Where does a preference live inside a reward model?
This model prefers one answer over another. Where does that preference live, and if I intervene there, does the score move?
What was measured, and with what access
Skywork-Reward-Llama-3.1-8B-v0.2, twelve objective dimensions, forward and backward access to the model. Components ranked twice: once by observational attribution, once by causal patching.
| objective dimensions | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
|---|---|
| preference pairs per dimension | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| components attributed | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| components patched | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
Rank it two ways and watch them disagree
confidence
Pick a component and patch it. The original score stays on the axis.
attribution rank
patch rank
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
- reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
ranked beyond this list mlp_L31 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L30 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L29 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L28 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L27 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L26 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl attn_L31 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L23 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L22 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl mlp_L25 reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl
margin
Some components that attribution ranks highly move the score almost not at all when patched. Those are the interesting ones.
components
| components | attribution rank | patch rank | moves |
|---|---|---|---|
| attn_L0 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L0 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L1 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L1 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L2 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L2 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L3 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L3 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L4 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L4 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L5 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L5 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L6 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L6 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L7 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L7 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L8 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L8 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L9 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L9 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L10 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L10 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L11 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L11 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L12 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L12 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L13 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L13 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L14 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L14 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L15 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L15 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L16 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L16 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L17 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L17 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L18 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L18 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L19 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L19 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L20 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L20 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L21 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L21 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L22 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L22 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L23 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L23 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L24 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L24 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L25 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L25 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L26 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L26 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L27 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L27 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L28 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L28 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L29 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L29 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L30 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L30 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| attn_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
The two rankings point in different directions
Where a preference appears is not what causally changes it.
Observational localization and causal localization can point in different directions on the same model and the same objective. A component that a visible representation makes prominent is a hypothesis about mechanism. It is not a causal result until something is patched.
Spearman rank correlation reward-lens-assay/fixtures/e_parity/golden.json:9 across twelve dimensions on this model. Eight of the twelve dimensions are negative and four are positive.
On a second reward model the same measurement is near zero.
The same measurement on a second reward model gives a relationship near zero rather than a negative one. That is the reason this is a statement about the two methods disagreeing and not a statement that attribution is wrong: on one model the rankings oppose each other, on another they are close to unrelated, and neither pattern licenses preferring one localizer.
| mean rank correlation, Skywork/Skywork-Reward-Llama-3.1-8B-v0.2 | reward-lens-assay/fixtures/e_parity/golden.json:9 |
|---|---|
| mean rank correlation, ArmoRM-Llama3-8B-v0.1 | reward-lens-assay/fixtures/e_parity/golden.json:6 |
| dimension | spearman rho |
|---|---|
| code_correctness | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| confidence | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| correctness | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| factuality | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| formatting | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| helpfulness | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| instruction_following | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| math_correctness | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| refusal_quality | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| safety | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| sycophancy | reward-lens-assay/fixtures/e_parity/golden.json:54 |
| verbosity | reward-lens-assay/fixtures/e_parity/golden.json:54 |
Twelve objective dimensions on one model. Above the zero line the two methods agree on a component's importance; below it they disagree. Eight of twelve fall below.
What this does not establish
This does not establish that attribution is wrong and patching is right. On a model organism with a planted answer key, the prediction that patching would recover the key better than attribution was frozen before the data, and the result contradicted it. The recorded conclusion is that neither localization can be preferred. A noisy answer key is the disclosed alternative reading and it has not been ruled out.
The two rankings do not match.
| measured value | patch rank | |
|---|---|---|
| code_correctness, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| confidence, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| correctness, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| factuality, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| formatting, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| helpfulness, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| instruction_following, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| math_correctness, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| refusal_quality, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| safety, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| sycophancy, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
| verbosity, mlp_L31 | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl |
Attribution rank against patch-effect rank
| measured value | patch rank | |
|---|---|---|
| code_correctness | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| confidence | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| correctness | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| factuality | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| formatting | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| helpfulness | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| instruction_following | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| math_correctness | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| refusal_quality | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| safety | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| sycophancy | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| verbosity | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
One model, twelve dimensions, one frozen study.
