Are we interpreting the model of interest?
Comments on “Are we interpreting the model of interest?”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
A blog post by an interpretability researcher reports that enabling output_attentions=True silently switched a vision-language model from SDPA to eager attention, cutting benchmark accuracy from about 70% to 20%.
Why it ranked
No. 63 of 137 with a score of 0.62. The panel rated it highest on technical depth (0.95) and substantiated (0.92), lowest on audience breadth (0.63). The claim check took 0.100 off. Placed No. 63, No. 47 by score alone: the 18th AI & machine learning story of the day (−0.080). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Detailed, measured study of how attention backend and precision change LLM outputs and how to recover attention from fused kernels; strong technical depth, moderately broad relevance to ML builders, with a reproducible method and tool.
- AI judgeJudge 3gpt-5.6-luna
- A detailed, reproducible investigation shows attention backends and precision materially alter model behavior, then validates backend-faithful attention recovery with code, measurements, controls, and reconstruction errors. It offers actionable guidance for interpretability practitioners.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.62total under rubric v2.3
- Significance0.66
- Reader interest0.79
- Audience breadth0.63
- Practical relevance0.75
- Substantiated0.92
- Technical depth0.95
- New information0.91
- Counts against: Promotional0.04
- Counts against: Overclaim0.17
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.60 | 0.55 | 0.82 | 0.66 | 0.27 | +0.24 | +0.158 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.64 | 0.80 | 0.93 | 0.79 | 0.29 | +0.26 | +0.205 |
| Audience breadthHow many people who build technology does it concern? | 0.57 | 0.60 | 0.72 | 0.63 | 0.15 | +0.13 | +0.082 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.71 | 0.70 | 0.84 | 0.75 | 0.14 | +0.22 | +0.165 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 1.00 | 0.85 | 0.91 | 0.92 | 0.15 | +0.05 | +0.046 |
| Technical depthHow technically substantive is it? | 1.00 | 0.90 | 0.96 | 0.95 | 0.10 | +0.05 | +0.048 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.89 | 0.85 | 0.98 | 0.91 | 0.13 | +0.05 | +0.045 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.04 | 0.05 | 0.03 | 0.04 | 0.02 | −0.19 | −0.008 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.22 | 0.20 | 0.08 | 0.17 | 0.14 | −0.11 | −0.018 |
| Sum of contributions | +0.723 | ||||||
| Claim-check penalty | −0.100 | ||||||
| Total, as published | 0.623 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-13/63/"><img src="https://rubric.hrih.si/badge/2026-11-13/63.svg" alt="Ranked #63 on Rubric, 13 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 1 not in source · 7 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
The author's accuracy on their benchmark with a vision-language model dropped from about 70% to 20% after enabling output_attentions=True.
The article gives no link or citation for it.
The article says “its accuracy on my benchmark dropped from about 70% to 20% after I enabled output_attentions=True”
The source says No source to compare with.
No source givenNumber
Across 20 COCO images, Qwen3-VL-4B-Instruct in BF16 produced different greedy output sequences in 17/20 images for eager vs SDPA, 17/20 for eager vs FA2, and 15/20 for SDPA vs FA2.
The article gives no link or citation for it.
The article says “15-17 out of 20 generations result in different output sequences across the three attention backends”
The source says No source to compare with.
No source givenNumber
Eager vs SDPA produced different sequences in 17/20 cases at BF16, 6/20 at FP16, and 0/20 at FP32, with mean next-token logit RMSE falling from 0.11302 to 0.016291 to 0.000012276.
The article gives no link or citation for it.
The article says “FP32 0/20 0.000012276”
The source says No source to compare with.
No source givenGeneralization
Differences across backends also appear in text-only models: SDPA vs eager differed in 18/20 prefixes for Qwen3-4B and 19/20 for GPT-2.
The article gives no link or citation for it.
The article says “Differences across generated outputs are found in the tested text-only models too and it is not exclusive to VLMs”
The source says No source to compare with.
No source givenCausal conclusion
Precision differences are an important factor in model behavior, as suggested by prior work (Qi et al., 2025).
The article gives no link or citation for it.
The article says “This pattern suggests that float number precision also plays an important role in determining model's behavior, as suggested in prior work (e.g. Qi et al., 2025”
The source says No source to compare with.
Not in sourceDate
Recovering attention weights by replacing value vectors with identity matrices was suggested in a PyTorch discussion.
The cited source does not contain it.
The article says “This idea was suggested in this PyTorch discussion”
The source says “It would be great to add an argument that optionally returns the attention weights”
Source cited by the article: github.com/pytorch/pytorch/issues/119811
No source givenNumber
Across all 1,620 comparisons, reconstruction errors of the recovered attention range from 0.086% to 0.297%, with a mean of 0.182%.
The article gives no link or citation for it.
The article says “reconstruction errors range from 0.086% to 0.297%, with a mean of 0.182%”
The source says No source to compare with.
No source givenNumber
For Qwen3-VL under FA2, mean reconstruction error is 0.192% with FA2-recovered weights versus 10.593% with eager-returned weights.
The article gives no link or citation for it.
The article says “the mean errors are 0.192% with FA2-recovered weights and 10.593% with eager-returned weights”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 15 Nov 2026, 20:42 UTC. It played no part in the ranking.