# Intrinsic Discovery (Induction Labs)

> Source: https://aiwiki.ai/wiki/intrinsic_discovery
> Updated: 2026-08-28
> Fact-checked: 2026-08-28
> Categories: AI Agents, AI Benchmarks, AI Research, Reinforcement Learning, World Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Intrinsic Discovery (Induction Labs)." aiwiki.ai, 28 Aug 2026. https://aiwiki.ai/wiki/intrinsic_discovery
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Intrinsic Discovery** is an experimental training method introduced by Induction Labs in August 2026. It uses an intrinsically rewarded [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) policy to generate terminal interactions, then trains a [world model](https://aiwiki.ai/wiki/world_model) on the accumulated command and observation pairs. The company presented the work as early research rather than a released product or a generally capable agent.[1]

Induction Labs tested the method with three base [large language models](https://aiwiki.ai/wiki/large_language_model): gpt-oss-20b, Qwen3.6-35B-A3B, and gpt-oss-120b. Its reported evaluations concern prediction of terminal output, not successful completion of software tasks. The company released a modified 297-example evaluation dataset, AgentWorldBench-Terminal-V2, but did not release the discovery implementation, generated training data, trained weights, or a paper describing the experiments.[1][2]

## Method

Intrinsic Discovery separates exploration from the final model being evaluated. At each training step, an exploratory policy samples a terminal action and the environment returns an observation. The method assigns a novelty reward

`r(x_t) = 1 - max sim(x_t, x)`

where the maximum compares the new observation with observations already stored in an archive. An observation that is less similar to the archive receives a larger reward. Induction Labs says it updates the policy with group relative policy optimization, adds new observations to the archive, and repeats the process. Because rewarded observations enter the archive, repeating them becomes less valuable in later iterations.[1]

The output of this process is the accumulated interaction data, not the exploratory policy itself. Induction Labs used the generated command-observation pairs to train a world model initialized from the same base model. It says this second stage used reinforcement learning with verifiable rewards. Each of the three base models underwent its own discovery and world-model training process.[1]

The release does not identify the representation or model used by the similarity function, its calibration, the archive's storage or retrieval procedure, the group relative policy optimization settings, the number of training updates, or the total compute used. It also does not provide the reward implementation. These omissions prevent an independent reproduction of the method from the release alone.[1]

## Relationship to prior research

Novelty and surprise rewards are established approaches to exploration in reinforcement learning. Surveys of intrinsic motivation distinguish novelty-based rewards from surprise, information gain, and skill-learning objectives, and describe how the choice of state representation affects what an agent treats as new.[5][6] Intrinsic Discovery applies a novelty reward to language-model interactions in a terminal and uses the resulting experience to train a language world model. The release does not establish that novelty-based exploration itself originated with Induction Labs.

The evaluation builds on AgentWorldBench from the Qwen-AgentWorld project. Qwen-AgentWorld defines language world modeling as prediction of an environment's next state from an interaction history and evaluates models across terminal, web, operating-system, software-engineering, Android, search, and Model Context Protocol domains. Intrinsic Discovery was tested only in the terminal domain.[3][4]

## Experimental setup

Induction Labs compared discovery policies with frozen policies based on the same starting models and described the data budgets as matched. It measured exploration diversity by counting unique shell commands and Python libraries across 20,000 executed commands. The company reported roughly three to five times as many unique items for its Qwen3.6-35B-A3B and gpt-oss-120b experiments, and summarized the overall result as up to five times more diverse states.[1]

Those counts are proxies rather than direct measurements of meaningful state coverage. Induction Labs notes that they can be gamed, for example by issuing random commands. The published page does not disclose its command-normalization rules, library-identification method, uncertainty estimates, number of repeated runs, or complete numerical series. The diversity result is therefore an author-reported observation under the company's measurement procedure.[1]

The company also reported an internal ablation with gpt-oss-20b. On an unpublished evaluation called TerminalWorldBench, training on discovery data reached 64.8 while a static-policy data baseline plateaued near 50.9. Because the evaluation set and full protocol were not released, this comparison cannot be independently checked.[1]

## AgentWorldBench-Terminal-V2

AgentWorldBench evaluates a model's predicted environment observation using an LLM judge. The original Qwen-AgentWorld protocol scores format, factuality, consistency, realism, and overall quality, and used GPT-5.2 as the judge.[3][4] These scores measure simulated next-state fidelity. They are not measures of whether an autonomous agent completed a terminal task, wrote correct software, or operated safely.

