Today's Overview
- DoSA Safely Skips Candidates That Cannot Reach Top-K. Provable bounds avoid full scoring for rejected candidates, delivering up to 1.94x and 1.50x time-to-first-token speedups over SGLang and LiteTopK, respectively.
- Real Behavior Traces Can Train General Abilities. Vela generated 100,000 tasks from anonymized behavior records of 100,000 users and retained roughly 75,000 after filtering. Fine-tuning gained up to 11.5 points on user-related tasks and averaged four points on general benchmarks. Reuse still requires consent, representativeness, and bias controls.
- Deep Research Agents Still Struggle With Everyday and Professional Work. LifeSearch evaluates 203 tasks derived from real-user query seeds against more than 4,000 detailed rules covering numbers, dates, and URLs. The best evaluated agent scores only 63.4%.
Featured
01 Sparse Attention Still Scans Everything
Sparse attention reads only a small set of keys. Yet its indexer often scores every candidate to find them, leaving the full-scan cost intact.
DoSA first uses seed candidates to establish a lower bound for the Top-K cutoff. It then calculates provable upper bounds on each attention head's remaining contribution. These bounds cover uncomputed positive contributions and possible negative deductions. DoSA skips a candidate only when its best possible score remains below the cutoff. Otherwise, it completes the score and selects exactly the same keys as the original method.
Savings depend on the candidate score distribution and how clearly the Top-K boundary separates winners from losers. The complete indexer runs up to 2.74x faster than SGLang and 1.69x faster than LiteTopK across English, Chinese, and code workloads. On GLM-5.2 and DeepSeek-V3.2, time to first token improves by up to 1.94x and 1.50x over those respective baselines. The full paper must show whether those peaks hold at common context lengths and batch settings. Teams deploying long-context models should examine indexing costs, not just how many keys sparse attention reads.
Key takeaways:
- Sparse reads do not mean sparse retrieval. The indexer's full scan may be the hidden bottleneck.
- DoSA skips scoring only when a provable upper bound falls below the Top-K cutoff, preserving the selected keys.
- Evaluate both indexer speed and first-token latency. Peak local gains do not guarantee consistent end-to-end gains.
Source: DoSA: Dots Sparse Attention
02 Product Logs Can Train General Ability
Vela turns real behavior sequences spanning weeks or years into training tasks and trajectories. It generated 100,000 tasks from anonymized behavior records of 100,000 users across 91 real-world task templates. About 75,000 tasks remained after filtering.
Fine-tuning Qwen3.6-35B-A3B on this data raised user-related task scores by up to 11.5 points. General benchmark scores increased by four points on average. Fine-tuning before reinforcement learning also beat reinforcement learning directly from the base model by up to 7.7 points. High-quality behavior traces can provide supervision that static text and synthetic data miss.
Anonymity does not equal permission. It also cannot automatically remove sampling gaps, platform-induced behavior, or historical bias. Teams with product logs should establish consent, usage boundaries, and offline validation before deciding which traces to reuse. User behavior is not free training data.
Key takeaways:
- Product logs can become model supervision, not just personalization context.
- Behavior traces may improve both user understanding and general capabilities.
- Resolve consent, representativeness, and bias validation before reusing real behavior.
Source: Vela: User Behavior as Supervision Sparks LLM Intelligence
03 The Best of 12 Evaluated Agents Scores 63.4% on LifeSearch
LifeSearch tests 12 agents on 203 Chinese- and English-language tasks derived from real-user query seeds. The best system reaches only a 63.4% aggregate score.
The benchmark decomposes those tasks into more than 4,000 detailed rules. It checks coverage and accuracy while verifying numbers, units, dates, and URLs individually. Current systems still struggle to deliver results that users can act on and verify item by item.
Its scoring correlates closely with human rankings, reaching a Spearman coefficient of 0.89. Specific failure modes and domain-level differences still require confirmation from the full paper. Deep research products should prioritize structured fact-checking and source tracking over more polished prose.
Key takeaways:
- Academic question-answering scores cannot substitute for reliability on real decisions.
- Agent evaluations should verify numbers, dates, units, and source links individually.
- Product work should prioritize fact verification and evidence tracking over presentation quality.
Source: LifeSearch: Benchmarking Deep Research Agents for Real-World Life and Productivity
