Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-07-13 · source-backed
Test-time training for long-context LLMs is highly sensitive to which spans you train on. Random spans degrade accuracy because most are irrelevant (arXiv:2607.09415). S-TTT has the model first identify relevant evidence passages, then run adaptation only on those, for up to 15% relative gains on LongBench-v2 and LongBench-Pro with Qwen3-4B-Thinking and Llama-3.1-8B. The builder takeaway: TTT is practical for long documents if you gate what the model learns from at inference time. Don't adapt on the whole haystack.
Each link below shares sources, entities, or timing with this story.
Meta released Llama / Shared entities / Earlier coverage
Linked by a graph relationship (Meta released Llama); both cover Llama, Test; earlier Llama coverage from 2026-05-01.
Linked by a graph relationship (Meta released Llama); both cover Llama, Qwen3; earlier Llama coverage from 2026-04-02.
Meta released Llama / Shared entity: Llama / Earlier coverage
Linked by a graph relationship (Meta released Llama); both cover Llama; earlier Llama coverage from 2026-06-15.
Linked by a graph relationship (Meta released Llama); both cover Llama; earlier Llama coverage from 2026-05-02.
Meta released Llama / Shared entity: Test / Earlier coverage
Linked by a graph relationship (Meta released Llama); both cover Test; earlier Test coverage from 2026-04-28.
Meta released Llama / Shared entity: Llama / Earlier coverage
Linked by a graph relationship (Meta released Llama); both cover Llama; earlier Llama coverage from 2026-03-12.
Meta released Llama / Shared entity: LLMs / Earlier coverage
Linked by a graph relationship (Meta released Llama); both cover LLMs; earlier LLMs coverage from 2026-03-11.
Meta released Llama / Same source domain
Linked by a graph relationship (Meta released Llama); reported by the same outlet (arxiv.org).