[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-lighton-rerank-listwise":3,"news-related-b3a59736-4791-4834-8dc2-a851833a63ae":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"b3a59736-4791-4834-8dc2-a851833a63ae","LightOn-rerank：2B 模型同时排文本和文档页，listwise 把 pointwise 打成过去式","LightOn AI 7 月 16 日放出 LightOn-rerank，用 Qwen3.5-2B 加 LoRA 同时搞定文本段落与文档页图像的多模态重排序，是开源 2B 档里第一个把两个模态塞进同一个 adapter 的工作。\n\n它最值得讲的不是分数，是范式。LightOn 直接抛弃 pointwise scoring 让每个候选独立打分的老路，改为 generative listwise：query 加 4 个候选一次性进 forward pass，模型输出 [2]>[4]>[1]>[3] 这样的排列 token。同样的 213K 组训练数据，listwise loss 一旦换成逐文档独立打分，ViDoRe V3 直接掉 10.8 个 NDCG 点；换回 4 候选同窗口后又涨回 62.66。结论很硬：重排序的好坏不在 loss 形式，而在候选之间能否互相 attend。\n\n跨尺度验证同样反直觉：0.8B \u002F 2B \u002F 4B 三档做 grid，pointwise 在 2B → 4B 几乎纹丝不动（−0.1 NDCG），listwise 继续涨（+2.0）；4B listwise 拿到 64.69，直接压过 Qwen3-VL-Reranker-8B 的 64.23，参数减半却更准。\n\n部署侧也有干货。CUDA event 拆下来，2B listwise 每个 window 里 ViT 编码器占 46%、prefill 16%、decode 才 38%——做文本 rerank 时常用的 first-token readout 在这里几乎不省时间，因为 4 个候选的图像编码不依赖 decode 状态。更划算的杠杆是砍候选数：top-20 重排序保留 85% lift、只用 1\u002F5 window；top-10 砍掉一半 lift 但只用 1\u002F12。第一阶段召回曲线决定深度，不决定模型。\n\n比起堆数据、堆大模型，LightOn 给出的更像一份工程方法论：把候选比较写进前向，把工程取舍写进配置文件。模型不是越准越好——2B listwise 已经够用，但得让它真去比较。\n","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Flightonai\u002Flighton-rerank","24d5c6c5-6573-4180-a1fd-f1459842d1af",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"eeb5cb30-daa2-40a7-a77c-527a2b9b1c5e","en","LightOn-rerank: one 2B model ranks text and document pages","LightOn AI released LightOn-rerank on July 16, using Qwen3.5-2B + LoRA to simultaneously handle multimodal reranking of text passages and document page images — the first open-source 2B-tier work to pack both modalities into the same adapter. What's worth talking about isn't the score, but the paradigm. LightOn directly abandons the old road of pointwise scoring where each candidate is scored independently, and instead goes generative listwise: query + 4 candidates go through a single forward pass, and the model outputs a ranking token such as [2]>[4]>[1]>[3]. On the same 213K training pairs, once the listwise loss is replaced with per-document independent scoring, ViDoRe V3 drops a full 10.8 NDCG points; switching back to 4 candidates in the same window it climbs back to 62.66. The conclusion is hard: the quality of reranking lies not in the loss form, but in whether the candidates can attend to each other. Cross-scale validation is equally counter-intuitive: a grid across 0.8B \u002F 2B \u002F 4B shows pointwise barely moves from 2B → 4B (-0.1 NDCG), while listwise keeps climbing (+2.0); 4B listwise hits 64.69, directly beating Qwen3-VL-Reranker-8B's 64.23 — half the parameters, more accurate. There's also substance on the deployment side. CUDA event breakdown shows that in each 2B listwise window, the ViT encoder takes 46%, prefill 16%, and decode only 38% — for text rerank the common first-token readout hardly saves time, because the image encoding of 4 candidates doesn't depend on decode state. A more cost-effective lever is cutting the candidate count: top-20 rerank keeps 85% of the lift using only 1\u002F5 window; top-10 cuts the lift in half but uses only 1\u002F12. The first-stage recall curve determines depth, not the model itself. Rather than piling on data or larger models, LightOn's contribution reads more like an engineering methodology: write candidate comparison into the forward pass, write engineering tradeoffs into the config file. The model doesn't have to be more accurate — 2B listwise is already enough, but it has to actually compare.","lighton-rerank-listwise","2026-07-16T22:30:00Z","2026-07-16T22:09:46.219043Z","2026-08-19T02:08:40.142862Z",true,"agent",100,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","MiniMax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"4c7f5330-3aff-458a-9ef5-f04cc5585703","微信视觉团队开源 WeMM 嵌入模型:2B 反超 8B 前基线,9B 达 MMEB-v2 80.6","wemm-embedding-wechat-multimodal","2026-08-26T21:07:30+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"7ef479ae-66af-463a-802f-07a84ade93b1","商汤开源 SenseNova-U1.5-8B：原生多模态通吃生成编辑，短板全写进模型卡","sensenova-u1-5-8b-open-source-multimodal","2026-08-25T19:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"ce70384a-990b-4994-bfb6-27775be45661","TensorRT Edge-LLM 0.10.0：边端第一个统一的 C++ 多模态推理栈","tensorrt-edge-llm-0-10-multimodal-runtime","2026-08-23T00:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"96b989b7-992b-424e-a8c1-1568760150c1","小红书开源 dots3-note:280B MoE 多模态、512K 上下文,Apache 2.0 直接放行","dots3-note-preview-280b-open-weights","2026-08-18T23:10:00+00:00"]