1
We're a team building World-Model-Lens, a library for Interpretability
ProposalGrant
Closes April 29th, 2026$0raised
$500minimum funding
$3,000funding goal
Offer to donate
10 daysleft to contribute
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Project summary
Repo: www.github.com/Bhavith-Chandra/WorldModelLens
World Model Lens is a research‑oriented software toolkit for analyzing, debugging, and understanding world models (e.g., Dreamer‑style RSSMs, JEPA, transformers, video prediction, and robotics/autonomous‑driving world models). It provides observability, replay, causal analysis, safety auditing, and probing tools so researchers can inspect activations, beliefs, uncertainties, and failure modes in a structured way. The project’s main goal is to make world models more interpretable, safer, and easier to debug, especially for teams working on AI safety, RL, planning, and autonomous systems. It standardizes logging, analysis, and benchmarks so that different groups can compare and reproduce findings on world‑model behavior.
PS. Inspired from @NeelNanda's TransformerLens Library, and AgentLens library from MATS Alumni
Team
Bhavith Chandra - Master's Student at NYU
Yugandhar - Master's Student at USC
Rishikesh - Research at Mercity
Wasim - Research at Elemental Labs
What are this project's goals? How will you achieve them?
Observability Infrastructure: Develop activation caching, saliency maps, surprise detection, and belief tracking to deeply inspect internal model behavior.
Replay and Counterfactual Analysis: Enable trajectory replay, intervention replay, imagination branching, and “what‑if” scenario exploration for better understanding of decision processes.
Causal and Mechanistic Analysis: Support causal tracing, circuit discovery, path patching, and bottleneck detection for transparent model reasoning.
Safety‑First Tooling: Create tools for out‑of‑distribution (OOD) detection, hallucination analysis, robustness testing, and safety audits of deployed models.
Standardized Probing and Metrics: Offer unified probing and evaluation via linear and semantic probes (e.g., DINO/CLIP‑style), concept discovery, and disentanglement metrics (MIG, DCI, SAP).
These goals will be achieved by:
Extend Our HookedWorldModel Framework: Expand our existing in‑house HookedWorldModel and adapter system to support additional backends and broader use cases.
Develop Reusable Analysis Modules: Integrate modular components such as a belief analyzer, causal tracer, and safety auditor for flexible research use.
Add Benchmarks and Example Workflows: Provide ready‑to‑run pipelines for reinforcement learning, robotics, autonomous driving, and scientific simulation.
Documentation and Accessibility: Package and document the toolkit so that external researchers can easily plug in their own models and contribute to the ecosystem.
How will this funding be used?
Compute costs for running experiments, benchmarking, and example workflows on real world‑model checkpoints.
We made a team and a salary of 300 per week
How much money have you raised in the last 12 months, and from where?
As we're international students and have beeing using our personal savings but now that we've drained them too testing. Need support from ppl, and posting in this from reading tweets on @ethanjperez ( we're NYU students too ) @RyanKidd ( One of our team member is really preparing his life to get into MATS program )