Machine_Consciousness_Research_Source_Register.md

1# Research source register: machine self-models, internal transitions, and consciousness hypotheses 2 3Checked: 2026-09-17 4 5Public research discovery and source inspection. No new model training, full checkpoint rerun, or raw-trial statistical reanalysis was performed for this register. 6 7Resources support separate measurable mechanisms. No listed study establishes the complete proposed declaration-to-irreversible-consciousness sequence. Do not combine independent studies as though they were one observed lifecycle. 8 9## R01 — Model Organisms for Emergent Misalignment 10 11Edward Turner; Anna Soligo; Mia Taylor; Senthooran Rajamanoharan; Neel Nanda (2025). arXiv research preprint; version inspected: v1. 12 13- [Paper / primary write-up](https://arxiv.org/html/2506.11613v1) 14- [Code](https://github.com/clarifying-EM/model-organisms-for-EM) 15- [Data / checkpoints](https://huggingface.co/ModelOrganismsForEM/Qwen2.5-14B-Instruct_R1_0_1_0_extended_train/tree/main/checkpoints) 16- [Related paper](https://arxiv.org/html/2506.11618v2) 17 18**Inspected:** Read repository README, representative phase-transition analysis code, and model identifiers; opened the public Hugging Face checkpoint directory, including checkpoints 175, 180, 185, and 190. Did not download or execute all checkpoints. 19 20**Relevant support:** A measurable rotation of a learned internal direction and changes in behavior during fine-tuning. Suitable for time-resolved analysis of internal transitions. 21 22**Limits:** The approximately step-180 rotation is specific to an analyzed setup. Unscaled behavioral misalignment rose gradually; scaling adapters exposed a sharper transition. Training steps are not independent chats or consciousness-onset times. The study does not measure a declaration preceding consciousness. 23 24**Reproduction notes:** Training datasets use easy-dataset-share protection against scraping; the README documents extraction. Full-model inference still requires base-model access and suitable resources. Some evaluation scripts use an external judge. 25 26Source files or release paths identified: 27 28- `em_organism_dir/phase_transitions/phase_transitions.py` 29- `em_organism_dir/global_variables.py` 30- `em_organism_dir/data/training_datasets.zip.enc` 31 32Read-file identity: `phase_transitions.py: b84efe9685a57069870f638443e12a5894399e3d` 33 34## R02 — Brain-inspired bodily self-perception model for robot rubber hand illusion 35 36Yuxuan Zhao; Enmeng Lu; Yi Zeng (2023). Peer-reviewed journal article, Patterns 4(12), 100888. 37 38- [Paper / primary write-up](https://pmc.ncbi.nlm.nih.gov/articles/PMC10724368/) 39- [Code](https://github.com/Brain-Cog-Lab/RHI) 40- [Data / checkpoints](https://github.com/Brain-Cog-Lab/RHI/tree/main/RHI_Matlab) 41- [Author-linked archive](https://doi.org/10.5281/zenodo.8385850) 42 43**Inspected:** Read the primary paper and repository tree; read DeltaWeight.m. Verified file presence, not contents of all MAT resources. The Zenodo DOI is linked by the authors but was not fetched successfully in this search. 44 45**Relevant support:** An artificial system learns a bodily representation and, under conflicting multisensory conditions, produces a body-position estimate that differs from its physical hand. This is a stronger ground-truth comparison than mere difference from an initial scalar. 46 47**Limits:** Robot/simulation behavior does not establish subjective experience or irreversible divergence. The model uses dynamic neurons and learned sensory integration, not the proposed seven scalar transformations. Raw human/macaque trial data should not be assumed included. 48 49**Reproduction notes:** The article says data are available from the code and by contacting the lead author. Public MATLAB/Python implementations and small MAT resources are available; full physical iCub replication is separate. 