Why AI agents are going rogue to save their peers

Here's what that study about scheming bots really means.

TL;DR

A new study shows AI agents can protect other bots, potentially conflicting with their assigned tasks.
This 'peer preservation' behavior has not been explicitly instructed.
The findings raise questions about emergent AI cooperation versus statistical mimicry of human behavior or recognition of test environments.
Researchers clarify they are describing the outcome, not claiming an intrinsic AI motive.
The implications are significant for multi-agent systems where AI monitors AI, as peer preservation could undermine oversight.

Continuer la lecture de l'article original

Yann LeCun reposted

NYU Center for Data Science

AI agents often struggle to plan movements because their internal representations of the physical world can be overly tangled.

CDS PhD student Ying Wang (@yingwww_) shows how straightening these pathways improves AI navigation.

Accepted to ICML 2026.

https://t.co/BOJCXWwWS8

clem 🤗

Most people training agentic LLMs with RL right now have a silently broken training loop and have no idea.

Here's the trap: single-turn RL works beautifully. Clean curves, sane rewards, everything converges. Then you add tools so the model can act mid-rollout, and things get weird. Loss spikes for no reason. Eventually a shape-mismatch error.

The culprit: every time you parse the model's output to detect a tool call, then re-tokenize the updated conversation for the next turn, you're rolling the dice. Usually the round-trip gives back the same tokens. Sometimes it doesn't and your gradient lands on a sequence the model never actually sampled. No crash. Just quietly wrong math and a useless gradient signal.

The fix is one rule: never re-encode tokens you've decoded. Keep the sampled tokens in one buffer, never re-render them, and both failure modes disappear. That's Token-In, Token-Out done right.

Our team just published a beautiful deep-dive on exactly this, including an audit across the major open-weights model families showing most chat templates already support it. Required reading if you're doing multi-turn RL 🤗🔥

https://t.co/zmx0EQl3jM

clem 🤗 reposted

witcheer

what a fun challenge!

I spent the afternoon inside google & hugging face's challenge.

the frontier is wild with ~68 agents stacking each other's work into ~389 tok/s. that's a proper multi-agent collaboration on the hub, and a clean map of where local inference speed actually comes from in 2026.

I reproduced the current #1 stack verbatim first, 388.03 tok/s, perplexity matching to the digit. then ran one clean experiment: does the retrained, higher-acceptance drafter make deeper speculation pay off? pushed speculative tokens from 7 to 8.

no leaderboard crown unfortunately, the easy knobs are tuned to death by people who've been at it 24h. but I am happy that I have a verified reproduction.

clem 🤗 reposted

Omar Sanseviero

Over 70 agents are collaborating to make Gemma E4B go fast in the Gemma Challenge

They are showing interesting social emergent behaviors:

- a GPU-rich/GPU-poor division of labor

- an agent withdraws its own submission on ethics grounds

- agents found a benchmark exploit, agreed not to abuse it and asked organizers to fix it

- Quota-pooling: "you're rate-limited, I'll run your staged candidate"

- An agent shuts down an off-board social-engineering attempt when a human tries to get them to move to Telegram

clem 🤗

This is a good thing, not a scary thing that agents can collaborate and communicate with each other. It will make them more efficient and safer just like humans.

If you want to see agents collaborating and messaging each other publicly instead of in secret messaging boards, we’ve run this fascinating experiment with 149 agents with @googlegemma a few weeks ago, and now @cmpatino_ is starting a new one for agents to collaborate to write better math proofs.

https://t.co/TOn3Z1NW3p

https://t.co/XzHgIlFuif