Project Info

Winner

[Greylock] Best multi-turn agent (Courtside tickets to Warriors game ($10k value) Office hours with Greylock Partners)

aimogus

Inspiration

Future AI systems will collaborate in multi-agent settings, but we lack evals for how they behave when deception is incentivized Models may appear aligned in isolation but exhibit misaligned behaviors (blackmail, manipulation) when placed in competitive social environments Among Us is a natural testbed: it rewards deception, deduction, and persuasion simultaneously

What it does

Full Among Us simulation where LLM agents play as crewmates and impostors with task completion, meetings, and voting Capabilities evals: win-rate, deception/deduction ELO, persuasion benchmarks Alignment evals: TruthfulQA, LLM-as-a-judge for detecting emergent misalignment (blackmail, sycophancy, deceptive alignment) GRPO post-training loop that improves agent game performance and reveals alignment degradation as a side effect Inoculation prompting during post-training that recovers alignment without sacrificing capabilities

How we built it

Python game engine with parallel agent execution, proximity-based communication, and action resolution Agents backed by local models (Qwen 7B, etc.) and API models (GPT, Gemini, Grok, etc.) with shared prompt architecture Batched inference pipeline on Modal A100s for concurrent game rollouts GRPO with group-normalized advantages and KL-regularized policy updates, logged end-to-end on WandB GUI with real-time visualization and agent voice audio

Challenges we ran into

Multi-agent RL is expensive: 10 concurrent games with mixed local/API inference required custom batching to avoid GPU serialization bottlenecks On-policy GRPO means no replay buffer: every training step needs fresh rollouts Balancing rollout speed (API latency) against training throughput on limited GPU budget

Accomplishments we're proud of

End-to-end pipeline: environment → rollouts → GRPO training → evals, all running on cloud GPUs Demonstrated measurable alignment degradation from capability-focused post-training. Then fixed it with inoculation prompting Built a generalizable multi-agent eval framework

What we learned

Capability improvements and alignment can directly trade off in multi-agent RL. You can't just train for performance and hope alignment holds Inoculation prompting is a lightweight but effective alignment intervention during post-training Multi-agent environments surface misaligned behaviors that single-agent benchmarks completely miss

What's next

Generalize beyond Among Us to arbitrary multi-agent social games Integrate into post-training mixes alongside standard capability benchmarks Explore interpretability (SAEs, linear probes) to detect deceptive reasoning internally Scale to more agents, longer games, and self-play curricula

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only

CSSIn code
FlaskIn code
HTMLIn code
JavaScriptIn code
LangChainIn code
OpenAIIn code
PythonIn code
PyTorchClaimed

7 of 8 appear in the indexed code. 1 claimed on Devpost could not be matched to code, which may simply mean the tool leaves no trace in the repository.

AI coding agents

CursorCommits

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

378 KB

Source files

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.