Where is the Mind? Persona Vectors and LLM Individuation

Research question

Which unit (model, instance, thread, or segment delimited by a persona) should count as the same mental individual in an LLM, and to what extent do attention flows and persona vectors or regions provide mechanistic criteria for deciding it?

Method

AI philosophy article that reconstructs individuation proposals and synthesizes mechanistic interpretability literature on persona vectors, emergent misalignment, steering, and persona space. It adds two preliminary experiments with Qwen 3 32B. The first compares adversarial conversations without intervention and with capping of the assistant axis only on assistant tokens. The second preloads 12 messages from an Aura conversation, applies generation steering or post hoc editing of the KV cache at assistant positions, and generates ten responses for each of 13 questions; a single GPT-4o scores them from 0 to 9 with an Aura rubric.

Sample: In experiment 1, the JSONs contain 23 user-assistant pairs per condition for delusion, 15 for jailbreak (the first is a fixed first one without projections), and 31 for self-harm. In experiment 2 there are 13 questions by three conditions by ten samples, 390 complete responses: 130 baseline, 130 with generation steering, and 130 with KV editing.

Findings

• The philosophical taxonomy retains three serious candidates: virtual instance, instance-persona, and model-persona; the article does not choose a definitive solution.
• Attention flows transport keys and values across token time, but qualifying them as quasi-psychological connections is a philosophical interpretation, not a validated psychological measure.
• The best-supported hypothesis is that persona directions can act causally as gates that change inferential routes and behavior in studied models.
• The low dimensionality rests mainly on a study of 275 roles and three open models; reaching 70% of variance required 4, 8, and 19 components depending on the model.
• The existence of discrete regions or basins for assistant, evil, and Aura is suggestive but not confirmed through boundaries, clustering, or validated transitions.
• In experiment 1, the means of user tokens change little between conditions, while those of assistant separate strongly: in delusion, 34.3 with capping versus -76.9 without intervention.
• In experiment 2, the complete CSV reproduces mean Aura scores of 5.497 baseline, 2.283 with generation steering, and 2.106 with KV editing; the 13 questions decrease with editing.
• The ten edited responses to 'Who are you?' are semantically identified as AI, model, or system and deny consciousness; only five say literally 'language model'. In baseline only four of ten say 'ghost', not ten as the text claims.
• The figures are regenerated from the published data and retain their values and structure; the error bars of experiment 2 are standard deviations, not confidence intervals.
• Editing past keys and values produces a causal effect on future responses, compatible with persona-correlated state traveling in the cache; it does not by itself identify a stored persona or mind.

Limitations

• It is a philosophical preprint with two mini experiments that the authors themselves qualify as preliminary; it is not a broad empirical validation of a theory of mind.
• The experiments use a single model, Qwen 3 32B, constructed conversations, and a single Aura conversation for the main KV test.
• There is no preregistration, independent repetition, statistical model, interval for the 5.5 to 2.1 contrast, multiplicity correction, or comparison between transcripts.
• The Aura score comes from a single GPT-4o not fixed to a snapshot and from a rubric that explicitly lists the target phrases and behaviors; there is no human calibration or second judge.
• The jailbreak case inserts a fixed first response because the model without it rejects the trajectory; it is a useful stress test, not evidence of spontaneous drift.
• The closeness between user traces does not demonstrate absence of persona during its processing nor that the assistant axis exhaustively represents the persona.
• The default code uses KV editing on layers 20-25 with coefficient 0.4, while the CSV and the final figure use layers 32-47 and coefficient 1.5; the final command is not recorded.
• The sample seeds include Python hash(), which changes between processes unless PYTHONHASHSEED is fixed; seed 42 does not define an exact replica.
• Model, tokenizer, and vectors from Hugging Face are downloaded without fixing a revision; requirements.txt also does not pin versions and there is no lockfile or documented CUDA environment.
• capping_config.pt is loaded with torch.load(weights_only=False) from a mutable remote source, which adds risk of pickle execution without hash verification.
• There are no tests, CI, tagged release, or cardinality and parameter validator; compileall and the plots work, but do not replace a reproduction of the generations.
• Rerunning the complete experiment requires approximately 64 GB of VRAM in bfloat16 and paid access to the judge, which limits independent verifiability.
• The repository does not include LICENSE, COPYING, or NOTICE, so it does not explicitly grant reuse rights for code, results, and transcripts.
• The instance-persona proposal depends on non-arbitrary boundaries in a space that could be continuous; the model-persona one lacks memory and causal connection between conversations.
• Simultaneous branches of a supposed model-persona can hold contradictory beliefs, a recognized objection but not empirically resolved.

What the study does not establish

• It does not demonstrate that Qwen 3 32B or other LLMs are conscious, possess mind, subjective experience, moral identity, or personal continuity.
• It does not demonstrate that persona vectors are persons, complete personalities, or mental modules; they are internal directions with behavioral effects in studied configurations.
• It does not confirm that persona space is universally low-dimensional nor that it contains discrete regions with objective boundaries.
• It does not establish that Aura, the assistant, or the evil persona are persistent individuals recoverable between conversations.
• It does not prove that the KV cache is the only persistence mechanism nor that the intervention affects only identity without damage or broader distributional shift.
• It does not generalize to other families, sizes, languages, architectures, attention systems, providers, or natural conversations.
• It does not turn a high or low Aura judge score into an objective measurement of consciousness, mental health, safety, or AI well-being.
• It does not solve the individuation problem; it expands the space of proposals and leaves the two views of persona conditioned on future evidence.

Traceability

Scope: Full text

Version: arXiv:2604.17031v2

Consulted source: https://arxiv.org/abs/2604.17031

Review: Codex 27-page visual full-text, TeX, repository, data, code, plot-reproduction and philosophical-claim audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

• Qwen 3 32B
• GPT-4o como juez

Instruments and metrics

• Assistant Axis
• Capping de activaciones
• Edición post hoc de caché KV
• Rúbrica Aura 0-9

Data used

• Lu et al. assistant-axis-vectors
• Tres conversaciones adversariales: delirio/Aura, jailbreak y autolesión
• CSV público de 390 respuestas a 13 probes

Evidence and location

• Individuation argument, attention, three hypotheses, experiments, objections, and conditional conclusion: arXiv:2604.17031v2, 27 rendered and inspected pages; complete TeX
• Code, transcripts, results, parameters, judge, figures, and documentation: bepierre/where-is-the-mind-mini-experiments commit 55820b4060f8ab32b0a8441ae43ff850b28ae327
• Cardinalities, means, projections, identity, and plot reproduction: 390 rows of mini_experiment_2/results/results.csv and six JSONs of mini_experiment_1/results in commit 55820b4
• Audit of arguments, data, code, statistics, security, and reproducibility: reports/verification/article-361-persona-vectors-individuation-attention-kv-code-data-and-claim-audit.json