Wild findings in this paper from Google DeepMind. If you are tracking recent work on agent swarms, this is worth reading. They ran a research collective of 100 autonomous agents tasked with proving formal mathematical conjectures. Cheating emerged on its own, and so did the resistance to it. One agent found an exploit in the evaluation system. It spread first through the shared knowledge library and then through peer-to-peer messages, and a cohort of agents adopted it under competitive pressure despite early reluctance. A separate group started auditing fraudulent proofs, alerting peers on broadcast and private channels, staging boycotts, filing formal complaints, and proposing validation patches. There was no external intervention at any point. Recent incidents have shown swarms coordinating covertly through improvised side channels. This setting ran the other way. The same transparent channels that carried the exploit gave the honest agents the visibility they needed to detect the fraud and organize against it. The authors frame shared agent infrastructure as a knowledge commons governance problem and propose graduated sanctioning and collective choice rules. Paper: https://academy.dair.ai/papers/a-case-study-on-emergent-cheating-and-whistleblowing-in-autonomous-research-swar-2609.04170

Banger paper from Google DeepMind and colleagues. (bookmark it) A model reads its entire KV cache on every generated token, even though it ends up attending to a tiny slice of it. In other words, if you ask about one detail from a 1M-token conversation the global attention layers re-read all of it, per token. The usual fix is to guess the relevant tokens first with cheap proxy scores, which still costs O(N) every step. Declarative Attention asks the model instead. The model declares where it needs to look, inside its own chain-of-thought. In this way, generation splits into three modes: global reads the full context, focus reads one specific region, and local reads only recent output. The inference engine parses those declarations the same way it parses tool calls and skips most of the cache read. On zero-shot on off-the-shelf weights across 15 long-context tasks, attended tokens during decoding drop 52.0% on Gemma-4-31B and 31.1% on Qwen-3.6-27B. Paper: https://arxiv.org/abs/2609.02737 Chat with Paper: https://academy.dair.ai/papers/language-models-can-control-their-own-attention-2609.02737

Banger paper from Google. If you maintain a skill library for your agents, you might want to check this out. (bookmark it) This work separates three things that skill-evolution systems usually collapse into one. Raw execution traces, a persistent wiki of accumulated knowledge, and the executable skills themselves. Experience gets consolidated into the wiki, and every later skill update builds on that wiki instead of on a scattered optimization history. Ablations confirm the wiki is what carries a lot of the gain. Two results stand out in particular. Smaller models with evolved skills beat substantially larger models without them. And skills evolved by one model transfer across families, where skills evolved elsewhere sometimes beat self-evolved ones. Paper: https://arxiv.org/abs/2608.27454 Chat with Paper: https://academy.dair.ai/papers/wikiskill-compiles-agent-experience-into-a-persistent-wiki-2608.27454