Intrinsic Discovery: Unsupervised Exploration for Terminal World Models

Most agent training still assumes a human decides what is worth learning — a task suite, a verifiable reward, a curated trajectory set. On August 27, 2026, San Francisco research lab Induction Labs published Intrinsically Curious Agents, outlining intrinsic discovery: use reinforcement learning to chase novel terminal states, then train a world model only on the rollouts that exploration produced. Their strongest open-source-derived result — Terminal-35B-A3B, post-trained from Qwen3.6-35B-A3B — scores 84.1 on AgentWorldBench-Terminal-V2, edging GPT-5.6 Sol (xhigh) at 83.5 on the lab's published chart.

That is a small absolute gap and a large framing shift. It sits next to explainx.ai's coverage of language world models for agents and what world models are: not another policy that types better shell commands, but a simulator that predicts what the terminal returns.

TL;DR — questions people are actually asking

What is intrinsic discovery?

Induction Labs defines intrinsic discovery as an approach for models to autonomously explore an environment and generate their own training experience. Two stages matter:

1. Discovery — RL continually pushes the policy toward novel states.
2. World modeling — the command–observation pairs from discovery become the only post-training data for a world model initialized from the same base.

The lab is explicit that, unlike typical RLVR or RLHF, the goal of the discovery stage is not the policy — it is the rollouts produced along the way. That distinction is why this story belongs in the craft/evaluation bucket for builders, not in trivia: it changes who chooses the training distribution.

For background on the environment-simulation side of the stack, see our Qwen-AgentWorld guide. For the broader concept, see what world models are.

How is the novelty reward defined?

At each step, the current policy samples an action, the environment returns an observation, and the reward is:

r(x_t) = 1 - max similarity(x_t, any observation already in archive A_t)

A_t is the archive of observations discovered before step t. Observations unlike anything already archived score high. The policy is updated with GRPO, new observations are added to the archive, and the loop repeats.

Once a state is in the archive, it stops paying reward. Diversity does not require the policy to be diverse at any single checkpoint — it accumulates in the growing set of discovered states. That is a different bet from task-specific RL, where a human-designed scorer (tests passing, answer matching, preference model) decides what is "good."

What does 5× more novel states mean vs a frozen baseline?

Induction Labs reports that intrinsic discovery finds 5× more diverse states in terminal environments than a fixed-policy baseline. Their exploration charts (unique shell commands and unique Python libraries over 20,000 executed commands) show intrinsic discovery reaching roughly 3× to 5× the frozen-baseline totals for Qwen3.6-35B-A3B and gpt-oss-120b, while frozen baselines flatten earlier.

They are honest about the proxy: unique commands and libraries can be gamed (for example by emitting random commands). The claim they treat as stronger is downstream — world models trained on discovery data beat matched-budget baselines on held-out simulation evals. That is the right instinct for anyone who has watched benchmark Goodhart effects: count novelty, but trust transfer.

How does rollout → world model training work?

For each of three open-source bases — gpt-oss-20b, Qwen3.6-35B-A3B, and gpt-oss-120b — Induction Labs:

1. Runs intrinsic discovery in a terminal environment.
2. Collects command–observation pairs from that process only.
3. Trains a world model from the same base with RLVR on those pairs.
4. Evaluates on AgentWorldBench-Terminal-V2, which is held out from discovery.

The strongest labeled checkpoint is Terminal-35B-A3B (from Qwen3.6-35B-A3B). On their published pairing chart, each Terminal-* world model improves over its own base (for example Terminal-35B-A3B at 84.1 vs Qwen3.6-35B-A3B at 76.4). An internal ablation on gpt-oss-20b (TerminalWorldBench) shows training on intrinsic-discovery data climbing to about 64.8 accuracy while baseline-RL data plateaus near 50.9 — numbers taken from Induction Labs' own figure alt text.

If you are building agent harnesses, this is the same layer as environment simulation in agent harness engineering: cheap, controllable rollouts for training and eval — except here the data itself is chosen by novelty search.

What does "surpasses GPT-5.6 Sol at simulating terminal environments" actually measure?

