arXiv:2603.21396. v1 submitted 22 Mar 2026; v5 (read, HTML) 10 Jun 2026; no
venue version listed. Macar and Yang are co-first authors (Anthropic Fellows
Program); Wang is at MIT, Wallich at Constellation, and Ameisen and Lindsey
(advising) at Anthropic. Code: https://github.com/safety-research/introspection-mechanisms.
The v1 abstract frames three findings and gives the two elicitation gains as
"53pp" and "75pp". The v5 abstract writes them as "+53%" and "+75% on held-out
concepts", and adds the DPO-versus-SFT result, the two-stage circuit, its
absence in base models and robustness to refusal ablation, and the separation
of identification from detection. Only the abstracts were compared; the v1
body was not read.
A mechanistic follow-up to Lindsey's concept-injection paradigm, run mainly on
Gemma3-27B instruct (injection layer 37 of 62, strength 4, 500 concepts, 100
trials each, GPT-4.1-mini judge), with Qwen3-235B for prompt robustness,
OLMo-3.1-32B Base/SFT/DPO/Instruct checkpoints for training stage, and five
models (Gemma3-27B plus four 8B–32B models) for a logit-proxy replication of
refusal ablation. Detection holds
at 0% false positives across several prompt and dialogue variants. The base
model does not discriminate injection from control (FPR 42.3%, TPR 39.5–41.7%
for strength ≤4). In LoRA ablations on the OLMo SFT checkpoint, contrastive
preference objectives produce discrimination and SFT variants do not (Table 3).
Using Gemma Scope 2 transcoders, the authors trace detection to early
post-injection "evidence carrier" features that suppress later-layer "gate"
features promoting a default "No"; ablating top gates cuts detection from
39.5% to 10.1%.
Refusal-direction ablation raises Gemma3-27B detection from 10.8% to 63.8% at
strength 2, with false positives rising from 0.0% to 7.3%. The abstract's
"+53%" is this percentage-point difference. A learned bias vector at layer 29
raises held-out detection by 74.7 and introspection rate by 54.7 (the paper
writes these with a % sign; the baseline is plotted in Figure 17, not
printed). Appendix S reports that the same vector cuts CoT faithfulness
(MMLU 37.5% to 21.5%, GPQA 44.6% to 19.4%) and prefill detection (36.3% to
16.1%). Figure-only values (per-stage OLMo curves, per-variant TPR/FPR bars,
Figure 17 baselines) are not cited.
The authors state that the results concern one controlled setup and should not
be read as evidence of subjective experience or consciousness. They note it is
hard to distinguish simulated from genuine introspection, that most work is on
one model, and that they did not identify how post-training produces the
circuit.