AI Paper Club
A book club, but for AI papers.
This month’s papers
1. The Model: The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation
Li et al., 2026 · Presented by Chandhru Karthick
Instead of copying an RL-trained teacher’s outputs, RIDE pushes the student along the direction RL moved the teacher’s hidden states, and past it. Across four base and RL-teacher pairs, students match or beat the RL teacher, while output-space extrapolation can make them worse.
For discussion: if RL post-training is mostly a direction in representation space, how cheap can we make it?
2. The Harness: HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
Wu et al., 2026 · Presented by Gourob Podder
A benchmark that scores the scaffolding, not just the answers: models build an agent harness from a minimal seed, then keep revising it. Self-built harnesses trail human-engineered ones on coding and search but hold their own on writing and ML experiments, and their self-improvements are unstable and transfer poorly across models.
For discussion: which parts of your harness would you trust a model to rewrite?
3. Production: Measuring Agents in Production
Pan et al., ICML 2026 · Presented by Arindam Sharma
20 in-depth case studies and a survey of 86 practitioners across 26 domains. Real agents are simpler than the hype: 68% take 10 steps or fewer before a human steps in, 70% prompt off-the-shelf models instead of tuning weights, and 74% lean mainly on human evaluation. Reliability is the top challenge.
For discussion: what would it take to trust an agent with step 11?
Agenda
6:00pm: Food, drinks and intros.
6:15pm: The Papers
Three short talks, each followed by discussion.
7:30pm: Networking
Keep talking and meet like-minded peers!
Want to lead a paper?
We’re looking for engineers to present at future sessions. If there’s a recent AI paper you’d like to walk the group through, email Diana at diana@solsticehealth.co.
Hosted by Solstice
Solstice is the precision commercialization platform built for life sciences. Our mission is accelerating the knowledge of life-saving medicines. Backed by a $21M Series A led by Transformation Capital, we build agentic AI that pharma teams rely on in a domain where accuracy and compliance aren’t optional.
We’re growing our engineering team in New York. See Open Roles: https://jobs.ashbyhq.com/solstice
