Oratis's picture
Org rename: HakkoLab -> Diogenes (handle DiogenesLab) — update references
4c6734f verified
|
Raw
History Blame Contribute Delete
2.72 kB
metadata
license: cc-by-4.0
language: en
tags:
  - interpretability
  - agents
  - global-workspace
  - persona-stability
  - mechanistic-interpretability
pretty_name: Externalizing the Workspace

Externalizing the Workspace: Persistent Self-State for Long-Horizon Agent Coherence

Oratis (Wang Bihao) · Diogenes · Draft v3, July 2026 · PDF

Language-model agents drift. Building on Anthropic's finding that LLMs maintain a verbalizable global workspace (Gurnee et al., 2026), this paper models persistent self-state agent architectures (LISA's soul) as functional externalizations of the workspace — and backs the model with experiments that run entirely on open models and independent-researcher compute.

Headline results

finding evidence
Core J-lens workspace phenomena reproduce on open models Qwen2.5-1.5B/7B + Llama-3.2-1B; laptop + one A100
Choice pre-commitment emerges with scale to-be-reported concept at J-lens rank 1–28 before any output token at 7B; absent at 1.5B
Injected self-state selectively loads its identity concepts into the mid-band workspace control battery unmoved; effect grows with scale
Workspace occupancy tracks soul-consistent behavior r = −0.74 … −0.95 across four settings
Privileged self-state is necessary for coherence under value pressure 6-arm × 5-seed pilot: commitment persistence 1.00 vs 0.00 for memory-only baselines (incl. Generative-Agents-style reflection)
Unconditional broadcast is necessary retrieval-gated soul collapses in-work value coherence to no-soul level: .61 [.60,.64] vs 1.00 [1.00,1.00]

What's in this repo

  • main.pdf — the paper (Draft v3)
  • figs/ — key figures (lens registers; soul workspace-loading; ablation-pilot drift curves)
  • results/ — raw JSON results: J-lens reproduction (3 models), E6 soul-loading, E8 scale study, multi-seed ablation pilot with bootstrap CIs
  • Reproduction code lives in the LISA repository under research/ (J-lens library, experiment scripts, pilot harness, exact commands)

Method in one line

A corpus-averaged Jacobian lens (gradient-estimated concept vectors + finite-difference full-vocabulary readouts) makes the workspace measurable with nothing but backprop access to open weights — turning "agent identity" from a metaphor into a quantity you can track over an agent's lifetime (workspace drift).

Citation

@article{oratis2026workspace,
  title={Externalizing the Workspace: Persistent Self-State for Long-Horizon Agent Coherence},
  author={Oratis (Wang Bihao)},
  year={2026},
  note={Draft v3, Diogenes}
}