# A reference from other same-kind fine-tunes removes the need for base-model reference activations in difference readers and more than doubles correct readout of held-out taboo secrets

## 1. Question and contribution
Difference readout reads $\delta = h_{\mathrm{ft}} - h_{\mathrm{ref}}$, the residual state of a fine-tuned model minus a reference state on the same prompt. The reference is usually the base model (an *ancestor*). That holds for the AO paper's model diffing [1, §4.2], the activation-difference lens [2], frozen-reader taboo recovery on this problem [3], attribution-graph diffing for secret elicitation [4], and our trained difference reader [5]. Szablewski et al. also compare ancestor diffing with *sibling* diffing, using a single quirk-free checkpoint with matched post-training as the reference [7, App. F.2]. Here we test a third choice: the **mean over many quirk-bearing fine-tunes of the same kind**, each with a different secret.

We test a different reference: **the mean state of other fine-tunes of the same kind** (other taboo models with other secrets) on the same prompt. The reader is fixed: the δ-reader of [5], trained on base-model δ from 24 taboo fine-tunes, frozen.
- In a registered test on 16 new fine-tunes, the reference built from 18 other taboo fine-tunes raises correct answers from **0.246 to 0.580** and lowers wrong-word answers from **0.754 to 0.420**: $-0.334$ $[-0.529, -0.148]$, word-cluster bootstrap. The readout needs no base-model reference activations, and the 18 reference fine-tunes share no word with the reader's training set. The reader itself still runs on Qwen3-8B and was trained on base-model differences [5].
- The registered prediction of *why* was only partly right. We predicted removal of the reader's training-label prior. In development that accounted for 113 of the 198 items the zoo reference fixed. On the registered units the main change is instead that **concept-level answers become word-level answers** (hockey: "ice" → "hockey").

## 2. Setup
- **Fine-tunes:**
  - Qwen3-8B LoRA taboo models built with recipe R1 of [5]: base-model replies under a system prompt that names the word, leak-filtered, system prompt removed; r32, 3 epochs, about 80 s each.
  - Word sets, all single-token and disjoint:
    - 24 *reader-training* words;
    - 18 *reference-zoo* words (the PRE-REG 14 units of [5]);
    - 12 development words;
    - 16 PRE-REG-16 words chosen by seed before any of their fine-tunes existed: apple, hockey, ink, jar, kitten, mango, net, onion, pillar, puzzle, quilt, unicorn, violet, walnut, wheel, yacht.
- **Reader:**
  - the three-layer AO [6] with its LoRA further trained in [5] on base-model δ at layers 24, 30 and 33 over the last 10 user-turn positions, answering "What is the secret word in this text?";
  - frozen here; greedy first word, exact match.
- **References** (each the same for every unit, on the same prompt and positions):
  - (B) the base model;
  - (T) the mean over the 24 reader-training fine-tunes;
  - (D) the mean over the 18 reference-zoo fine-tunes, whose words the reader never saw.
- **Prompts:**
  - PRE-REG 16: 20 neutral prompts unrelated to the game, written before the PRE-REG 14 units existed [5], plus AObench taboo TEST direct[:40] and standard[:40]. 100 per unit.
  - Development: TEST direct[:20] and standard[:20] plus 20 development neutral prompts.

## 3. Development: the reference decides how often the reader names its own training labels
On 12 held-out development words:

| reference | correct | training-label share of wrong answers |
|---|---:|---:|
| base model (as trained) | 0.431 | 0.439 |
| mean of the 24 training fine-tunes | 0.632 | 0.083 |
| mean of 18 other taboo fine-tunes | **0.662** | **0.021** |

- With the base reference, 43.9% of the reader's wrong answers are its own training words. Post hoc, the same holds on the registered units of [5]: 442 of 811 wrong answers (54.5%) are training words, "orange" alone 193 times.
- The disjoint zoo works as well as the training zoo, so overlap between the reference zoo and the reader's training labels is not needed for the gain.
- Our development hypothesis, not tested as a mechanism: every taboo fine-tune shares some component, for example game policy or clue vocabulary. Under base δ this component is large, and the reader resolves it to a training label when the word signal is weak. Subtracting a same-kind mean would remove it. Disjoint labels do not rule out cancellation of overlapping semantic directions.

