Pith Number

pith:53I72JDJ

pith:2026:53I72JDJ3YOHUVBXSVUNM2JII5

not attested not anchored not stored refs resolved

Probing Persona-Dependent Preferences in Language Models

Daniel Paleka, Oscar Gilg, Patrick Butlin, Pierre Beckmann

A single linear direction in residual-stream activations predicts and steers task preferences across LLM personas, including opposing ones.

arxiv:2605.13339 v1 · 2026-05-13 · cs.CL · cs.AI

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{53I72JDJ3YOHUVBXSVUNM2JII5}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest claim

This preference representation is largely shared across personas: a probe trained on the helpful assistant predicts and steers the choices of qualitatively different personas, including an evil persona whose preferences anti-correlate with those of the Assistant.

C2weakest assumption

That the linear probe direction identified in residual-stream activations represents the genuine causal preference mechanism rather than a correlated but non-causal feature, and that steering along it produces clean changes to choices without major unintended effects on other capabilities.

C3one line summary

Linear probes on residual-stream activations extract a preference vector that tracks and steers pairwise task choices across personas in Gemma-3-27B and Qwen-3.5-122B, including anti-correlated evil personas.

References

12 extracted · 12 resolved · 0 Pith anchors

[1] Language models as agent models 2022

[2] Emergent Misalignment: Narrow finetuning can pro- duce broadly misaligned LLMs.Nature, 649(8097):584–589, January 2026 2025 · doi:10.1038/s41586-025-09937-5

[3] doi: 10.1111/phpr.12395. David J. Chalmers. What we talk to when we talk to language models. PhilArchive, https://philpapers. org/archive/CHAWWT-8.pdf, · doi:10.1111/phpr.12395

[4] Sam Marks, Jack Lindsey, and Christopher Olah 2026

[5] 2023 , journal = · doi:10.1038/s41586-023-06647-8

Formal links

2 machine-checked theorem links

Receipt and verification

First computed	2026-05-18T02:44:48.414574Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

eed1fd2469de1c7a54379568d6692847513a3ecb45ae3a9a8ad5b460c9684bd0

Aliases

arxiv: 2605.13339 · arxiv_version: 2605.13339v1 · doi: 10.48550/arxiv.2605.13339 · pith_short_12: 53I72JDJ3YOH · pith_short_16: 53I72JDJ3YOHUVBX · pith_short_8: 53I72JDJ

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/53I72JDJ3YOHUVBXSVUNM2JII5 \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: eed1fd2469de1c7a54379568d6692847513a3ecb45ae3a9a8ad5b460c9684bd0

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "6fe760520f2841531bf19dc72c1b1f484d7fd6d120aa6862a3e878d5d9d5b20a",
    "cross_cats_sorted": [
      "cs.AI"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.CL",
    "submitted_at": "2026-05-13T10:57:37Z",
    "title_canon_sha256": "3f390671a8ed420207080c00efc3d0de30108a11c5efba3fc9cfe67400e7869e"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2605.13339",
    "kind": "arxiv",
    "version": 1
  }
}