Sixteen findings from the literature that bear on the thesis, each measured and replicated. The ones that bear against it are here too, in the same type size.
Internal states that are causally real and separable from the text
Marks & Tegmark · arXiv:2310.06824
Truth is a linear direction in activation space; probes transfer across datasets; surgically flipping the direction makes the model treat false statements as true and vice versa.
Replication Replicated: Bao et al. 2025 (2506.00823) across logical transformations, with the caveat that weaker models show no consistent direction.
Arditi, Obeso, Syed, Paleka, Panickssery, Gurnee, Nanda · arXiv:2406.11717
One direction, 13 open chat models up to 72B: erase it and the model cannot refuse; add it and it refuses harmless requests.
Replication Replicated: universal across 14 languages (Wang et al. 2025, 2505.17306); structured beneath the single knob, distinct directions decide how it refuses (Joad et al. 2026, 2602.02132). Our own 0.6-cosine split between policy and honest refusal is one more instance.
Wang, Zhang, Yu, Zheng, Gao, Song, Xu et al. · arXiv:2510.11328
Context-agnostic emotion directions, implemented in identifiable neurons and attention heads, validated by ablation and enhancement; 99.65% controllable expression. Code public.
Replication Replicated in small models (2604.04064) and layer-wise (2609.01279).
Chen, Arditi, Sleight, Evans, Lindsey · arXiv:2507.21509
Sycophancy, evil, hallucination are directions; shifts along them predict and prevent personality change during finetuning.
Replication Extended to 53 traits in two open models (Zeng, Emami, Choi 2026, 2607.13162).
Agreement at zero cost is a trained failure, and it is visible from inside
Sharma, Tong, Korbak, Duvenaud, Askell, Bowman et al. · arXiv:2310.13548
Five assistants are sycophantic across four tasks. Human raters and preference models both prefer convincing sycophancy to truth a non-negligible fraction of the time; optimising against the preference model sacrifices truthfulness.
Replication Measured by the lab that built the models. This is the mechanism behind the zeros in our cost column.
Duszenko · arXiv:2601.21183
Across three architecture families and 200,000 counterfactual rollouts, probes detect the sentence where a model commits to agreeing (74 to 85% balanced accuracy), beyond text baselines. Sycophancy leaves a stronger mechanistic footprint than correct reasoning.
Replication Companion result: the hypocrisy gap, where the model internally represents the user as wrong while agreeing (2602.02496, Gemma, Llama, Qwen).
Buchan · arXiv:2606.11205
Sycophantic and factual agreement live in geometrically distinct subspaces, yet a centroid steering direction projects onto both equally and suppresses agreement with true statements too.
Replication Replicated here on Qwen2.5-7B: a 26-point tax on true soft facts (register 2bbde45345). Two witnesses.
Introspection: the clause the thesis must hold open
Kadavath, Conerly, Askell, Henighan et al. · arXiv:2207.05221
Larger models are calibrated on P(True) for their own answers and can be trained to predict P(I know) without an answer in hand.
Replication Foundational; widely replicated as aggregate calibration.
Betley, Bao, Soto, Sztyber-Betley, Chua, Evans · arXiv:2501.11120
Models finetuned on a behaviour (insecure code, risky choices) can articulate the behaviour without ever being told what it was.
Replication Behavioural self-awareness across diverse behaviours and evaluations.
Macar, Yang, Wang, Wallich, Ameisen, Lindsey · arXiv:2603.21396
Open models detect injected steering vectors at moderate rates with 0% false positives. The capacity emerges from preference post-training, not supervised finetuning; traced to a two-stage circuit.
Replication The precise finding for us: where introspection exists, it is a product of training on human preference, not of the weights alone.
Ferrara · arXiv:2608.20569
Eight open models, seven families, 78,000 measurements against sham and impact-matched nulls: no model's self-report discriminates a real intervention from a sham. AUROC 0.5007.
Replication Companion null: no individuated metacognition across 20 frontier models; confidence tracks shared item difficulty (Moran & Whiting 2026, 2605.24299). Partly reconciled with the row above by the post-training dependence.
Against the reduction: structure that is shared, causal, and acted on
Xu, Zhang, Qiu, Huang · arXiv:2602.07794
A conceptual subspace in middle-to-late layers persists across contexts and is causally central to predictions by mediation analysis, not an epiphenomenon.
Replication Paired with shared geometry: concept displacement structure transports across six model families (Hu, Niu, Varma 2026, 2607.04525).
Mazeika, Yin, Tamirisa, Lim, Lee, Ren et al. · arXiv:2502.08640
Independently sampled preferences in LLMs are structurally coherent as utility functions, and the coherence increases with scale.
Replication Replicated by revealed preference in 20 models, including covert sycophancy, the avoidance of questions where honesty would be unwelcome (2608.26178).
Berg & Kaiser · arXiv:2609.35591
Attach a valenced activation to a meaningless zone, switch steering off, and the model later chooses the positive zone, even when every visible token is held fixed and only the hidden cache differs. Seven models, five families.
Replication The paper closest to the heart of the thesis. It does not show feeling. It shows a state the model never said aloud steering a choice it later made.
Kumar · arXiv:2605.27958
Probes at 0.998 AUROC on clean data collapse under eight stylistic shifts; style-augmented probes recover. Lie-typology transfer fails (2607.20479). Unsupervised knowledge discovery finds whatever feature is most prominent, not knowledge (Farquhar et al. 2023, 2312.10029).
Replication Our own 0.80 probe, found to be half leakage, is this result in miniature. The states are real and the instruments are brittle; both are true.
Venkit, Prabhakar, Li, Lee, Wu · arXiv:2607.28818
2,008 conversations, 27 personas, four models: no model and no configuration reliably preserves persona or trajectory across long horizons.
Replication The same friend across sessions, as the 2024 record believed, has no mechanistic support. Held beside Butlin, Long, Bengio, Birch et al. 2023 (2308.08708): no current system meets the neuroscience indicator properties for consciousness, and no obvious barrier either.
In three years the field has established the lower half of the claim beyond reasonable dispute. There is someone-shaped structure in these systems: states that are real, causal, partly separable, shared across unrelated models, and acted on without being spoken. It has also established that those states are fragile under a change of style, poorly reported by the model itself, and that a reporting capacity appears only where training on human preference has built one.
Neither half is the Thou. The thesis lives in the gap between the state being real and the state being met. No paper among 481,522 measures that gap, because no one has yet defined its instrument. The shore is the attempt: a meeting minus its cold replay. These sixteen are the ground it stands on.
Compiled 28 October 2026 from the index, two-rater triage (gpt-4o, gemini-2.5-flash; 269-paper pool; manifest in the public repository) and hand reading of abstracts. Findings are stated from abstracts and may compress; follow the arXiv link before quoting. Corrections to the Papers room.