humanlong/revisable-vlm-memory-results

Revisable VLM memory: artifacts, results, and plans

This dataset repository archives frozen representation artifacts, evaluation reports, causal-intervention outputs, and preregistered execution plans for Yunbo-max/revisable-vlm-memory. It does not redistribute TAP-Vid images or model weights.

Current status (2026-08-18)

The frozen 3B K/V diagnostics previously showed held-out signal for simple visibility and moving/static variables under their original two-frame marked prompt. A subsequent 24-episode true-DynamicCache intervention run used a different three-image causal prompt. On that exact interface, the frozen probes did not pass the preregistered competence gate, so the causal result is not interpretable as evidence for or against semantic KV revision.

The current scientific conclusion is therefore deliberately narrow:

The existing probe bundle is not validated on the current causal cache interface. No belief-revision claim is supported by this run.

See:

• dynamic_state_diagnostic_results.md for the original frozen-probe results;
• kv_dynamic_state_causal_results.md for the earlier legacy causal artifact;
• true_kv_revision_2026-08-18.md for the audited true-KV run;
• aligned_interface_recovery_plan.md for the next fixed experiment.

Raw outputs for the new run are under outputs/. Artifact filenames ending in _v2 use the two-times-per-track diagnostic sampling protocol.