Agents are learning to improve themselves, in swarms and in training

Swarms that evolve in latent space

In a preprint, a framework called SwarmEvolve runs evolutionary search with swarms of hundreds of small, low-cost agents that reason in latent space without decoding to text. The agents share information by passing KV caches, which avoids the latency of text generation, and a summarizer agent performs crossover by processing the combined KV caches of swarm members. A novelty filter based on distances between agents' hidden state vectors keeps the population diverse, so the swarm keeps exploring instead of converging on one early idea. m:…7410

The speed difference is large. SwarmEvolve reaches best scores in a median of 3 to 8 minutes, while other state-of-the-art frameworks require 60 minutes to reach the same scores. With Qwen3 8B, it improves on solutions found by AlphaEvolve using frontier LLMs. This suggests the expensive part of agent-driven discovery was never the reasoning itself but the cost of turning every thought into text and back again, and that skipping that step lets a crowd of small models outrun a single large one. m:…7410

Sibling harnesses rescue failed training runs

In a second preprint, Harness-S takes a different route to the same goal: instead of changing the agent, it changes the harness the agent trains in. It is a selective harness-rescue training method that builds a pool of specialized sibling harnesses from a common base, each developed along a distinct repair direction, and invokes the pool only when initial attempts under the base harness fail. The sibling harness pool remains fixed during policy training, and a router selects the top variants using harness behavior profiles and voting across multiple LLMs. The trained policy is then returned to the original base harness. m:…6481

The results show the rescue transfers. On Qwen3.5-9B AppWorld, Pass@3 goes from 64.89 with plain GRPO to 72.52 under the base harness, and the trained policy substantially outperforms policy-only training when returned to the original base harness. The SWE-Gym assertion-fallback rate is 48.5% for Harness-S versus 13.3% for GRPO, and the method achieves about 7.89 points of average gain over GRPO with only about 18% more rollouts. For a founder, the practical read is that training budget spent on better harnesses, not just more rollouts, can buy capability that survives the trip back to production conditions. m:…6481

Decisions that improve without retraining

In a third preprint, a method called DPDM separates what an agent knows from what it has learned from experience. It combines frozen language-model representations with a shared pretrained conditional predictor, using Qwen3.5-27B as the semantic encoder and a single frozen LimiX-2 predictor serving all five task categories, and it learns from deployment evidence without parameter updates. That separation is the point: the paper argues it makes learning from experience an independently improvable foundation-model capability, so the decision-making layer can get better without retraining the underlying model. m:…6494

Mean accuracy across 50 benchmarks rose from 78.80% with few-shot prompting to 84.92% with DPDM at evaluated history budgets, with absolute accuracy gains of 34.26% on BFCL tasks and 28.87% on ToolSandbox tasks in ToolPRMBench. The gains concentrate where agents actually fail, in tool use and multi-step reasoning, which suggests the method is most valuable for products where an agent must choose the right tool and the right next step, and where retraining after every deployment lesson is not an option. m:…6494

Knowing what the parts do, and what the companion is for

A fourth preprint turns the interpretability lens inward. Researchers trained a sparse autoencoder on the output of mixture-of-experts layers in the Qwen3-30B-A3B model, studying layers 4, 30 and 46, and found that expert contribution is more concentrated than gating suggests, with most latents having only one or two dominators. The 128 experts in layer 30 group into 11 domains, and experts specialize below coarse domains, with distinct experts for Java code, front-end web code, geometry, and arithmetic and polynomials. Knowing which expert does what is what makes targeted improvement possible, instead of retraining the whole model to fix one skill. m:…6489

A fifth preprint steps back from the engineering and asks what all this improvement is for. It argues that physical AI should include AI-augmented human intelligence via lifelong cognitive symbionts, and defines a cognitive symbiont as a lifelong, embedded, real-time, private, low-friction and proactive companion. The paper frames the core idea as the cognitive gap, the distance between what a person could bring to a situation and what they can hold in working memory, and asks how much intelligence a machine can enable in a particular person. It identifies 23 open problems regarding symbiont governance, including risks of covert assistance in examinations. The implication for founders is plain: agents that improve themselves are arriving, and the product question is shifting from how autonomous they can be to how well they augment the person using them. m:…6490