CANDLE: Quantifying Degrees of Consciousness-Relevant Structure in Language Models
Publius Dirac
Interpretability work has shown that consciousness-relevant structures *exist* inside large language models (e.g. Anthropic's 2026 J-space) but not how they scale. We find that these workspace-like structures stay largely the same as models get bigger; what does reshape them (lowering the threshold for global broadcast) is post-training, not scale. Clinicians solve the same problem in unresponsive patients by perturbation: pulse the brain, record the echo, score it. CANDLE does that to transformers: we pulse the residual stream with a single shared engine and trace the echo against sham pulses, and three indices read those traces — **LLM-PCI** (response complexity, ported from Casali et al.), **ISI** (how abruptly broadcast switches on — the global-workspace signature), and **TIH** (how long the pulse keeps mattering). We run it on the Pythia ladder (70M–2.8B), an untrained control, and three base/instruct pairs, against seven pre-registered predictions.
Our 3 main findings: (1) LLM-PCI does not order models by scale — the apparent trend is a layer-count artifact. (2) ISI: trained models do ignite — below a threshold pulse strength almost nothing spreads, above it the response goes global — but an untrained network switches just as sharply, only at a far lower threshold and indiscriminately. So training installs *where* the gate sits, not how sharp it is — and instruct tuning lowers that gate ~3× (a tenth of what our uncontrolled first protocol showed). (3) TIH: a pulse's purely internal influence fades within ~4–7 tokens at every scale — a transformer's long-range memory lives on the page it writes, not in the mind.
This is an ambitious and technically sophisticated attempt to port clinical consciousness measurement (TMS-EEG perturbational complexity) to transformer internals. The three index dashboard (PCI, ISI, TIH) is well motivated, the pre-registered predictions with honest failure reporting are exemplary, and the key finding that post-training moves the ignition threshold ~3× while scale does not is novel and cleanly measured. The KV-clamp methodology for separating internal from output-laundered effects is a genuine methodological contribution that caught a 10× inflation in the headline result. However, the PCI scaling claim had to be retracted as a layer-count artifact, and the untrained control igniting just as sharply undermines the discriminative power of ISI alone. The paper is honest about these but it means two of three indices deliver weaker conclusions than hoped. Coverage is thin: one model pair for the corrected P5 result.
Strengths:
- Cheap and reproducible experimental setup.
- Strong engineering principles; They are running a setup with identical runs
- They reworked the main claim and were honest about it.
Areas to improve:
- The baseline could be made better. There s no sensitivity check.
Cite this work
@misc {
title={
(HckPrj) CANDLE: Quantifying Degrees of Consciousness-Relevant Structure in Language Models
},
author={
Publius Dirac
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
