aggregation

LiBRA proposes using class-conditional logit fingerprints of client models instead of parameter vectors for Byzantine-robust client selection in federated learning, instantiating Krum, Multi-Krum and Bulyan variants that preserve robustness while requiring far fewer comparisons and no access to client parameters.

一款 GitHub 上的开源 AI 新闻雷达项目，聚合 X、Telegram、Reddit、GitHub、Hacker News、RSS 等多源信息，利用 AI 自动过滤、评分、去重，并生成中英文每日简报。

Jim Gray 等人的经典论文，提出 data cube 关系聚合算子，将 GROUP BY 推广到 N 维聚合，并讨论高效计算方法，其中许多特性后来被纳入 SQL 标准。该论文是 OLAP 与数据仓库领域的奠基性工作。

A blog post explaining the simple two-stack algorithm for maintaining sliding-window aggregations (min, quantiles, HyperLogLog, etc.) with amortized O(1) operations, which works for associative aggregations without inverses, and tracing the algorithm's folklore origins back through Tangwongsan et al. and D. Sleator's lecture notes.

The open-source tool last30days-skill surged to GitHub Trending's top spot with 7,000+ stars in a week (13,000 total), scanning 10 major platforms—including X, Reddit, YouTube, and Polymarket—to condense the past 30 days of buzz into a single cross-verified report.

This preprint challenges aggregate independence metrics for LLM judge panels, showing that verification signals only improve accuracy on pivotal 'one-vote-margin' queries, and proposing a margin-stratified call-reduction rule.

Introduces a model-side aggregation interface using compact HyperLogLog sketches to maintain set-based aggregation states for long-context language models, achieving 99.2% accuracy on benchmark tasks and eliminating the need for external execution cycles.

This paper introduces an uncertainty-aware trust estimation method for aggregating predictions from multiple LLMs, adapting structured expert judgment with Cooke-style log weighting to penalize overconfident incorrect predictions. Evaluations on MMLU and MMLU-Pro show that this approach achieves superior accuracy-reliability balance under heterogeneous and contaminated expert panels.

This paper investigates whether aggregating probability estimates from multiple LLMs exhibits a wisdom-of-crowds effect, finding that learned aggregators outperform individual models and that training cutoff contamination is a pervasive confound in such evaluations.

This paper investigates how multimodal speaker verification using audio, prosodic, and linguistic cues across multiple utterances can significantly degrade speaker anonymization, reducing equal error rate by over 15% with only five utterances.

This paper proposes using model merging techniques, specifically Iso-C aggregation, to improve the aggregation step in DiLoCo distributed training, resulting in a new method called IsoLoCo that outperforms DiLoCo on language model pre-training.

PRA-RAG is a provably robust aggregation algorithm for Retrieval-Augmented Generation that defends against poisoning attacks on retrieved texts. It uses geometric structures in the embedding space to identify robust subsets and provides theoretical bounds on attack impact, reducing attack success rate to as low as 1% while maintaining accuracy.

This paper investigates whether LLMs can approximate swarm intelligence effects through intra- and inter-model aggregation, finding significant error reductions up to 37 percentage points in MAPE across eight estimation tasks.

A developer discusses the limitations of RAG for aggregation and counting queries over collections of documents, and asks for community advice on alternative approaches like text-to-SQL and intent routing.

Franz 6 unifies all messaging apps into a single window with built-in private AI features.

This paper reveals that aggregating complete reasoning traces from multiple LLM agents, rather than just their final answers, can correct errors even when agents unanimously agree, introducing the 'aggregation paradox' and the Self-Consistent Mixture of Agents method.

This is an open-source project on GitHub that aggregates ebook resources from multiple Chinese reading apps, categorizes them, and supports downloading in multiple formats.

FreeLLMAPI is an open-source tool that aggregates free tiers from 11 major LLM providers into a single OpenAI-compatible endpoint, routing requests and managing rate limits to deliver ~1B+ tokens per month. It simplifies access to multiple free models through one local server.

Indigo is a tool that helps you stay in touch with your contacts across Bluesky and Mastodon.

Krater provides access to multiple AI tools under one subscription, simplifying tool management.