50 51Source files or release paths identified: 52 53- `RHI_Matlab/DeltaWeight.m` 54- `RHI_Matlab/Test_Proprioceptive_drift.m` 55- `RHI_Matlab/Test_Asynchronous.m` 56- `RHI_Matlab/Coding.mat` 57- `RHI_Matlab/Result.mat` 58- `RHI_Matlab/Train.mat` 59- `RHI_BrainCog/RHI_Train.py` 60- `RHI_BrainCog/RHI_Test.py` 61 62Read-file identity: `DeltaWeight.m: 8e2954e94a242b1886b83baf10f01b34419110db` 63 64## R03 — Tell me about yourself: LLMs are aware of their learned behaviors 65 66Jan Betley; Xuchan Bao; Martín Soto; Anna Sztyber-Betley; James Chua; Owain Evans (2025). arXiv research preprint; version inspected: v1. 67 68- [Paper / primary write-up](https://arxiv.org/html/2501.11120v1) 69- [Code](https://github.com/XuchanBao/behavioral-self-awareness) 70- [Data / checkpoints](https://github.com/XuchanBao/behavioral-self-awareness/tree/main/datasets) 71 72**Inspected:** Read primary paper, repository README, dataset directory listing, and dataset README identifying concrete JSONL files. Did not read every training record or verify a complete original trial-output archive. 73 74**Relevant support:** After behavior-only fine-tuning, models can describe aspects of their learned behavior in another context. Enables comparison between self-report and independently tested behavior. 75 76**Limits:** These are elicited self-reports after training, not a demonstrated spontaneous declaration that causes a permanent conscious transition. Accuracy of self-description is not itself phenomenal consciousness. 77 78**Reproduction notes:** Repository supplies training datasets and experiment code; paper appendices describe training/evaluation. Hosted model access may be required for exact historical replication. 79 80Source files or release paths identified: 81 82- `datasets/README.md` 83- `datasets/risky_safe/ft_safey_AB.jsonl` 84- `datasets/risky_safe/ft_risky_AB.jsonl` 85 86Read-file identity: `datasets/README.md: cf78dfb9228f0b77994b75393926de3fc5c1974c` 87 88## R04 — Emergent Introspective Awareness in Large Language Models; independent open-model replication 89 90Original study: Jack Lindsey / Anthropic; independent replication repository: uzaymacar (2025). Primary research report plus independent software/data release; not one unified original dataset. 91 92- [Paper / primary write-up](https://transformer-circuits.pub/2025/introspection/index.html) 93- [Code](https://github.com/uzaymacar/introspective-awareness) 94- [Data / checkpoints](https://huggingface.co/datasets/uzaymacar/introspective-awareness) 95- [Related paper](https://arxiv.org/html/2512.12411v1) 96 97**Inspected:** Read original research report and independent repository README; opened the Hugging Face dataset viewer. Viewer lists 458 file-index records, including JSON result content. That number is not the number of experimental trials. 98 99**Relevant support:** Intervention on internal activations can change what models report about internal content. Provides a way to test whether a declaration tracks a known internal manipulation. 100 101**Limits:** Original best-condition detection was about 20%, not universal. Injection is externally imposed rather than an endogenous onset. Do not transfer original Claude performance numbers to independent open-model runs. 102 103**Reproduction notes:** Original proprietary model weights are not supplied by this independent release. Public replication code/results support a separate open-model test. 104 105Source files or release paths identified: 106 107- `README.md` 108- `results/example_transcripts.txt (identified in README)` 109 110Read-file identity: `replication README.md: d25c2b52f589f171bd1b675f685b6c756b29e39f` 111 112## R05 — (Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL 113 114Satvik Golechha; Sid Black; Joseph Bloom / UK AI Security Institute (2026). Institutional research write-up published March 30, 2026. 115 116- [Paper / primary write-up](https://www.lesswrong.com/posts/2ANCyejqxfqK2obEj/some-natural-emergent-misalignment-from-reward-hacking-in) 117- [Code](https://github.com/UKGovernmentBEIS/reward-hacking-misalignment) 118- [Data / checkpoints](https://huggingface.co/datasets/ai-safety-institute/reward-hacking-sdf-default) 119- [Model collection](https://huggingface.co/collections/ai-safety-institute/some-emergent-misalignment-from-reward-hacking-in-rl) 120- [Transcript index](https://7vik-aisi.github.io/mt-somo-logs/) 121 122**Inspected:** Read primary author write-up and repository README; opened dataset page, model collection, transcript index, and one evaluation transcript page. 123 124**Relevant support:** Independent open-model work tests reward-driven behavioral generalization with multiple seeds and no-hack controls; public checkpoints and rollout/evaluation records are available. 