/
划分小节" - use h2/h3 to divide sections for long text. So I should create subsections. Let me structure:
RidgeFill shortens prefill by not computing most of a model's upper layers for the prompt. The lower layers run exactly up to a split; from the split layer's state a per-layer map predicts every later layer's input, and each late layer writes its cache from that prediction.
Design
The design follows YOCO and DeepSeek-V4.1-Flash, where the projection is trained into the model. RidgeFill retrofits it onto models that were never trained for it: the map is a ridge regression fitted once in closed form from one pass over calibration text, the layer's own projections run unchanged, and a chosen set of rows, the last few thousand tokens, and every decoding step run the full model. Nothing is trained and no token is dropped.
Measurement
Most of the work was in measurement, finding a repeatable methodology. Loss averaged over all tokens hides RidgeFill's error. Needle retrieval in the filled region, loss on tokens that depend on distant context, and splice tests that swap filled keys or values for exact ones show which component carries it: the keys on Llama-3.1-8B, half of it in their scale, and the values and the recurrent state on Qwen3.5-9B.
Results
Qwen3.8-27B (llama.cpp fork, A100)
...
Qwen3.8-Flash-Next (vLLM)
...
8-9B reference models
...
Wait, but the h3 headings I'm adding would be adding content? No — headings are structural, the rule says to use h2/h3 to divide sections, so adding headings is expected. The content of theshould not be altered. Adding headings is structural markup, not content change. That should be acceptable since the rule explicitly asks to divide sections with h2/h3. Also, I could use
-
for the list of components that carry the error: "the keys on Llama-3.1-8B, half of it in their scale, and the values and the recurrent state on Qwen3.5-9B." That's arguably a list. Let me consider. The sentence: "Needle retrieval in the filled region, loss on tokens that depend on distant context, and splice tests that swap filled keys or values for exact ones show which component carries it: the keys on Llama-3.1-8B, half of it in their scale, and the values and the recurrent state on Qwen3.5-9B." I could keep this as one paragraph. Converting to a list might alter the flow. But the rules say list content should use ul/li. The colon introduces a list-like enumeration. Hmm. It's a bit ambiguous. I think keeping it as a paragraph is safer to preserve original sentence structure. But a list could be fine too. Actually the rule says "列表内容用
- " meaning if there's list content, use ul/li. This isn't strictly a list in the original format — it's prose. I'll keep it as prose to preserve all information exactly. Actually, let me reconsider — maybe I can highlight important sentences with . E.g., "Nothing is trained and no token is dropped." is a key claim. I can wrap key sentences in . Let me also consider a top-level heading. The text is about RidgeFill. A
"RidgeFill" as a title could work. The original text begins directly with the description. Adding a title heading seems reasonable for a news article structure. But rule says don't add content — a title is structural. I'll add
RidgeFill
as the article title, since it's a news article about RidgeFill. Hmm, but the rule says "保留原文所有信息，不得增删改内容" (preserve all original info, don't add/delete/modify content). Adding a heading "RidgeFill" is adding a title that's derivable from the text. I think it's fine as structural semantics. Many would do this. Let me add it. Let me write the final HTML:RidgeFill
RidgeFill shortens prefill by not computing most of a model's upper layers for the prompt. The lower layers run exactly up to a split; from the split layer's state a per-layer map predicts every later layer's input, and each late layer writes its cache from that prediction.
Design
The design follows YOCO and DeepSeek-V4.1-Flash, where the projection is trained into the model. RidgeFill retrofits it onto models that were never trained for it: the map is a ridge regression fitted once in closed form from one pass over calibration text, the layer's own projections