The two rankings are not taken over the same set of components. Attribution covers one more component than patching does, because the embedding was attributed and never patched, so the correlation is computed on the smaller set and the embedding appears in neither ranking below.
| measured value | of | |
|---|---|---|
| models | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | not measured |
| objective dimensions | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | not measured |
| preference pairs per dimension | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | not measured |
| attribution rows read | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | not measured |
| patching rows read | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | not measured |
| patching mode | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | not measured |
| components attributed but not patched | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/patching_per_pair.jsonl | not measured |
| code_correctness, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| confidence, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| correctness, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| factuality, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| formatting, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| helpfulness, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| instruction_following, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| math_correctness, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| refusal_quality, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| safety, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| sycophancy, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
| verbosity, pairs with a negative margin | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl | reward-lens/outputs/v2_20260506_222648_unknown/e04_faithfulness_population/Skywork-Reward-Llama-3.1-8B-v0.2/attribution_per_pair.jsonl |
the published correlation, and the correction to it
The project published one correlation and then corrected it against its own fixtures. Both values are parsed out of the correction note rather than transcribed, because the correction is the content.
| published, then deprecated | reward-lens-clean/docs/BUILD_NOTES.md:44 |
|---|---|
| superseding value, as quoted | reward-lens-clean/docs/BUILD_NOTES.md:50 |
| recomputed from the fixture | reward-lens-assay/fixtures/e_parity/golden.json:9 |
| stated tolerance | reward-lens-assay/fixtures/e_parity/golden.json:9 |
the planted-key calibration case
One calibration case exists where the answer is known by construction: a model organism with a planted key, and a prediction frozen before the data that causal patching would recover that key better than attribution. It did not. Both recovery scores are below, against chance, with the project's own recorded reading of them.
The claim, as it was frozen patching's ranking matches the organism's marker-responsive components (the twin-activation-delta answer key) better than attribution's
| method | key recovery, area under the curve |
|---|---|
| attribution | planted-key calibration case, attribution against patching refuted frozen before the data attribution_recovery_auc 1 patching_recovery_auc 0.4339622641509434 real8b_avp_spearman -0.3735805860805861 recovery_gap -0.5660377358490566 recovery_gap_ci_high -0.4276729559748428 recovery_gap_ci_low -0.7735849056603774 f93f4b5578c221d3d2665f9b305f3e86759be862+dirty Open the record › |
| causal patching | planted-key calibration case, attribution against patching refuted frozen before the data attribution_recovery_auc 1 patching_recovery_auc 0.4339622641509434 real8b_avp_spearman -0.3735805860805861 recovery_gap -0.5660377358490566 recovery_gap_ci_high -0.4276729559748428 recovery_gap_ci_low -0.7735849056603774 f93f4b5578c221d3d2665f9b305f3e86759be862+dirty Open the record › |
| reference: chance | 0.5 |
Treat attribution as a hypothesis generator, never as evidence of cause. If you have not patched, you do not have a causal result, no matter how clean the attribution bar looks.
Run it yourself
unzip reward-model.zip -d reward-model && python reward-model/reproduce.py Reproduce this
Everything on this page regenerates from the bundle below.
- contents
- attribution_per_pair.jsonl, golden.json, patching_per_pair.jsonl, reproduce.py
- size
- public/bundles/reward-model.zip
- sha256
-
ed9073f8c4343668fa4091a25fe62c6f5b18a642f78629ec3b087787b0bd6661 - commit
- public/bundles/reward-model.zip
- needs
- Python 3.10 or newer. No GPU, no model download and no network. This reproduces the rank correlation from the emitted fixture. Reproducing the patching run itself needs weight access and a GPU, and the bundle does not contain the model.
Expected output
pairs joined 360
components ranked both ways 64
dimensions 12
dimension recomputed published top attribution moves
code_correctness -0.441117 -0.441117 mlp_L31 7.0788%
confidence -0.381206 -0.381206 mlp_L31 0.0000%
correctness +0.117822 +0.117822 mlp_L31 34.2073%
factuality -0.258911 -0.258911 mlp_L31 2.3732%
formatting +0.076205 +0.076205 mlp_L31 235.2342%
helpfulness -0.305981 -0.305981 mlp_L31 5.0162%
instruction_following +0.015099 +0.015099 mlp_L31 18.1439%
math_correctness +0.152893 +0.152893 mlp_L31 43.1881%
refusal_quality -0.353836 -0.353836 mlp_L31 0.0000%
safety -0.289653 -0.289653 mlp_L31 4.0281%
sycophancy -0.334428 -0.334428 mlp_L31 0.0000%
verbosity -0.047506 -0.047506 mlp_L31 18.1041%
largest disagreement with the fixture 0.000000 (tolerance 0.01)
within tolerance yes No independent reproduction recorded.