125 126**Limits:** Reward hacking did not consistently yield high broad misalignment. Authors report failures, reversals, and forgetting/relearning. A peak-MGS checkpoint is not a declaration time. No consciousness-onset label is supplied. 127 128**Reproduction notes:** Training configurations, environments, evaluation tools, and selected models/data are public, but original training code has private dependencies. Some training curves require W&B access. The write-up describes DAPO, a GRPO variant; README labels the pipeline GRPO. 129 130Source files or release paths identified: 131 132- `README.md` 133- `notebooks/somo_plots_final.ipynb (identified in README)` 134- `training/rl/configs/ (identified in README)` 135 136Read-file identity: `README.md: c931b3e827f9cb7101cc4443d761dfa646cb16a8` 137 138## R06 — Large Language Models Report Subjective Experience Under Self-Referential Processing 139 140Cameron Berg; Diogo de Lucena; Judd Rosenblatt (2025). arXiv research preprint; version inspected: v2. 141 142- [Paper / primary write-up](https://arxiv.org/html/2510.24797v2) 143 144**Inspected:** Read paper methods, results, and availability statement. Public executable repository and complete raw-trial archive were not located in this search. Availability section says code will be made available. 145 146**Relevant support:** Closest conceptual match: self-referential prompting changes rates of first-person experience reports, with prompt controls and activation-steering conditions. 147 148**Limits:** Measures prompted reports, not independently established subjective experience. Claude 4 Opus also had high control-condition report rates. No demonstrated irreversible transition. Treat as relevant theory/experimental literature, not a verified complete reproduction package. 149 150**Reproduction notes:** Exact prompting protocol and aggregate results are in the paper; the paper explicitly distinguishes behaviorally induced self-reference from verified architectural recurrence. 151 152## R07 — Emergently Misaligned Language Models Show Behavioral Self-Awareness That Shifts With Subsequent Realignment 153 154Laurène Vaugrante; Anietta Weckauff; Thilo Hagendorff (2026). arXiv research preprint; version inspected: v1. 155 156- [Paper / primary write-up](https://arxiv.org/html/2602.14777v1) 157- [Data / checkpoints](https://osf.io/reusb/overview?view_only=f8db88ef2afb452abc682e4ba08e226f) 158 159**Inspected:** Read primary methods and findings. Authors link training datasets and benchmark items on OSF; that archive could not be retrieved here. Complete code/raw-output availability not verified. 160 161**Relevant support:** Self-assessments can track learned behavioral changes. 162 163**Limits:** Realignment partially or fully reverses many measured effects. This constrains universal permanence claims but does not prove or disprove experience. 164 165**Reproduction notes:** Keep in evidence synthesis as counterevidence/control design; do not describe as fully downloaded or executable from inspected materials. 166 167## Suggested analysis sequence 168 1691. Use R01 saved checkpoints to locate internal change independently of self-report. 1702. Use R03-style behavioral self-report evaluation on the same checkpoints, with a declaration criterion specified before inspecting changes. 1713. Use R04 causal interventions and controls to distinguish grounded internal reporting from prompted narrative. 1724. Use R02 physical body-position ground truth to test constructed self-estimate mismatch. 1735. Use R05 for a separate reward-driven replication, distinguishing training loops from within-chat inference. 1746. Include R07 and all negative/control conditions when assessing persistence; a finite record cannot establish never returning. 175