Natasha Jaques

This account is verified. Learn more

@natashajaques

Assistant Professor leading the Social RL Lab socialrl.cs.washington.edu @uwcse and Staff Research Scientist at @GoogleAI.

Seattle, WA natashajaques.ai Joined June 2009

1.1K Following 36K Followers 1.6K Posts

Latest posts 5

Pinned post

Natasha Jaques @natashajaques 4d

Everyone is suddenly obsessed with multi-agent swarms for solving long-horizon tasks, but most of this work looks like the development or evolution of multi-agent harnesses that build in some coordination structure. Can we instead let agents learn an emergent coordination and communication protocol purely driven by task performance?

Our new paper is the first work to show how to get cooperative multi-agent RL post-training of swarms to work better than single-agent best-of-N sampling. Using asynchronous multi-agent training with no prescribed roles, we show that agents learn coordination and communication protocols that improve token efficiency and success rate over compute-balanced baselines by as much as 166%, and generalize to up to 3x larger swarms.

Extremely excited for the research agenda around improving multi-agent communication and coordination this will unlock!

Kunal Jha Verified account @kjha02 5d

What if a swarm of AI agents could learn to work together on its own? 🐝

No leader. No roles. No turn order. Just one shared reward.

Our new paper s³: multi-agent RL lets LLM swarms self-organize, beat best-of-N, and scale past their training size!

scalingswarms.com 🧵
View quoted post on X

Natasha Jaques @natashajaques 4d

If you want to learn from humans, you should probably know something about how humans work. We found that prior work on robot learning from interventions relied on flawed assumptions about how humans give feedback. We ran human studies, analyzed the issues, and built a much more accurate human simulation model. Using this, we developed a state-of-the-art method for learning from humans interventions during deployment, allowing robots to adapt online to complex tasks like unbuttoning a shirt!

pranav Verified account @_pranavnt 4d

Excited to share the first paper of my undergrad, Q-OIL, which we’ll be presenting at #CoRL2026 in Austin!

Interventions provide a natural source of feedback for robots to learn from in the real world. But, humans intervene suboptimally, which needs to be accounted for in both simulation benchmarking and algorithms.

We analyze human intervention data and measure intervention suboptimality in both timing and execution. Thus, rather than using interventions as supervision, we use them as soft exploration guidance.

(1/8)

View quoted post on X

Natasha Jaques @natashajaques Oct 23

More UED ideas, this time for automatically generating tool-use tasks. Can we make longer and more complicated tasks that maximize a solver’s learning potential? We show that using our approach we can increase the proportion of hard tasks on AppWorld from 19.8% to 52.5%, and make more semantically & functionally diverse tasks, while still keeping things realistic (e.g. the user asks the agent to “email my roommate” instead of “email Cory”).

Alex Nam Verified account @hjalexnam Oct 23

Evaluating tool-use capabilities is expensive and domain-dependent. Once existing test data become saturated, expert humans need to design new tasks. Alternatively, LLMs can auto-generate challenging tasks from existing toolsets and task primitives through evolutionary search using multiple agents.

With @MichaelD1729, @natashajaques, @minsuk_chang.
Paper: arxiv.org/abs/2610.27307 for how to auto-generate tasks from existing benchmarks and use them for failure diagnosis

View quoted post on X

Natasha Jaques @natashajaques Oct 14

Our new paper pushes multi-agent recursive self-improvement even further. This time the proposer agent builds whole Harbor sandbox environments to challenge the solver agent. It learns from the solver’s successes and failures to update its own memories, skills, and tools, and build increasingly challenging---yet valid---tasks.

We show powerful benefits both when using a proprietary model as the proposer, and in self-play, when both the solver and proposer are the same open-weight model. In this RSI setting, Qwen3.6-35B-A3B doubles its score on Terminal Bench 4, and performs so well on HiL-Bench it actually exceeds Claude Opus-5.

More evidence of the power of multi-agent co-adaptation to unlock continuous improvement and emergent capabilities.

Simon Yu Verified account @simon_ycl Oct 14

Recursive self-improvement needs agents that can build their own training environments. But researchers still part of the loop, making it hard to scale.

We introduce AutoEnvScaling, which turns Data Flywheel into a terminal task: a proposer agent builds complete Harbor tasks and sandbox. It builds just like human researchers. This enables self-play RL for terminal agents: one model improves as both builder and solver, without human-in-the-loop.

View quoted post on X

Natasha Jaques @natashajaques Oct 13

Proud to have signed this letter. As AI gets more powerful, we urgently need better guardrails.

Nicolas Papernot @NicolasPapernot Oct 13

Over 400 academics in 30+ countries signed our open letter calling for immediate changes in AI development and deployment practices to ensure that cybersecurity incidents do not continue to occur, as future incidents will be far more damaging to our critical infrastructure.

View quoted post on X