File size: 1,708 Bytes
848636f
 
2314b29
 
 
 
 
 
 
 
848636f
2314b29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-4.0
language:
- en
tags:
- interpretability
- activation-patching
- feature-steering
- mechanistic
pretty_name: ConsisitencyBench-Interpretability Extension
---

# ConsistencyBench-Interpretability Extension

White-box mechanistic analysis of logical inconsistency using Qwen/Qwen2.5-1.5B-Instruct (local,
full activation access) as a dedicated interpretability testbed, distinct from the
17-model black-box leaderboard.

## Contents

- `layer_probe_results.csv` - per-layer logistic-regression probe accuracy for decoding
  "will this response be inconsistent?" directly from residual-stream activations
- `activation_patching.csv` - literal patching results: copying a consistent run's
  activation into an inconsistent run and checking whether the output flips
- `steering_dose_response.csv` - IR at increasing diff-in-means steering strength (a
  causal dose-response curve, not just a correlational probe)
- `hint_sensitivity.csv` - flip rate and hint-induced inconsistency under misleading
  epistemic pressure (across the 6-model API subset used for this analysis)
- `metadata.json` - best probe layer, probe accuracy, model config

## Method Summary

1. Train a linear probe at every layer to decode inconsistency from the residual stream
   at the final prompt token (Alain & Bengio, 2017 methodology)
2. Take the best layer's probe direction (and the diff-in-means direction) as the
   "shortcut direction", in place of a pretrained SAE, which does not exist for this
   checkpoint
3. Run two causal tests: literal activation patching between matched consistent/
   inconsistent pairs, and a steering dose-response sweep with a specificity check on
   unrelated control questions