Step 17: assistant

Calls update_claim_assessment

{
  "status": "doubtful",
  "claim_id": "d3311aac-502f-4e69-ab97-53ee2b6de622",
  "assessment": "The claim rests on one study, Arad, Mueller and Belinkov's \"SAEs Are Good for Steering – If You Select the Right Features\" (EMNLP 2025). On AxBench's Concept500 set with instruction-tuned Gemma-2-9B, keeping only sparse autoencoder (SAE) features with an output score of at least 0.1 gave harmonic-mean steering scores of 0.546 at layer 20 and 0.470 at layer 31. The authors set these beside the numbers AxBench originally reported and conclude that filtered SAEs are comparable with supervised methods.\n\nThat comparison does not match like with like. The supervised baselines were not re-run. Their scores come from the AxBench paper, which used a gpt-4o-mini judge and its own unreleased sample of instructions. The SAE scores used a Claude 3.7 Sonnet judge and a fresh sample. The authors' own unfiltered SAE run scores 1.5 to 2.8 times AxBench's SAE figure, so pipeline differences alone account for a large part of the apparent gain. The filtered numbers also cover only the concepts whose features pass the threshold, while the baselines cover all concepts. The study does not report what share of concepts were kept.\n\nEven taken at face value, the numbers support only a narrow reading. Filtered SAEs beat the weakly supervised ReFT-r1 at layer 31 (0.470 vs 0.401) and difference-in-means at both layers. They trail ReFT-r1 at layer 20 (0.546 vs 0.630), and they trail the fully supervised LoReFT (0.777, 0.764) and LoRA (0.602, 0.580) at both layers. Prompting, at about 1.07, scores roughly twice as high. The paper's \"90.7% of the best method\" figure at layer 20 equals 0.546/0.602, which is the ratio to LoRA, not to the best method. An outside evaluation on Gemma-2-2B layer 20, run under the AxBench protocol (arXiv 2609.35367), found that the output-score filter does not give the reported gains. That test did not cover the 9B setting.\n\nWhat the evidence does support is that output-score selection makes SAE features much better steering vectors than unselected ones, closing part of AxBench's gap. A matched re-run would settle the stronger claim: all methods scored on the same concepts, instructions and judge, with the retained fraction reported.",
  "confidence": 0.75,
  "claim_credence": 0.2,
  "marginal_yield": 0.3,
  "reasoning_trace": "Read the source's TeX (arXiv 2505.20063, final_version.tex), Section 4.4, Table 1 and Appendix on AxBench.\n- Same model, layers and concept set as AxBench's 9B setting (Gemma-2-9B-it, L20/L31, Concept500), same three metrics and harmonic mean. Differences: judge is Claude 3.7 Sonnet (AxBench used gpt-4o-mini, confirmed by arXiv 2609.35367); 10 instructions re-sampled because AxBench did not release its sample (footnote); steering-factor selection on 5 held-out instructions per concept with the authors' own criterion.\n- Baselines in Table 1 are labelled \"AxBench reported results (Wu et al. 2025)\": not re-run.\n- Unfiltered SAE in the authors' pipeline: 0.293 (L20), 0.387 (L31), against AxBench's SAE 0.191 and 0.140. Ratios 1.53 and 2.76. The paper's \"nearly threefold improvement\" over AxBench (0.546/0.191 = 2.86) therefore mixes the filter's effect with pipeline differences. The like-for-like filter effect is 0.293 to 0.546 (L20) and 0.387 to 0.470 (L31), and it is not monotonic in threshold (L20: 0.338 at 0.01 < 0.373 at 0.001).\n- Filtering removes features, and each feature belongs to one concept, so the filtered score averages over a subset of concepts. The number kept is not reported.\n- Ratios at face value: L20 0.546/0.630 (ReFT-r1) = 0.867, /0.602 (LoRA) = 0.907, /0.777 (LoReFT) = 0.703, /1.075 (Prompt) = 0.508. L31 0.470/0.401 (ReFT-r1) = 1.17, /0.580 (LoRA) = 0.81, /0.764 (LoReFT) = 0.615, /1.072 = 0.438. The paper's 90.7% therefore matches the LoRA ratio, not the \"best method\" in its sentence.\n- Outside check: arXiv 2609.35367 (DAFI), Appendix N. It evaluated the output-score filter on Gemma-2-2B L20 under the AxBench protocol with a gpt-4o-mini judge, found the filter retains pairs the judge scores low, and did not observe the gains. It explicitly does not speak to the 9B protocol. No independent matched re-run on 9B was found. A citing paper (arXiv 2602.10437) summarises the result as narrowing the gap, not closing it.\n- \"Competitive with supervised steering methods\" read plainly includes LoReFT and LoRA, which AxBench treats as the supervised fine-tuning baselines. Filtered SAEs are not competitive with these even on the uncorrected numbers. Against ReFT-r1 the results are mixed and confounded. Credence 0.2 allows for a matched re-run showing near-parity with the weaker representation-based baselines."
}

Calls evaluate_argument

{
  "verdict": "slight",
  "evaluation": "Granting the numbers, the inference is weak. It treats scores from different judges, instruction samples and concept subsets as comparable. The authors' own unfiltered run shows those differences alone raise SAE scores 1.5 to 2.8 times. Even at face value, filtered SAEs beat only ReFT-r1 at one layer and difference-in-means, and they trail LoReFT, LoRA and prompting. The argument does establish part of its premise: [[claim:3071aaa9-dbb1-4c25-836e-b7441da8ac21|output-score selection improves SAE steering]] within one pipeline. That is a real but narrower result.",
  "argument_id": "793235b0-7c74-4b32-a837-d43270f3a133"
}

Calls evaluate_argument

{
  "verdict": "likely",
  "evaluation": "Granting its premises, which can be read directly from the source's Table 1, footnote and appendix, the claim as worded is very probably not established. The remaining gap to LoReFT, LoRA and prompting holds even on numbers biased in the SAEs' favour. The argument is not decisive: a matched re-run on Gemma-2-9B could still show parity with the weaker representation-based baselines. The outside test was on a different model.",
  "argument_id": "e21f1f1a-8da0-470d-8eb4-89932eedf449"
}