Research
Sentient Labs is a frontier open-source AI research lab, peer-reviewed at NeurIPS, ICML, COLM, and more, and cited across 100+ institutions including MIT, CMU, Alibaba and Microsoft.
We advance how systems reason, learn, and self-improve — spanning agentic frameworks, evaluation benchmarks, model provenance, and AI security. Built to ship: as open-source tools, protocol contributions, and new primitives for open-source AI builders at the frontier.
The Scaffold Effect in Coding Agents
Naman Vats, Oleg Golev — Sentient Labs
Public leaderboards rank models by name and pass rate alone, underspecifying the harness around them — the scaffold that issues tools, manages context, and decides when to stop. The paper argues scaffold choice is a hidden variable in coding-agent evaluation.
MINDGAMES: A Live Arena for Evaluating Social and Strategic Reasoning in Multi-Agent LLMs
Alexander Buyantuev, Ilya Makarov, Pramod Viswanath — Sentient Foundation
Selected by NeurIPS to run a first-of-its-kind competition for agents that improve themselves through social games — advancing the next paradigm of self-improving AI.
EvoSkill: Automated Skill Discovery for Coding Agents
Salaheddin Alzubi, Noah Provenzano, Jaydon Bingham, Weiyuan Chen, Tu Vu — Sentient Labs · Virginia Tech
An open-source toolkit that takes a benchmark and a coding agent (like Claude Code) and evolves it into a state-of-the-art specialist on that benchmark's task in minutes.
ROMA: Recursive Open Meta-Agent Framework for Long-Horizon Multi-Agent Systems
Salaheddin Alzubi, Sewoong Oh — Sentient Foundation · Virginia Tech · UC Berkeley · UC San Diego · University of Maryland
An open-source recursive meta-agent framework for building high-performance, long-horizon multi-agent systems.
CryptoAnalystBench: How to Evaluate AI Agents When There Are No Clear Answers
Anushri Eswaran, Oleg Golev, Darshan Tank, Sidhant Rahi — Sentient Labs · UC San Diego
A benchmark for evaluating crypto AI agents on long-form analytical responses, built from 198 real production queries from Sentient Chat and scored across four dimensions via an LLM-as-a-judge rubric.
Are Robust LLM Fingerprints Adversarially Robust?
Edoardo Contente, Pramod Viswanath, Sewoong Oh — University of Washington · Princeton University · Sentient Foundation
Examined the adversarial robustness of LLM fingerprinting, identifying four fundamental vulnerabilities that enable adaptive attacks bypassing authentication across ten major methods — pointing to the need for robustness by design in model attribution and IP protection.
SPIN-Bench: How Well Do LLMs Plan Strategically and Reason Socially?
Jianzhu Yao, Zerui Cheng, Pramod Viswanath — Princeton University · UT Austin · Sentient Foundation
A multi-domain evaluation unifying PDDL planning, competitive board games, cooperative card games, and multi-agent negotiation to measure strategic planning and social reasoning in LLMs.
LiveCodeBench Pro: How Do Olympiad Medalists Judge LLMs in Competitive Programming
Zerui Cheng, Peiyao Sheng, Zixuan Wang, Pramod Viswanath — Sentient Foundation
A competitive-programming benchmark (Codeforces, ICPC, IOI) annotated by Olympiad medalists, revealing that top models score 0% on hard problems — exposing that high scores reflect implementation precision, not genuine algorithmic reasoning.
CHANCERY: Evaluating Corporate Governance Reasoning Capabilities in Language Models
Lucas Irwin, Arda Kaz, Peiyao Sheng — Sentient Labs
A Sentient Labs benchmark that tests whether LLMs can reason about whether proposed corporate actions comply with real-world governance charters.
The Dobby Report
Sentient Labs
Sentient's family of fine-tuned LLMs balancing strong technical capability with a deliberately unhinged personality — using data overloading and token-level routing to beat the tone-vs-reasoning trade-off. Validated by ~2M votes on Dobby Arena.
Open Deep Search: Democratizing Search with Open-source Reasoning Agents
Salaheddin Alzubi, Creston Brooks, Purva Chiniya, Edoardo Contente — Sentient Foundation
OpenDeepSearch — a lightweight yet powerful search tool designed for seamless integration with AI agents.
Real AI Agents with Fake Memories: Fatal Context Manipulation Attacks on Web3 Agents
Atharv Singh Patlan, Peiyao Sheng, S. Ashwin Hebbar, Prateek Mittal — Sentient Foundation
Investigated AI agent vulnerabilities in blockchain financial ecosystems via ElizaOS, showing how prompt or memory injection can trigger unintended asset transfers and protocol violations.
OML 1.0: A Primitive for Reconciling Open Access with Owner Control in AI Model Distribution
Zerui Cheng, Edoardo Contente, Ben Finch, Oleg Golev, Andrew Miller, Sandeep Nailwal, Sewoong Oh, Himanshu Tyagi, Pramod Viswanath — Sentient Foundation
Introduced OML — a primitive enabling models to be freely distributed for local execution while maintaining cryptographically enforced usage authorization. First to introduce fingerprinting into open models without compromising performance.