Chinese contentMirendil 发布首个模型 Fern 0.1，上下文窗口达 1000 万 token
新公司 Mirendil 出了第一款模型 Fern 0.1，主打 1000 万 token 超长上下文，RULER 单针测试满分，不过写代码的智能体能力还比不过前沿模型。
#RULER
9 stories · 2 in 7 days · 5 in 30 days
Stories, products and papers tagged "RULER", newest first.
Oct 27
Oct 23
Papers11:27
Chinese contentTTR-Compressor：用测试时回归重做序列压缩一篇方法论文：把压缩变成回归问题，RULER 成绩超过注意力压缩器，视觉 token 还能砍掉一半多。
Oct 15
Chinese contentLatent Compaction：用测试时训练压缩 KV cache，32 倍压缩下 RULER 准确率达 96.7%
清华 THUNLP 开源了个 KV cache 压缩方法，压到 1/32 还能在 RULER 上保持 96.7% 准确率，比 KVZip+ 的 40% 高出一大截，长文档检索场景可以直接试。
Oct 6
LatentIndex cuts sparse-attention indexer cache by 61.1% on DeepSeek-V3.2
Sharing continuous latent representations instead of discrete token selections resolves the usual trade-off between cache reuse and per-layer accuracy in sparse attention. Engineering teams running long-context inference on DeepSeek-V3.2 or GLM-5 can apply the training-free calibration directly to cut indexer memory and speed up decoding. The 2.3-2.7 times decode speedups at 128K contexts make it especially relevant for local long-document and agent workloads.
Oct 5
Sep 24
Chinese contentKV-Cache 淘汰策略可按风险目标认证：Llama 上 SnapKV 认证通过 75% 保留率
这论文把 KV-cache 压缩变成了能签合同的事：给定风险目标，自动算出能砍多少缓存，砍不了就回退全 KV。做长文本推理部署的可以看看。
Sep 23
Chinese content论文提出 LYRA 方法破解长上下文检索中的 Proximity Trap
一篇很实在的论文：作者发现长文档检索差的真正元凶是近处无关文字抢注意力，给出 LYRA 方法外加 ProxBench 基准，做了 RAG 的话值得看他们的实验数据。
RULER: Rubric-Based RL Rewards for SVG Code Generation
Rubric-derived rewards offer a template for RL on any open-ended generation task where scalar metrics like CLIP trigger reward hacking, extending beyond SVG to other code-to-visual domains. The method needs no human preference labels, so small teams can replicate it with a VLM judge alone. Practitioners training generative models without ground truth should examine the rubric-design ablations, which the paper identifies as the decisive factor.
Jul 13