MONDAY · SEPTEMBER 21, 2026 · ISSUE 14 · SINCE FRIDAY No. 14
Daily News — 2026-09-21
273 papers indexed on arXiv·~10,000 packages released on PyPI·24 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
A 552B MoE that cuts KV cache storage fourfold in HBM and eightfold on SSD without sacrificing benchmark parity—the engineering case for memory-first model design.
The central problem DeepSeek-V4.1-Flash attacks is not intelligence but memory: KV caches have become the dominant cost in long-context serving, straining HBM, SSD, and interconnect bandwidth simultaneously. The paper's…
The capability primitive — composable, decomposable, deferrable — is what separates this from a model-switching wrapper.
The capability abstraction is the real organizing idea here. A capability bundles tools, instructions, hooks, and model settings into a single composable unit that attaches to any agent with one list entry.
Papers
A 552B MoE that cuts KV cache storage fourfold in HBM and eightfold on SSD without sacrificing benchmark parity—the engineering case for memory-first model design.
DeepSeek-V4.1-Flash (552B MoE) cuts global KV cache to 1/4 and persistent KV cache to 1/8 of V4-Flash via CSA2 cross-layer reuse and FP4 caching.
A benchmark that exposes the gap between supporting omni-modal inputs and actually reasoning across them — MiniMax-H3 clears that bar less than half the time.
MiniMax-H3 hits 41.97% overall on a new physical-world reasoning benchmark, with audio-based tasks its weakest point at 27.40%.
Four harness mechanisms discovered by an AI search loop cut token traffic by nearly half with minimal score loss—and the savings compound across agent swarms.
SoL-Pi is a coding-agent harness that reduces token traffic by 44.7-49.0% and API cost by about one third on the 51-task EdgeBench evaluation.
LLMs are measurably worse at social reasoning when they hear about events through a user's biased retelling — and more turns don't reliably fix it.
Fuse benchmarks LLM social reasoning through user-mediated simulations, revealing that biased framing and mediation itself compound model errors.
Reframes reasoning efficiency as allocation rather than compression, with a reward mechanism that actually preserves hard-problem accuracy while cutting easy-problem token waste.
A post-training framework that allocates reasoning depth per problem difficulty, boosting AIME24 Pass@3 by 10.0% while cutting token usage by 27.9%.
Tools & packages
A tight, honest market-making loop that treats the 300 ms block window as a hard constraint and engineers every component around it.
AI agent firing one trade decision per Monad block on the Kuru MON-USDC market.
A multi-hop translation chain that uses linguistic distance as a deliberate obfuscation mechanism — clever architecture, but its durability against adaptive detectors is an open question.
An open-source tool that rewrites AI-generated text to bypass AI detectors such as Turnitin or GPTZero.
The capability primitive — composable, decomposable, deferrable — is what separates this from a model-switching wrapper.
pydantic-ai brings Pydantic's validation model directly into AI agent construction, keeping structured I/O as a first-class concern.
Knitted window borders with per-app hand-picked colourways, built on private macOS APIs and honest about every rough edge.
A macOS menu-bar app that dresses windows in knitted borders.
A well-organized, employer-specific question bank that maps what 35 AI companies actually probe in interviews, with enough loop detail to calibrate preparation depth.
A collection of AI engineering interview questions and answers organized by company.