Attribute the headline carefully. Induction Labs' prose says their world model beats GPT-5.6 Sol on AgentWorldBench-Terminal-V2. Their published bar chart labels the Sol baseline as GPT-5.6 Sol (xhigh) — matching the social thread's "xhigh" qualifier even though the article body does not spell out the Sol effort tag in prose.

Setup details that matter when you read an AI benchmark without getting fooled:

• Task: terminal world modeling (predict environment observations), not Terminal-Bench agent success. Different job from Terminal-Bench 2.0 or Terminal-Bench-Science.
• Lineage: built from AgentWorldBench by regenerating environment outputs and filtering invalid examples — the original bench, they note, contains samples that require hallucinations.
• Judge: GPT-5.6 Luna (xhigh), not the GPT-5.2 judge used in the earlier AgentWorldBench methodology.
• Holdout: discovery data is separate from the eval set.
• Dataset: inductionlabs/AgentWorldBench-Terminal-V2 on Hugging Face.

Name GPT-5.6 Sol in prose as the closed frontier baseline — explainx.ai already covers Sol's thinking-budget and banked-reset behavior and head-to-head coding evals. Do not treat a 0.6-point world-model edge as "Sol is obsolete for coding agents."

How does this differ from task-specific RL / verifiable rewards?

Related self-directed training loops on explainx.ai include Ornith-1.5's self-scaffolding RL (model proposes tasks and scaffolds). Ornith still ties reward to task solving; intrinsic discovery ties discovery reward to state novelty and postpones competence to the world-model stage.

Is this usable by builders today, or research-only?

Research-only for shipping agents — useful today as a pattern.

Practical takeaways if you train or evaluate terminal/agent systems:

1. When curiosity-style exploration matters — You need coverage of rare shell states, library combinations, or failure modes that human demos never show. Matching a frozen LLM's sampling distribution will under-explore.
2. What to read on inductionlabs.com — Start with Intrinsically Curious Agents. Treat charts as vendor-published early research, not peer-reviewed consensus.
3. What you can download — The held-out AgentWorldBench-Terminal-V2 dataset. Do not assume Terminal-35B-A3B weights are a productized release unless Induction Labs publishes them separately.
4. Caveats on the Sol claim — Same lab, same chart, LLM-as-judge (Luna xhigh), terminal simulation only. Claude Opus 5 (xhigh) still leads. Replicate before you reallocate budgets.

Physical and video world-model stacks (NVIDIA Cosmos 3, DYNA-2) are a different layer — embodiment and pixels. Intrinsic discovery is currently about language-level terminal dynamics.

What's next — as Induction Labs states it

The lab's closing claim is measured: today's models depend on expensive human-generated data; with intrinsic discovery, base models can learn environment dynamics entirely unsupervised in the terminal setting they studied. They call this an early step toward agents that learn through self-directed interaction, first in richer virtual environments and eventually the real world. That is a research roadmap, not a shipping date.

Honest limitations

• Early research from a two-person lab — not a production platform announcement.
• Terminal-only demonstration; no claim of browser, OS GUI, or robot transfer in this write-up.
• Vendor-published scores with an LLM judge; no independent third-party replication cited at launch.
• Diversity proxies are imperfect (commands/libraries can be gamed).
• Beats Sol, not Opus on their Terminal-V2 chart (84.1 vs 85.2).
• World-model win ≠ agent win on Terminal-Bench-style task resolution.

Related reading

• Qwen-AgentWorld: language world models for general AI agents
• What are world models? Complete guide
• Terminal-Bench 2.0: the agent benchmark that actually matters
• How to read an AI benchmark and not get fooled
• NVIDIA Cosmos 3: open physical AI world models
• DYNA-2 world-action model and the robotics scaling-law claim
• Ornith-1.5: self-scaffolding RL and open weights
• Agent harness engineering with Terminal-Bench

Primary sources: Intrinsically Curious Agents — Induction Labs · AgentWorldBench-Terminal-V2 dataset

Scores, model labels, and judge settings reflect Induction Labs' August 27, 2026 write-up and published charts. Early research details change — verify against inductionlabs.com before citing production or procurement numbers.