Inferring goals from failure: online Bayesian goal inference for boundedly-rational agents
xuanalogue · x · 2026-10-11
Sharing his PhD work (arXiv:2006.07532), xuanalogue presents an architecture that infers an agent's goals online from both optimal and non-optimal action sequences, modeling agents as boundedly-rational planners that interleave search with replanning. Represented as probabilistic programs, the models support efficient Bayesian inference via SIPS, a sequential Monte Carlo algorithm that incrementally extends inferred plans as actions arrive. Experiments show it beats Bayesian IRL baselines on trajectories with failure and backtracking, generalizing across domains with compositional structure and sparse rewards.
More from Research
- Was RLHF really OpenAI's invention? A debate over the 2017 precedent — binarybits · 2026-10-11
- LLVM Internals Series: Writing a Simple LLVM Pass From Scratch — sh4dy_0011 · 2026-10-11
- Multi-harness RL makes harness selection obsolete, claims zainhas — zainhas · 2026-10-11
- StructureGS-SLAM: structure-aware Gaussian Splatting SLAM with planar instances — rsasaki0109 · 2026-10-11
- Alignment researcher points to Rohin Shah's value learning sequence and IRL model misspecification — xuanalogue · 2026-10-11
- Looped LM paper: 1.6B model matches full-cache baseline with 3x smaller KV cache — rupspace · 2026-10-11