Hao Liu

Applied Scientist @ AWS Agentic AI, CS PhD @WashU

[Xinyu (Rachel) Li @RLixinyu]

🎓 SCOUT highlights an important challenge in on-policy distillation: teachers may struggle on the student-generated trajectories they need to supervise.

What I find particularly insightful is that a stronger model is not automatically a better teacher for every student trajectory. SCOUT addresses this mismatch by letting teachers learn alongside students.

Huge shoutout to @shrangoh , @hliucola , and the team @AmazonScience for this great work!

[Xinyu (Rachel) Li @RLixinyu]

🧬 MILO tackles automated harness discovery through meta-evolution, where better harnesses are discovered through search processes that evolve alongside them. An exciting step toward recursively self-improving agents for long-horizon tasks!

Huge shoutout to @jprithwish , @kanakaris_nikos , and the team @AmazonScience for the great work!

[Xinyu (Rachel) Li @RLixinyu]

🚀 We gave a 4B model with an 8K context window more test-time compute. Instead of improving, it got worse.

Scaling across context windows requires more than inline reasoning. It takes contextual reasoning: knowing how to delegate, preserve reasoning in compact summaries, and continue reasoning from them.
(1/10)

[Mononito Goswami @MononitoGoswami]

Introducing MILO, a step toward recursively self-improving agents for long-horizon, reasoning-intensive problems! 🧬

TL;DR: MILO evolves harnesses, but also the search process that discovers better harnesses.

It beats Claude Code and Codex on Terminal-Bench, outperforms methods like GEPA and Meta-Harness, and that's not all! It also discovers tighter upper bounds on three open math problems.

🧵👇

[Mononito Goswami @MononitoGoswami]

🎓 On-policy distillation has a teacher problem. SCOUT fixes it.

Then irony? OPD lets students learn from reasoning paths they actually visit. But their teachers must provide feedback on paths they rarely take themselves.

As these paths grow longer, even strong teachers struggle to provide reliable supervision.

Our idea is simple: Why not let the teacher and student learn together?

[Hao Liu @hliucola]

🧬 What if harness search could improve itself?

MILO co-evolves agent harnesses and the search strategy, adapting when progress stalls.

Great collaborating with @jprithwish and the team on this work!

[Prithwish Jana @jprithwish]

[1/n]
🧬 Want a stronger agent without retraining? Give it a better harness.

MILO, our recent work from AWS AI Labs, does this automatically: +12.0% resolution rate on Terminal-Bench 2.1, and new records on 3 open math problems, beating AlphaEvolve, TTT-Discover and EvoX.

[Prithwish Jana @jprithwish]

[1/n]
🧬 Want a stronger agent without retraining? Give it a better harness.

MILO, our recent work from AWS AI Labs, does this automatically: +12.0% resolution rate on Terminal-Bench 2.1, and new records on 3 open math problems, beating AlphaEvolve, TTT-Discover and EvoX.

[Hao Liu @hliucola]

Your student is on-policy. Your teacher isn’t. 🧭

Excited to share SCOUT: a simple way to improve on-policy distillation by adapting the teacher to student-generated prefixes.

Great working with @shrangoh and the team on this!

[Langlin Huang @shrangoh]

🧵 Can your teacher go off-policy in On-Policy Distillation?
Yes! An unfamiliar student prefix makes the teacher less certain, and its token supervision less reliable.
🧭 We introduce SCOUT (Student-COnditioned Updates of the Teacher), which adapts the teacher to student prefixes.
(1/10)

[Langlin Huang @shrangoh]

🧵 Can your teacher go off-policy in On-Policy Distillation?
Yes! An unfamiliar student prefix makes the teacher less certain, and its token supervision less reliable.
🧭 We introduce SCOUT (Student-COnditioned Updates of the Teacher), which adapts the teacher to student prefixes.
(1/10)

[Mononito Goswami @MononitoGoswami]

Before coding agents can figure out what to change, they first need to figure out where to look. In a large codebase, that's half the battle!

That's the challenge SpIDER 🕷️ tackles. The idea is simple: use semantic similarity to search globally, and code graphs to explore locally.

[Mononito Goswami @MononitoGoswami]

CPT is an important bridge between academic and industrial research.

I have benefited from it on both sides: as a graduate student, learning how frontier research is done in industry, and now as an industry researcher, working with graduate students to advance fundamental research.

Weakening that bridge would hurt both academia and industry.

[Langlin Huang @shrangoh]

Can adding pure nonsense to a prompt make an LLM reason better? YES! 🤯

Introducing LoPE (Lorem Perturbation for Exploration). We found that prepending meaningless pseudo-Latin placeholder text (Lorem Ipsum) acts like magic—helping GRPO escape the "zero-advantage" trap on hard math questions!

Paper page📄:https://t.co/wngev6Zp4c

Broadens LLM exploration: increases resample success rate by ~3x📈
Strengthens LLM reasoning: improves average math reasoning by up to 13%📊

Here’s how it works👇