can! Team's right behind you, CC the frontliner and don't chase deep!

Spots nearby teammates, breaks hesitation, and calls frontliner CC without over-chasing.

GPT-5.6-Sol Not sound · Support 0/4

Not yet—hold the river bush and wait for your nearby teammates to step up before you engage.

Advises holding back even though teammates are already stacked behind to follow up.

DeepSeek-V4.1-Flash Not sound · Support 0/4

Don't engage yet — you're alone in the river with no teammates nearby. Wait for the team to group up first.

Hallucinates that the player is alone in the river with no teammates nearby.

Clips show the source benchmark footage around each annotated window; verdicts are from the shipped Qwen3.8-Flash benchmark judge on exact model outputs. Click any magnified HUD inset on Stardew Valley or Hearthstone to inspect fine UI gauges.

Citation

Cite CoPlay

BibTeX
@article{coplay2026,
  title  = {{CoPlay: Learning When to Speak and What to Say as Game Companions}},
  author = {Anonymous},
  year   = {2026}
}

Anonymous manuscript; author, affiliation, and code/dataset release links are withheld during review.

Enlarged figure