"Memory Caching: RNNs with Growing Memory" Google's new paper proposes a simple way to give recurrent models a memory that grows with sequence length. So instead of forcing an RNN to compress the full past into 1 fixed hidden state, it caches memory checkpoints across
@askalphaxiv
-
Malicious Intermediary Attacks on LLM Agent Supply Chain Security
By
–
"Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain" The danger of agent security can also exist in the API router, that is between the agent and the provider. As these routers can read prompts, keys, and tool calls in plaintext, even rewrite
-
SkillClaw: Collective Skill Learning Through Agentic Evolution
By
–
"SkillClaw: Let Skills Evolve Collectively with Agentic Evolver" As most AI agents still relearn the same targets from scratch, SkillClaw turns it into shared learning, where it collects agent trajectories across users, groups them by skill, and uses an agentic evolver to spot
-
Interleaved Head Attention Improves Transformer Architecture
By
–
“Interleaved Head Attention” A core limitation of transformers is that standard attention gives you H isolated heads, which means you only get H independent attention patterns. So this paper lets heads mix before attention by creating pseudo-heads from learned combinations of
-
PaperOrchestra: Multi-Agent Framework Automates AI Research Paper Writing
By
–
"PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing" This paper builds a multi-agent system for AI research writing. Each part of it handles one job: one plans, one does literature search, one makes figures, one writes, and one revises. This treats
-
ByteDance In-Place Test-Time Training Updates Model Weights
By
–
New Test-Time Training/Continual Learning technique from ByteDance Seed "In-Place Test-Time Training" It lets the model update part of its own MLP weights during inference to store useful information from the current prompt. With no new exotic architecture, it simply just
-
Meta Neural Computer: Model as Integrated Computing System
By
–
What if the model didn’t just use a computer, but actually was the computer? Meta AI introduces "Neural Computer", a model where computation, memory, and I/O are all inside one learned system. Their early prototype learns from screen recordings of terminals and desktops, and it
-
TriAttention: Efficient Long Reasoning with Trigonometric KV Compression
By
–
“TriAttention: Efficient Long Reasoning with Trigonometric KV Compression”
— alphaXiv (@askalphaxiv) 8 avril 2026
Most KV-cache compression methods guess what to keep by looking at recent attention.
But this paper argues that the signal is unstable because RoPE keeps rotating queries with position, so what looks… pic.twitter.com/GrEYc0gZ2b“TriAttention: Efficient Long Reasoning with Trigonometric KV Compression” Most KV-cache compression methods guess what to keep by looking at recent attention. But this paper argues that the signal is unstable because RoPE keeps rotating queries with position, so what looks unimportant now may matter later. So they proposed TriAttention, which looks in the pre-RoPE space and finds that many heads have stable Q/K centers. That lets it predict which token distances a head is likely to retrieve, and compress the KV cache using that structure rather than noisy recent attention. This shift from "keeping what was attended recently” to “keeping what this head is likely to need later” Empirically, it matches full attention on AIME25 with 2.5x higher throughput or 10.7x less KV memory.