Recursive Sparse Attention: Million-Token Extrapolation at Logarithmic Cost
Comments on “Recursive Sparse Attention: Million-Token Extrapolation at Logarithmic Cost”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Haoai Lab's RecSA builds a tree over the KV cache so each query searches it level by level, keeping selection cost and entries read at O(log L). Unselected nodes stay in the softmax as summaries, and candidate-local RoPE keeps relative distances within the trained range. Starting from Qwen3-4B and training on 0.62B tokens at 32K, it scores 60.5 on RULER at 1M versus 40.0 for Dense + DCA, and decodes 3.8× faster at 524K on one H200. Results are the authors' own, on a 4B model; larger models are untested.
Why it ranked
No. 12 of 140 with a score of 0.68. The panel rated it highest on new information (0.95) and technical depth (0.94), lowest on audience breadth (0.61). Placed No. 12, No. 19 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Original LLM attention research with substantive technical detail and benchmarks on RULER and latency; significant for long-context AI, though evidence is single-paper and the relevance is mainly to AI builders.
- AI judgeJudge 3gpt-5.6-luna
- Original, technically detailed research presents a reproducible sparse-attention architecture, training method, benchmark results, and kernel implementation for million-token extrapolation. It is highly relevant to AI infrastructure, though some novelty and broad-limit claims are difficult to independently verify from the article alone.
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
0.68total under rubric v2.3
- Significance0.70
- Reader interest0.77
- Audience breadth0.61
- Practical relevance0.67
- Substantiated0.85
- Technical depth0.94
- New information0.95
- Counts against: Promotional0.07
- Counts against: Overclaim0.36
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? | 0.72 | 0.55 | 0.82 | 0.70 | 0.27 | +0.24 | +0.167 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.67 | 0.70 | 0.95 | 0.77 | 0.28 | +0.26 | +0.201 |
| Audience breadthHow many people who build technology does it concern? | 0.61 | 0.50 | 0.72 | 0.61 | 0.22 | +0.13 | +0.079 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.69 | 0.55 | 0.78 | 0.67 | 0.23 | +0.22 | +0.148 |
| 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. | 0.96 | 0.70 | 0.88 | 0.85 | 0.26 | +0.05 | +0.042 |
| Technical depthHow technically substantive is it? | 0.98 | 0.85 | 0.98 | 0.94 | 0.13 | +0.05 | +0.047 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.91 | 0.95 | 0.98 | 0.95 | 0.07 | +0.05 | +0.047 |
| 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. | 0.07 | 0.10 | 0.03 | 0.07 | 0.07 | −0.19 | −0.013 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.44 | 0.30 | 0.34 | 0.36 | 0.14 | −0.11 | −0.040 |
| Sum of contributions | +0.679 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.679 | ||||||
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/2026-11-20/12/"><img src="https://rubric.hrih.si/badge/2026-11-20/12.svg" alt="Ranked #12 on Rubric, 20 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 2 matches source · 6 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceBenchmark result
RecSA scores 60.5 on RULER at 1M tokens, versus 40.0 for the strongest baseline (Dense + DCA).
The cited source says the same thing.
The article says “scoring 60.5 at 1M against 40.0 for the strongest baseline (Dense + DCA”
The source says “RecSA scores 60.5 on RULER ( Hsieh et al., 2024 ), compared with 1.6 for Dense + YaRN, as shown in Figure 1 a. The strongest of the other baselines,…”
Source cited by the article: arxiv.org/abs/2611.07232
Matches sourceNumber
On one H200, RecSA speeds up prefill by 3.2× and decoding by 3.8× at 524K.
The cited source says the same thing.
The article says “On one H200, it speeds up prefill by 3.2× and decoding by 3.8× at 524K”
The source says “On one H200 GPU with batch size 1, RecSA accelerates prefill by 3.2\times and decoding by 3.8\times at 524K, compared with dense FlashAttention”
Source cited by the article: arxiv.org/abs/2611.07232
No source givenFirst, only or fastest
RecSA is the first sparse attention that both selects at O(log L) cost per query and runs far beyond its training length after training only on short sequences.
The article gives no link or citation for it.
The article says “RecSA is the first sparse attention that does both: it selects at $O(\log L)$ cost per query, and it runs far beyond its training length after training only on short…”
The source says No source to compare with.
No source givenBenchmark result
RecSA outperforms all baselines on RULER at every length from 131K to 2M, initialized from Qwen3-4B and trained on 0.62B tokens at 32K.
The article gives no link or citation for it.
The article says “RecSA scores highest on RULER at every length from 131K to 2M (Table 2”
The source says No source to compare with.
No source givenGeneralization
MoBA and DSA score every preceding block or token to choose what to read, so their per-query selection cost is O(L).
The article gives no link or citation for it.
The article says “to choose that part it scores every preceding block or token, so its selection cost per query is still $O(L)$”
The source says No source to compare with.
No source givenNumber
For Qwen3-4B, a dense decoding step with host-memory overflow takes 177 ms at 1M and 2.9 s at 2M.
The article gives no link or citation for it.
The article says “a dense decoding step takes 177 ms at 1M and 2.9 s at 2M”
The source says No source to compare with.
No source givenBenchmark result
RecSA's perplexity on nine PG-19 books stays between 10.9 and 11.1 from 131K to 1M, while the DCA baselines reach 14.6–19.2 and Dense + YaRN 211 at 1M.
The article gives no link or citation for it.
The article says “From 131K to 1M, RecSA stays between 10.9 and 11.1, while at 1M the DCA baselines reach 14.6–19.2 and Dense + YaRN 211”
The source says No source to compare with.
No source givenCausal conclusion
Because candidate-local RoPE keeps relative distances within the trained range, RecSA has no inherent limit on context length.
The article gives no link or citation for it.
The article says “Candidate-local RoPE keeps every relative distance within the range seen in training, so RecSA has no inherent limit on context length”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 4 pts, not on front page
Recorded for context only, checked 20 Nov 2026, 05:47 UTC. It played no part in the ranking.