Induction Labs did not evaluate against the original terminal split unchanged. The company said some original examples required outputs determined by environment details missing from the visible session history. It created AgentWorldBench-Terminal-V2 by replaying sessions in fresh Linux sandboxes, regenerating outputs, and filtering examples it considered invalid. The public test split contains 297 rows and retains the original schema. Its dataset card says each session begins with `date -u` and uses timestamps from the replay.[1][2]

The company otherwise followed the original evaluation methodology but replaced GPT-5.2 with GPT-5.6 Luna at xhigh reasoning effort as the judge. This change, together with regeneration and filtering, means scores from Terminal-V2 should not be directly compared with scores reported for the original terminal split unless both models were evaluated under the same modified protocol.[1][2][4]

## Reported results

All results in this section were reported by Induction Labs. No independent reproduction was available as of August 2026.[1]

| Model evaluated on AgentWorldBench-Terminal-V2 | Reported score |
| --- | ---: |
| Claude Opus 5, xhigh | 85.2 |
| Terminal-35B-A3B | 84.1 |
| GPT-5.6 Sol, xhigh | 83.5 |
| Qwen3.6-35B-A3B base model | 76.4 |
| Terminal-120b | 79.7 |
| gpt-oss-120b base model | 74.6 |

Terminal-35B-A3B is Induction Labs' post-trained derivative of Qwen3.6-35B-A3B, while Terminal-120b is its derivative of gpt-oss-120b. Under the company's modified benchmark and judge, Terminal-35B-A3B scored above GPT-5.6 Sol and below Claude Opus 5. The comparison is limited to terminal next-observation prediction under that protocol. It does not establish a general model ranking.[1]

The paired results show increases of 7.7 points for the 35B-A3B experiment and 5.1 points for the 120b experiment relative to their respective base models. Induction Labs attributes these differences to training on self-discovered interactions. The release does not report confidence intervals, statistical tests, multiple training seeds, or a matched alternative-data ablation on the public Terminal-V2 benchmark.[1]

## Availability and evidence limits

AgentWorldBench-Terminal-V2 is available through [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the Apache License 2.0. Its 297-row test split allows researchers to inspect the regenerated prompts and target observations and to use the original AgentWorldBench evaluation schema.[2] The upstream Qwen-AgentWorld project separately publishes its paper, benchmark, evaluation code, and model weights under open licenses.[3][4]

Induction Labs had not released the Intrinsic Discovery training code, exploration rollouts, world-model checkpoints, similarity model, filtering log, complete evaluation outputs, prompts, hyperparameters, or compute accounting as of August 28, 2026.[1][2] The public dataset therefore supports inspection of part of the evaluation, but not reproduction of the discovery training or reported model scores.

The experiments cover one virtual environment type and use an LLM judge. Regenerating and filtering the benchmark may correct under-specified examples, but it also creates a company-defined test set. Judge errors, filtering choices, data contamination, and sensitivity to the novelty representation remain possible sources of bias. The release does not evaluate real-world operation, transfer beyond terminal simulation, agent safety, or whether the method improves task-solving agents. Claims about broader autonomy or intelligence remain research goals rather than demonstrated results.[1]

## References

1. David Li and Jonathan Li. "Intrinsically Curious Agents." Induction Labs, August 27, 2026. https://www.inductionlabs.com/news/intrinsic-discovery
2. Induction Labs. "AgentWorldBench-Terminal-V2." Hugging Face dataset card, accessed August 28, 2026. https://huggingface.co/datasets/inductionlabs/AgentWorldBench-Terminal-V2
3. Yuxin Zuo, Zikai Xiao, Li Sheng, et al. "Qwen-AgentWorld: Language World Models for General Agents." arXiv:2606.24597, June 23, 2026. https://arxiv.org/abs/2606.24597
4. Qwen Team. "Qwen-AgentWorld." GitHub repository and evaluation code, accessed August 28, 2026. https://github.com/QwenLM/Qwen-AgentWorld
5. Arthur Aubret, Laetitia Matignon, and Salima Hassas. "A Survey on Intrinsic Motivation in Reinforcement Learning." arXiv:1908.06976, 2019. https://arxiv.org/abs/1908.06976
6. Arthur Aubret, Laetitia Matignon, and Salima Hassas. "An Information-Theoretic Perspective on Intrinsic Motivation in Reinforcement Learning: A Survey." Entropy 25, no. 2 (2023): 327. https://doi.org/10.3390/e25020327