Interpretable by Design Tabular Prediction via In-Context Local Linear Learning

Okay, the user wants me to take a plain text news/abstract and format it as semantic HTML. Rules: only HTML fragment, no markdown fences, no explanations, no doc-level tags. Each natural paragraph in

, use

/

for sections,

• for lists,

  for quotes,

   for code,  for important sentences. Preserve all original content, no additions/removals/changes.
  
  Let me identify the structure of the text. It's an academic abstract about TICL³ (Tabular In-Context Local Linear Learning). Looking at the flow:
  - First sentence: introduces in-context learning (ICL) for tabular foundation models
  - Second: strengths and difficulty (accuracy but interpretability hard)
  - Third: example in biology
  - Fourth: post-hoc methods limitations
  - Fifth: introduces TICL³
  - Then details of method
  - Then results
  
  The abstract naturally divides into: (1) background/motivation, (2) method description, (3) results. I can use 
   for the section on the proposed method and results. Let me check what makes sense. The main body is one continuous paragraph. I could split it with a 
   before "We introduce TICL³" or before the results. Actually, the method and results could be separate 
   sections to give it structure.
  
  Let me plan:
  - Opening paragraph: ICL intro, accuracy, interpretability difficulty, biology example, post-hoc limitations → wrap in 
  
  - 
   heading for "We introduce TICL³..." → method
  - 
   for method description
  - 
   heading for results → "Across 386 classification..."
  - 
   for results
  
  Now, important sentences to highlight with . The key claims: "strong predictive accuracy", "interpreting how individual features contribute... remains difficult", the TICL³ name and description, "The shared model captures broad effects, while the local model corrects errors among similar samples", "the same model makes and explains the prediction", and the quantitative results.
  
  Math notation: TICL$^3$ appears multiple times. In HTML, I can write it as TICL3 or keep the $ notation? The user said preserve all original info. The $^3$ is LaTeX-ish. I should convert to 3 for proper HTML rendering, or keep as TICL$^3$? Converting $^3$ to 3 is a formatting translation that preserves meaning. Hmm, but the rule says don't change content. The superscript is content. I'll use 3 since that's the HTML-semantic equivalent. Actually, wait — the user said "保留原文所有信息" (preserve all original info). Changing $^3$ to 3 preserves the information (superscript 3). That should be fine as it's HTML formatting. Alternatively, keep "TICL$^3$" literally to be safe. Let me think about what a "排版工程师" would do — they'd render the superscript. I'll use 3. Hmm, but there's risk. Let me reconsider. The original text uses "TICL$^3$" twice. In HTML, writing TICL3 is the semantic equivalent and preserves information. I'll go with 3.
  
  Actually, to be maximally safe and faithful, maybe keep it as TICL3 — yes, that's a faithful HTML rendering.
  
  Now, should I use a list? There's no explicit list in the text. The method has a sequence: "first learns... Then, for each query..." — that's a description, not a list. I shouldn't force a 
  . But the method steps could be... no, the rule says preserve content, don't add. Keep as paragraph.
  
  Quotes: none in original. No 
  
             needed. Code: none. No 
             
  .
  
  Let me also decide heading levels. Since this is a single abstract, maybe one 
   for the method and one 
   for results. Or an 
   for method and 
   for results with an 
   overall? The rule says "长文用
  /
  划分小节" — long texts use h2/h3 to divide sections. This is relatively short. I'll use 
   for the two subsections: "Method" and "Results"? But wait, I shouldn't invent headings — "不得增删改内容" (no adding/removing/changing content). Adding heading text would be adding content. Hmm, that's a constraint. If I add a 
   with "Method" text, that's adding words. But headings are structural. I could use empty headings? That's weird.
  
  Actually, I can use 
   tags to wrap/divide without inventing much — but any heading needs text. Alternatively, I can just use 
   tags with descriptive labels like "Method" and "Experimental Results" — that's adding minimal structural text. Many would consider that acceptable formatting, but the rule is strict: 不得增删改内容. Hmm.
  
  Wait, let me re-read: "长文用
  /
  划分小节" — use h2/h3 to divide sections for long texts. The instruction implies adding headings is part of the task. But also "保留原文所有信息，不得增删改内容". There's tension. A safe approach: use minimal heading labels derived from the content, or avoid headings if not needed. But the task says to use h2/h3 to divide sections. 
  
  Option A: Split the paragraph with 
   tags that contain actual content, e.g., take the first sentence as a 
  , then... no.
  
  Option B: Use headings like 
  Method
   and 
  Results
  . This adds words but provides structure. Given the task explicitly asks for h2/h3 section division, adding short section labels seems within scope. However, strict "no additions" might prohibit it.
  
  Option C: Don't add headings at all; just use 
   for each natural paragraph. The text is one continuous paragraph though. Could split into 2-3 paragraphs. The original has clear semantic breaks. I can split into: (1) background & motivation, (2) TICL³ method, (3) results. Each as 
  . That preserves content without adding anything.
  
  Hmm. The task says "长文用
  /
  划分小节" (for long texts, use h2/h3). This is a short abstract. I think the safest faithful approach is: split into logical paragraphs with 
  , and optionally use 
   for the method/results sections with minimal labels. 
  
  Let me think about what's most defensible. The user emphasized "保留原文所有信息，不得增删改内容" strongly (rule 4). Adding heading text like "Method" is technically adding. But the sectioning instruction (rule 2) suggests headings are expected. I could use headings that echo existing content, e.g., no...
  
  Actually, a clean solution: split the text into paragraphs by semantic breaks. The original text has these natural breaks:
  1. ICL intro + accuracy + interpretability difficulty
  2. Biology example + post-hoc methods
  3. TICL³ introduction + method
  4. Results
  
  Or keep it simpler. Let me look at the abstract structure again:
  
  "In-context learning (ICL) lets tabular foundation models make predictions without dataset-specific training or fine-tuning. These models achieve strong predictive accuracy, but interpreting how individual features contribute to the prediction remains difficult. In biology, for example, feature contributions can help identify biomarkers or risk genes. Post-hoc interpretability methods assign feature importance scores, but these scores do not indicate what the model actually uses and can lead to explanation errors. We introduce TICL$^3$ (Tabular In-Context Local Linear Learning), an interpretable-by-design predictor that uses a tabular foundation model to guide a simple linear prediction rule. TICL$^3$ first learns a shared linear model to capture dataset-level feature effects. Then, for each query, it retrieves similar training samples using representations from the foundation model and fits a second linear model to the residuals in that neighborhood. The shared model captures broad effects, while the local model corrects errors among similar samples. Their combined coefficients determine each feature contribution to the final query-specific linear score, so the