ValueAI — Open SLWM & Natural Noise
Open implementations and experiment plans for multimodal learning and uncertainty estimation.
Open research infrastructure2026 — Present
Overview
Public SLWM experiments and a study of how small attention perturbations affect model uncertainty.
Technical work
- Built SLWM experimental infrastructure with a GPT-2-small baseline, a multimodal comparator and text/code, audio and visual data pipelines. Registered hypotheses, baselines, ablations and failure cases before performance claims.
- Evaluation plans match data, parameter and compute budgets and measure accuracy, unsupported claims, contradictions, abstention and calibration. Pipeline checks are reported separately from research results.
- Designed Natural Noise to compare GPT-2 Large with and without bounded, non-trainable attention perturbations. Independent full-prefix Monte Carlo passes measure entropy, mutual information, disagreement and Jensen–Shannon divergence.
- The Natural Noise hypothesis remains preregistered; matched multi-seed runs are still needed before drawing conclusions.