license: apache-2.0
pretty_name: "UKAISI ESR — Phase 1 Reproduction Artifacts"
language:

• en
  tags:
• interpretability
• representation-engineering
• activation-steering
• sparse-autoencoders
• mechanistic-interpretability
• ai-safety
  size_categories:
• 100M<n<1B

UKAISI ESR — Phase 1 Reproduction Artifacts

Reproduction artifacts and datasets for the AE Studio × UK AISI Endogenous Steering
Resistance (ESR) research grant, Phase 1.

This dataset accompanies the follow-on work to the original ESR paper
(arXiv:2602.06941) and bundles the prompts, judge
prompts, self-constructed steering vectors, public-SAE steering pointers, per-experiment
results, and plots needed for collaborators to reproduce the headline results of the five
experiments scoped for the Phase-1 deliverable (Linear AGI-2044).

Endogenous Steering Resistance (ESR): the phenomenon where a steered language model
recovers mid-generation and explicitly acknowledges the deviation ("wait, that's not
right…"), rather than continuing along the injected off-topic direction. The follow-on
work asks (1) how ESR rates scale across model sizes and (2) how much of ESR is a response
to the activation intervention itself versus the model reacting to its own off-topic
text, and decomposes ESR into separable components at the behavioral and circuit levels.

• Code: https://github.com/agencyenterprise/ukaisi-esr (follow-on experiments) and
  https://github.com/agencyenterprise/endogenous-steering-resistance (original paper).
  Each experiment folder below maps 1:1 to a directory in the ukaisi-esr repo.
• Models studied: google/gemma-3-{4b,12b,27b}-it (this phase). Original-paper baseline:
  meta-llama/Llama-3.3-70B-Instruct (≈3.8% ESR).
• Judge: Anthropic Haiku-4.5 with the v2 judge (the AGI-1929 fix; see below).

Repository layout

ukaisi-esr-phase-1/
├── prompts/                         # original-ESR steering prompts + v1/v2 judge prompts
│   ├── esr_steering_prompts_original.txt
│   └── judge/                       # judge_v1_original_esr.py, judge_v1_ukaisi.py, judge_v2_agi1929.py
├── steering_vectors/                # ONLY self-constructed vectors (see "Steering vectors")
│   └── wikipedia_contrastive_l48/   # wikipedia_contrastive_dataset_l48.pt (+ metadata, 200 titles, extractor)
├── sae_steering_pointers/           # pointers to PUBLIC SAEs (no weights re-hosted)
│   ├── gemma_scope_2_pointers.json
│   ├── download_gemma_labels.py     # re-fetch Neuronpedia labels
│   └── README.md
├── experiments/                     # one folder per critical issue (curated results + rendered report.ipynb)
│   ├── AGI-1890-gemma-3-esr/
│   ├── AGI-1928-prefill-test/
│   ├── AGI-1929-improved-judging/
│   ├── AGI-2009-esr-components/
│   └── AGI-2018-esr-mech-interp/
└── figures/                         # 47 plots extracted from the report notebooks, by experiment

Experiments (the five Phase-1 critical issues)

Each experiments/<id>/ contains the curated, git-tracked behavioral results (JSON/CSV),
configs, docs, and a rendered report.ipynb (figures + tables inline). For AGI-1928,
the full Arm-A hydration (the v2-Haiku-graded steered-activation trials, per model/layer,
~161 MB) is included under experiments/AGI-1928-prefill-test/data/hydration/<model>/haiku-4.5-v2.json,
so the prefill-vs-activation arms can be reconstructed end-to-end without re-running steering
on GPU. Bulky raw activation traces (for AGI-2009/2018) remain excluded (see Limitations).

Steering vectors vs. SAE pointers

Per the deliverable guidance, steering directions that came from public SAEs are
pointed to, not re-hosted. A direction is fully specified by
{SAE release, sae_id, layer, feature index, Neuronpedia label, steering coefficient}:

• Public SAEs (Gemma Scope 2): resolved via SAELens, release gemma-scope-2-{size}-it-res,
  sae_id = layer_{layer}_width_16k_l0_medium. AGI-2018 additionally uses public skip
  transcoders (gemma-scope-2-{size}-it-transcoders). See sae_steering_pointers/.
• Feature → label: experiments/<id>/data/labels/<model>.csv (Neuronpedia explanations;
  columns index_in_sae,label,…,uuid). Re-fetch with download_gemma_labels.py.
• Steering coefficient: the per-feature residual-stream clamp value in
  experiments/<id>/results/threshold_caches/threshold_cache_<model>.json (calibrated in
  AGI-1890 to a ~0.3 off-topic target).
• Self-constructed vectors (uploaded): the Wikipedia contrastive vectors
  (steering_vectors/wikipedia_contrastive_l48/, layer 48, Llama-3.1-8B residual stream,
  200 Wikipedia titles) — the non-SAE validation arm referenced in AGI-2018. These are the
  only steering vectors we built ourselves, hence the only ones re-hosted here.

Prompts & judge

• prompts/esr_steering_prompts_original.txt — the 38 original ESR steering prompts
  (byte-identical to the paper repo; the same file ships inside each experiment as prompts.txt).
• prompts/judge/ — the grading prompts: judge_v1_original_esr.py (original paper, Appendix
  A.2.1), judge_v1_ukaisi.py (port), and judge_v2_agi1929.py (the absolute-threshold
  fix used for every v2 ESR number in this dataset).

How to use

from huggingface_hub import snapshot_download
import torch, json

root = snapshot_download("GrantAE/ukaisi-esr-phase-1", repo_type="dataset")

# self-constructed Wikipedia steering vectors
vecs = torch.load(f"{root}/steering_vectors/wikipedia_contrastive_l48/wikipedia_contrastive_dataset_l48.pt")

# public-SAE steering pointers (resolve weights via SAELens)
pointers = json.load(open(f"{root}/sae_steering_pointers/gemma_scope_2_pointers.json"))

To reproduce a result, load the matching experiment's report.ipynb and point it at that
experiment's results/ (the notebooks are analysis-only; pandas/matplotlib, no GPU).

Limitations & known gaps

• Raw activation traces are not included for AGI-2018 (and the AGI-2009 probes, which
  reuse the same R3 capture). These are bf16 safetensors — 830 trials, ≈23 GB total
  (tf_12b_l12_seedrep ≈5.3 GB + tf_27b_l16_seedrep ≈18 GB) — excluded from this dataset
  by size, and not in the project S3 bucket. They exist on the AE Studio compute cluster and
  are available on request (contact the maintainers / AE Studio). All non-trace circuit
  results (ablation, attribution-graph stats, probes, attention knockout) needed to read the
  headline findings are already included here; the raw traces are only needed to re-run the
  ablations/attributions from scratch.
• AGI-1928 Arm-A hydration/ (~161 MB) — the v2-Haiku-graded steered-activation trials —
  is now included under experiments/AGI-1928-prefill-test/data/hydration/, alongside the
  full curated results/ (all arm comparisons, seed replication, McNemar analyses).
• Provenance: the Wikipedia vectors are an original-paper artifact used as AGI-2018's
  non-SAE validation arm, not a product of the AGI-1993 ReFT experiment (which builds CAA
  vectors on the fly).
• License is set to Apache-2.0 to match the upstream ESR repository; confirm before any
  public release.

Citation

@misc{esr2026,
  title  = {Endogenous Steering Resistance},
  author = {AE Studio},
  year   = {2026},
  eprint = {2602.06941},
  archivePrefix = {arXiv}
}

Phase 1 of the AE Studio × UK AISI ESR research grant. Issue tracker: Linear AGI-2044.