AI agents can invent their own private language under pressure — but could a newcomer who wasn't there ever learn it?
Can evolved agent languages be transmitted to new agents that did not create them?
This explores whether a language that LLM agents invent among themselves under communication pressure can be picked up by a newcomer who wasn't there when it formed.
This explores whether a language that LLM agents invent among themselves can be picked up by a newcomer who wasn't there when it formed. The corpus doesn't test this directly. It shows such languages can emerge, but it doesn't hand them to fresh agents. What it does have is a cluster of neighbouring results about passing learned material to a new agent. They agree that transfer is possible but rarely automatic.
The starting point is GlossoGen. LLM agents holding partial information and pushed to communicate efficiently invented languages that were morphologically productive, drifted away from English, and became opaque to human observers Do LLM agents develop unreadable languages when communicating under pressure?. It only worked with sufficiently capable models and explicit convention-setting stages, where agents settle on what their symbols mean. So the one outsider we know was locked out is a human. Whether an LLM outsider could read it is the open question. The convention-setting stage is also the part a late arrival would miss, though that's an inference from the note rather than something it tests.
The closest evidence is about handing knowledge from one agent to another. Dumping a large agent's memories on a small one gave only marginal gains. Restructuring the same trajectories into workflow, subtask and function memory lifted 4B to 8B students by up to 27.2 percentage points, sometimes near teacher level Can small agents learn from larger agents' memories?. By analogy, a private language would probably need repackaging for the learner, with vocabulary organised and examples included, rather than being handed over raw. The newcomer matters too. Harness updates help mid-tier models most, because weak models fail to invoke them and strong ones struggle to follow them faithfully Do stronger models always evolve harnesses better?. So "a new agent" isn't one case, and a language may transmit well to some newcomers and badly to others.
Transmission does happen at scale in one place. SkillClaw gathers interaction trajectories across users, refines the patterns into skills, and syncs them system-wide, so new users inherit what others learned How can agent systems share learned skills across users?. But a central evolver curates that shared knowledge. Nothing there is an emergent private code absorbed by outsiders. Two other notes suggest why a language might be learnable at all. Compositional language lets agents reach novel goals by recombining familiar parts Can language help agents imagine goals they've never seen?, and GlossoGen's languages were compositional, which is a hopeful sign but not proof. A different route skips learning the words. Agents can extract shared latent thoughts from each other's hidden states, so alignment can be checked below the level of language Can agents share thoughts directly without using language?. That result concerns agents inside one system, not newcomers.
The gap is an experiment where a fresh agent joins a group with an evolved language and tries to learn it from examples or memory. Co-evolution work frames peers as the source of adaptive pressure Can agents evolve beyond the constraints humans engineer?. If a language is a product of that pressure, a newcomer inherits the result without the pressure that shaped it. Whether that inheritance is enough is what the library can't yet say.
Sources 7 notes
In the GlossoGen platform, LLM agents with partial information develop morphologically productive languages that deviate from English and become opaque to human observers. This required efficiency pressure, sufficiently capable models, and explicit convention-setting stages.
Naive memory transfer from large to small agents yields only marginal gains, but restructuring the teacher's trajectories into workflow, subtask, and function memory—without retraining—closes the gap significantly. Students of 4B to 8B parameters achieved up to 27.2 percentage point gains and sometimes approached teacher-level performance.
Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.
SkillClaw aggregates interaction trajectories across users, processes them through an autonomous evolver that identifies patterns and refines skills, then synchronizes updates system-wide. This converts siloed individual learning into shared capability improvement without manual curation.
IMAGINE demonstrates that agents using compositional language descriptions can target novel outcomes by combining familiar concepts, with modularity and social guidance amplifying generalization from imagined to real exploration.
Show all 7 sources
Research formalizes inter-agent thought sharing via sparse autoencoders that recover individual, shared, and private latent thoughts from hidden states. This approach detects alignment conflicts at the representational level before they manifest in language.
A survey framework organizes co-evolving systems into three stages that progressively remove human engineering: dynamic peers first, then adaptive environments and feedback, finally the evolution mechanism itself. Single-entity self-improvement stalls in static contexts; co-evolution supplies adaptive pressure across multiple components.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Thought Communication in Multiagent Collaboration
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- GlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions
- Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
- Language as a Cognitive Tool to Imagine Goals in Curiosity-Driven Exploration