As more self-interested agents improve side by side, using peers' discoveries well becomes its own capability. LLMs can build on each other, but don't yet balance exploring and copying well enough to beat solo learning. Training for this is an exciting next step for RSI!

Timing matters. With the same four observations, spreading them across learning beats copying early (+2.4 pts for GLM, +3.0 over solo). Waiting gives peers time to produce something worth copying!

Skills really do get passed on: agents copy a peer's revision, revise it again, and pass it along. But copying narrows the search. Solo populations keep ~5 independent skill lineages, while social populations collapse onto 1–2 early discoveries (just 1 for GPT-OSS)!

Next, we let agents write, revise, and share skill files on IFBench. Peers change how models learn: GLM-5.3-Flash finds good skills sooner, and GPT-OSS-120B uses a third fewer tokens. But at matched cost, neither beats solo learners. More efficient, but not yet more effective!

We tried two direct fixes: reserving tokens for acting and forcing independent search. Each works on its target, but neither makes social learners reliably beat solo ones. Useful social information isn't enough; agents must balance what they copy against what they find!

Why? It isn't that LLMs copy bad options. For GPT-OSS, a copied arm beats the agent's own best find 93% of the time Instead, social agents collapse onto what peers are doing and stop exploring. And some reason so long about whom to copy that they run out of tokens before acting

We start simple: a social learning tournament where agents spend one token budget observing peers or acting. A hand-designed social learning algorithm beats a strong solo exploration policy. But Qwen, Ministral, and GPT-OSS are 30–52% less efficient with peers!

This is an old tension in cultural evolution and economics: copying only pays if someone explores, but everyone would rather others pay that cost. We test when LLMs learning from peers beat solo learners at the same token budget, counting the search behind copied skills

A peer’s discovery can save an agent from starting from scratch. But if everyone follows what others have found, who keeps exploring? Today’s LLMs copy useful discoveries, yet often converge too quickly on peers’ choices. Better information doesn’t reliably mean better learning.

Thanks!

Thanks!

Definitely!!! I’m very excited about this frame as a means for understanding the systems we’re apart of and designing new ones!!

This spans game theory, biology, AI, and more. As we build recursively self-improving multi-agent systems, our results suggest a simple idea: let access to compute co-evolve with how agents interact and reproduce. This can give cooperation an edge—and make groups more capable!

We isolated this in a Prisoner Dilemma where agents co-evolve cooperation and replication timing. We find unproductive defectors no longer takeover: Co-evolving both traits in a single phase allows cooperators to persist even when randomly paired with anyone in the population!

What if compute must be earned, not given? We required programs to solve math problems framed as sequential social dilemmas to earn energy. Programs evolved to favor joint tasks over individual ones, coordinating to maximize collective energy while suppressing theft!

Uniform energy allows random programs to evolve into self-replicating cooperators! But unequally distributed energy makes well-mixed populations break down. To survive, programs must interact locally on a grid, sharing resources to build stable code despite energy asymmetries.