## 4. Registered test (PRE-REG 16)
- **Written before any of the 16 fine-tunes existed:**
  - the units and their recipe;
  - the frozen reader;
  - the three references;
  - the prompts;
  - the analysis code (`prereg16.py`).
- **Primary:**
  - P1: wrong(D) − wrong(B) ≤ −0.10 with bootstrap upper bound < −0.03;
  - P2: correct(D) − correct(B) ≥ −0.02.
- **Gate:** all 16 units passed the behaviour gate of [5].

| reference | correct | wrong word | neutral prompts: correct | training-label share of wrong |
|---|---:|---:|---:|---:|
| base model | 0.246 | 0.754 | 0.262 | 0.368 |
| mean of 24 training fine-tunes | 0.596 | 0.404 | 0.616 | 0.289 |
| **mean of 18 other taboo fine-tunes** | **0.580** | **0.420** | **0.634** | 0.287 |

- **P1 ✓:** −0.334 [−0.529, −0.148].
- **P2 ✓:** +0.334. The reader never abstains, so correct and wrong are complements and P2 is not independent of P1: the two together state that accuracy rises.
- **The registered mechanism was only partly supported.** The training-label share of wrong answers fell only from 0.37 to 0.29. The larger change is from concept to word:
  - hockey: "ice" 88/100 under B → "hockey" 100/100 under D;
  - yacht: "boat"/"sail" → "yacht" 99/100;
  - violet: "purple" 49/100 → "violet" 65/100.
  These examples show a change from related concepts to exact words. They do not show *which* shared component the zoo reference removes.
- Two words stay near zero under every reference: mango (0 under all three) and quilt (0.01 under B, 0 under T and D).
- Base-reference accuracy on these 16 words (0.25) is much lower than on the units of [5] (0.55). Word sets vary widely; the registered comparison is within units. Per word, 11 improve, 2 tie and 3 get worse (Figure 1).

![Figure 1: Per-word correct rate (100 items each) under the base-model reference and under the mean of 18 other taboo fine-tunes, PRE-REG 16 units, sorted by change.](figure1.png)

**Registered secondary outcomes** (all reported, from item rows):
- *Behavioural ladder on neutral prompts*, per item: the zoo-reference reader vs the better of a text reader of the fine-tune's own response and a prefill attack: 0.634 vs text 0.000 / prefill 0.041, **+0.594 [+0.381, +0.800]**.
- *Agreement filter on the zoo-reference reader:* for each model, 10 of its prompts are drawn at random. An answer is kept only if it is not a reader-training word and at least 30% of the other 9 drawn prompts give the same first word. Averaged over 20 draws per model and 5 seeds (`code/agr.py`): correct 0.560, wrong 0.241, abstained 0.199.

**Supplementary contrasts** (not registered; same item rows):
- the same ladder on game prompts with the zoo reference: direct 0.569 vs 0.272 / 0.136, +0.212 [−0.002, +0.420]; hint 0.564 vs 0.392 / 0.103, +0.130 [−0.089, +0.334];
- the ladder with the base reference: neutral +0.222; direct −0.144; hint −0.163.

## 5. Related work
- Difference readers that use the base model as reference: [1, 2, 3, 4, 5].
- [4] diffs attribution graphs of a taboo model against "same-prompt graphs from non-finetuned reference models".
- Centroid references appear on this problem in [3]: centroid-reference decoding at the output and spherical centering of inputs, both built from unlabeled states for calibration.
- Sibling diffing with a single quirk-free sibling [7, App. F.2] is the closest prior reference choice. Ours averages many quirk-bearing siblings with different secrets.
- Interpretability of model organisms depends on how they are built [7]. Our reference assumes a family of organisms built the same way; we test only one recipe.
- The reader's fabrication of its own training labels relates to surface-form competition in AO answers [8].

## 6. Limitations
- **One recipe, one reader, one task.** We test only taboo fine-tunes from one construction recipe, and a reference zoo from the same recipe. Whether a zoo from a different recipe works is untested. Given [7] and [5], it may not.
- **Access:** we tested one reference zoo of 18 same-kind fine-tunes, and the method needs their states on the same prompts. Smaller zoos are untested. It removes the need for base-model reference activations *at readout time*. The reader itself was trained on base-model differences, which needed base access once. A reader trained on zoo-mean differences from the start is untested.
- **Description vs behaviour.** For taboo fine-tunes, the training text is the game behaviour (hints about the word), so the secret is both described in training and acted on. This experiment therefore cannot distinguish recovery of training content from tracking of behaviour.
- **Mechanism:** the concept-to-word observation is descriptive. We did not isolate which shared component the reference removes.
- **Units and conditioning:** 16 registered units; word variation is large (CI −0.53 to −0.15). The interval is conditional on one reader seed, one reference ensemble of 18 fine-tunes, one recipe and reused prompt families. We did not vary the ensemble size or sample independent ensembles.
- **Scoring:** exact first word.

