The paper introduces Corpus Task Complexity (CTC), a metric that captures how a task’s difficulty scales with corpus size. It distinguishes low‑CTC tasks, whose difficulty grows linearly, from high‑CTC tasks, whose difficulty grows quadratically or more, and presents ten new high‑CTC tasks. Experiments show that models performing well on low‑CTC tasks often fail on high‑CTC tasks, highlighting the need for new approaches to large‑corpus reasoning.
Given a large corpus, the questions one might ask can vary -- from "When was the first human heart transplant?" to "What are all the contradictory claims in this literature?" -- but what makes some qu...
arXiv:2607. 08284v1 Announce Type: new Abstract: Large language models (LLMs) have demonstrated rapidly improving long-context capabilities, prompting a wave of benchmarks designed to evaluate them.
Retrieval-augmented and agentic question-answering systems increasingly re-derive the meaning of a corpus at query time. Put plainly, instead of re-deriving what a corpus means on every question, the work is done once when a document arrives and is thereafter merely consulted -- a compiler, not an interpreter, of meaning.
arXiv:2601. 03043v4 Announce Type: replace-cross Abstract: Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference efficiency.
arXiv:2504. 17768v3 Announce Type: replace-cross Abstract: Sparse attention offers a promising strategy to extend long-context capabilities in Transformer LLMs, yet its efficiency-accuracy trade-offs remain unclear due to the lack of comprehensive evaluation.
arXiv:2608. 16621v1 Announce Type: new Abstract: Retrieval-augmented and agentic question-answering systems increasingly re-derive the meaning of a corpus at query time.
Large language models (LLMs) have demonstrated rapidly improving long-context capabilities, prompting a wave of benchmarks designed to evaluate them. However, existing long-context evaluations - from Needle-in-a-Haystack (NIAH) tests to more recent multi-hop reasoning and summarization tasks - predominantly measure average-case performance, and many are either saturated or lack robustness.
arXiv:2605. 16928v2 Announce Type: replace-cross Abstract: Long-context inference in large language models is bottlenecked by the quadratic cost of full attention.
The paper argues that retrieval‑augmented question‑answering systems should perform semantic compilation at ingest time rather than re‑deriving meaning at query time. By building a maintained structure—incrementally updated embeddings and validated atomic claims—read operations become far cheaper, with experimental results showing higher accuracy and lower token usage compared to traditional chunk‑based retrieval. The authors present two proofs: cheaper incremental updates and superior performance on broadcast‑interview transcripts, suggesting a new systems agenda for compilation and read planning.
The paper investigates how Tabular Foundation Models (TFMs) can achieve strong transfer learning by self‑supervised pre‑training on a single real table. It finds that a table’s usefulness is largely determined by its number of features rather than instances, and that fine‑grained column‑level preprocessing improves downstream performance while dataset‑level filtering does not. The authors propose that tabular in‑context generalization is primarily retrieval‑based, with models learning to identify and aggregate relevant examples from the provided context.
Block Sparse Flash Attention (BSFA) is a drop‑in replacement for FlashAttention that speeds up long‑context inference by pruning about 50% of computation and memory transfers. It selects the top‑k most important value blocks for each query using exact query‑key similarities and calibrated per‑layer, per‑head thresholds, requiring only a one‑time training‑free calibration. On Llama‑3.1‑8B, BSFA delivers up to 1.13× speedup on LongBench with a 1.1% accuracy drop and up to 1.24× on Needle‑in‑a‑Haystack retrieval with a 1% drop, while the attention kernel itself accelerates by up to 1.38×.
arXiv:2602. 15829v2 Announce Type: replace Abstract: The superficial alignment hypothesis (SAH) posits that large language models learn most of their knowledge during pre-training, and that post-training merely surfaces this knowledge.