LIT-503: First Contact: Unsupervised Human-Machine Co-Adaptation via Mutual Information Maximization

Reddy, Levine and Dragan (2022) — ARXIV-2205.12381, read as NOTE-250.

Key takeaways

• The idea is one sentence and it is a good one. Whatever the user is trying to do, a more intuitive interface produces less noisy commands — so the mutual information I(x_t, (s_t, s_{t+Δ})) between the command and the induced state change scores the interface without knowing the task.
• It works in four domains out of five, across 540K examples of keyboard and eye-gaze interfaces for typing, simulated robot control and games: average Spearman ρ = 0.43 against ground-truth task completion. The abstract says “a variety of domains”; the body says 4 of 5.
• The fifth failed by inverting, and the reason is the most interesting thing in the paper. In the shared-autonomy Lunar Lander data the assistant helps by overriding the user to prevent crashes, which mechanically lowers one-step mutual information — while raising the user’s influence over later states, because they have not crashed. One-step MI gave a strong negative correlation with reward.
• The fix introduces the dependency the method was built to avoid. Generalizing to I(x_t, (s_t, s_{t+Δ})) and setting Δ to the episode length restores a positive correlation — and “choosing the value of Δ requires prior knowledge of the timescale of the user’s desired influence over the system”, which the paper names as “the primary limitation on the generality of our method”.
• Online, it is much stronger than offline. In the 12-participant cursor study the correlation between the MI reward and the true task reward is ρ = 0.87, against 0.43 across domains offline, and the learned interface approaches an oracle that always moves straight to the target.
• What it converges to is worth noticing. Over 12 users the learned perturbation angle θ piles up at two values: no perturbation (θ ≈ 0) and exact inversion (θ ≈ π). Both are consistent mappings. The objective rewards a channel the user can use reliably, which is not quite the “intuitive” the motivating sentence promises.
• Scale is small and the paper says so: the largest interface has 8 parameters, limited by user-study duration and the data efficiency of the Bayesian-optimization RL.

Standing in the anthology

Filed for SOTA-309, and specifically for the horizon result rather than the headline. An objective that inverts when the assistant is doing its job is a measurement trap of the kind this record has been accumulating all day — beside SOTA-305 on reconstruction FID, SOTA-307 on seed variance, and SOTA-308 on recall bought with relevance. This one is the sharpest of the four, because the sign of the correlation flips rather than its size.

The theory is filed as THEORY-091, separately from the practice. The paper explains the score by intuitiveness: a good interface yields less noisy commands. Its own results fit a narrower account. The score measures how reliably the operator’s input determines the outcome at the chosen horizon. That rewards any consistent mapping, including exact inversion, and it predicts the sign flip, which the intuitiveness reading does not.

The domain is assistive interfaces, and under ADR-026 the test is the kind of claim rather than where it was found. The claim — score a human-in-the-loop channel by how much the operator’s input explains the outcome, and pick the horizon — is one the eighteen topics express, and it reaches any setting where a person steers a system and nobody can label what they wanted.