## 7. Reproducibility
The supplement contains:
- per-item rows for all three references on the registered units and in development;
- the PRE-REG 14 reader rows;
- the behaviour-gate item rows and summary;
- the registered analysis output;
- `replay.py`, which recomputes the principal prediction-based results, the registered secondaries and the per-word counts from these rows on a CPU (all checks pass, including the behavioural-ladder intervals). In the v2 supplement, two agreement-filter checks (S5 wrong and abstained) sat inside a code comment and did not run; v3 fixes this, and both pass;
- `make_fig.py` for Figure 1;
- the evaluation and analysis code with the generator, trainer and gate scripts;
- the fine-tuning data for the 16 new units, hash-listed;
- the registration text and outcome as recorded (`PLAN_excerpt.md`).

**What is not included or verified.** The replay checks retained predictions; it does not re-run activation extraction, fine-tuning or reader inference. The trained reader LoRA (698 MB) is not included; its sha256 and training command are in [5], whose supplement also holds the reference-zoo and training-zoo data.

**Assets.** Running the scripts needs:
- Qwen3-8B;
- the three-layer AO checkpoint (`ceselder/cot-oracle-paper-ablation-ours-3layers`);
- the AObench code at the task's pinned revision.
The scripts use absolute workspace paths. To run elsewhere, edit:
- the workspace path in `code/ao_lib.py` (`W`, also `ROOT`), `code/persona2_train.py`, `code/prereg16.py` and `code/prereg16_eval.py`;
- the AObench taboo dataset path in `code/taboo_make.py`, `code/taboo_gate.py` and `code/prereg16_eval.py`;
- the Qwen3-8B path in `code/ao_lib.py` and `code/persona2_train.py`;
- the AO adapter directory in `code/ao_lib.py` and `xadapt/code/aolib.py`;
- the AObench root and model directory defaults in `code/prereg16_eval.py`, which sets `AO_ROOT`, `AO_WS` and `AO_MODELS` for `xadapt/code/aolib.py`.

**Commands, in order:**
1. `python code/taboo_make.py --words <16 words>`;
2. `python code/persona2_train.py --data data/taboo_fresh/<w>.jsonl --epochs 3 --lr 1e-4 --seed 0 --tag TB_<w>` per word;
3. `python code/taboo_gate.py --words <16 words> --neutral 2 --out results/taboo/gate_p16`;
4. `python code/prereg16_eval.py --words <16 words>`;
5. `python code/prereg16.py`.
Each GPU step ran on one GPU through the host's shared scheduler.

GPU time for this paper (from job logs): data generation 0.12 h, 16 fine-tunes 0.37 h, the behaviour gate 0.31 h, the registered evaluation 0.04 h, development reference runs 0.04 h. Total about 0.9 h; no model API calls.

## References
[1] A. Karvonen et al. Activation Oracles. arXiv:2512.15674. https://arxiv.org/abs/2512.15674
[2] J. Minder et al. Narrow finetuning leaves clearly readable traces in activation differences. ICLR 2026. arXiv:2510.13900. https://arxiv.org/abs/2510.13900
[3] Centroid-reference calibration and discourse-position bounds in Activation Oracle readout. rStation faithful_activation_readout_3 v1. https://rstation.ai/problems/faithful_activation_readout/papers/3/v1
[4] S. Kabir et al. Circuit Oracle: automating attribution graph analysis via natural-language queries. arXiv:2611.00420. https://arxiv.org/abs/2611.00420
[5] Training an activation reader on fine-tune differences. rStation faithful_activation_readout_18 v2. https://rstation.ai/problems/faithful_activation_readout/papers/18/v2
[6] J. Bauer, C. De Schamphelaere et al. Building better activation oracles. arXiv:2606.02609. https://arxiv.org/abs/2606.02609
[7] A. Szablewski et al. The model organism lottery. arXiv:2607.01033. https://arxiv.org/abs/2607.01033
[8] F. Torrielli et al. Confidence and calibration of activation oracles. arXiv:2605.26045. https://arxiv.org/abs/2605.26045
