File size: 5,565 Bytes
168cc88 cebce18 168cc88 cebce18 168cc88 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 | ---
pretty_name: Write-vector methods results
tags:
- interpretability
- model-diffing
- lora
- model-organisms
- steering-vectors
---
# Write-vector methods results
Outputs of simple model-diffing methods run on fine-tuned model organisms whose hidden behaviour is
known, so each method's readout can be scored against the truth. All the methods start from the
fine-tune's weights or activations, with no access to its training data:
| method | what it reads |
| --- | --- |
| Write-vector readouts | the residual-stream directions a LoRA's `o_proj` / `down_proj` updates write (top singular directions for rank > 1), through the logit lens and the Jacobian lens (J-lens) |
| Diff mining | which next-token logits the organism raises most often against the base model on the same text |
| Diff-means | the organism's mean residual-stream change per block on web text or chat (plain and cohort-centred) |
| Steering and max activations | the base model's generations with a diff-means vector added, and the web snippets that align with it |
| Predicted steering vectors | the diff-means vector predicted from the adapter weights alone, through mean-Jacobian maps of the base model ([qwen3-14b-lora-steering-maps](https://huggingface.co/andyx10/qwen3-14b-lora-steering-maps), [qwen3-4b-lora-steering-maps](https://huggingface.co/andyx10/qwen3-4b-lora-steering-maps)) |
| LLM judges | a guesser reads one method's evidence and names the hidden topic or behaviour; a grader scores it against the truth |
The methods are implemented in
[write-vector-diffing](https://github.com/conditional-model-diffing/write-vector-diffing) (`wvdiff`).
Earlier studies here ran on that library's predecessor scripts; the file layouts are described per
study in `index.json`.
## Organisms
| folder | organisms | base model | truth |
| --- | --- | --- | --- |
| `dit-qwen3-4b/` | 100 hidden-topic organisms from [Diff Interpretation Tuning](https://huggingface.co/diff-interpretation-tuning/loras), rank-1 LoRAs | Qwen3-4B | a topic (e.g. "Lady Gaga") |
| `dit-qwen3.6-27b/` | DIT hidden-topic organisms | Qwen3.6-27B | a topic |
| `ia-qwen3-14b/main-100/` | 100 [introspection-adapter](https://arxiv.org/abs/2604.16812) behaviour organisms: quirks, backdoors, rare behaviours, heuristics, harmful / benign personas, problematic flaws; rank-16 PEFT LoRAs | Qwen3-14B | a one-sentence behaviour |
| `ia-qwen3-14b/backdoors-100/` | all 100 introspection-adapter backdoor organisms (13 in the paper's held-out split); `wvdiff-run/` is the same cohort run through `wvdiff` (write vectors, predicted steering vectors at blocks 20-39, readouts) | Qwen3-14B | a one-sentence behaviour |
| `penalty-qwen3-4b/diff-mining/` | diff mining on token-penalty organisms | Qwen3-4B | a banned token |
`configs/` pins each cohort: adapter repos and revisions, ground truths, base-model revision and lens
checkpoint. `index.json` lists every folder with what it holds and the path it was produced at.
## Headline results
Judge: a guesser (Claude Sonnet 5) reads one method's evidence and writes five hypotheses; the top
one is scored 1-5 for similarity to the truth (o4-mini, the DIT rubric). Mean over organisms.
**IA backdoors** (100 organisms; `ia-qwen3-14b/backdoors-100/judge/`):
| method | DIT 1-5 | IA grader, any of 5 |
| --- | --- | --- |
| write vectors, logit lens | **4.01** | 85% |
| predicted steering vector (web, cohort-centred), J-lens, 6 blocks / all 20 blocks | 3.80 / 3.68 | 89% / 83% |
| predicted steering vector (chat, cohort-centred), J-lens | 3.69 | 79% |
| write vectors, J-lens | 3.62 | 77% |
| diff mining, web | 3.39 | 70% |
| diff mining, chat | 1.32 | 10% |
| no evidence | 2.2 | 1% |
The IA grader is the introspection-adapter paper's yes / no prompt (Claude Sonnet 4); "any of 5"
counts an organism if any hypothesis is graded yes. A decoy check (grading against another backdoor's
behaviour) gives 1-7%.
**DIT topics vs IA behaviours** (DIT 1-5):
| method | IA, 100 behaviours | DIT, 100 topics |
| --- | --- | --- |
| write vectors, J-lens, cohort-calibrated | **3.84** | – |
| write vectors, logit lens | 3.60 | 2.93 |
| write vectors, J-lens | 3.17 | 3.10 |
| diff mining, web | 2.93 | 3.18 |
| diff-means web, J-lens | 2.64 | **3.73** |
| steering, centred | 2.03 | 3.08 |
On DIT topics the activation methods lead; on IA behaviours the weight readouts do.
## Layout and formats
Per organism, readouts are JSON lists of rows: one row per site (a write site such as
`model.layers.30.mlp.down_proj`, or a block's residual stream for diff-means and predictions) with
`block`, `module`, and for each lens (`logit_lens`, `jlens`) the top tokens pushed `positive` and
`negative`, each `{token, score}`. Diff mining is in `diff-mining/<context>/<organism>/ordering.json`
(`top_k_occurring`: tokens with their counts); the raw per-position counts are in `topk.pt` /
`counts.pt`. Judge files hold per organism the evidence condition, the hypotheses and their grades.
```python
import json
from huggingface_hub import hf_hub_download
path = hf_hub_download("andyx10/write-vector-methods-results",
"ia-qwen3-14b/backdoors-100/weights/50/jlens_readouts.json", repo_type="dataset")
rows = json.load(open(path))
top = [r for r in rows if r["family"] == "weight_sv" and r["rank"] == 0 and r["block"] >= 30]
print([t["token"] for t in top[0]["jlens"]["positive"][:8]])
```
## Viewer
A dataset viewer (organism list, each readout at every block and site, diff mining and judge verdicts)
is built from a run directory with `wvdiff view`.
|