Paper Summary

Block rules

1. Background / existing techniques

• Name the prior art explicitly (give acronyms if the field uses them: PDC/PIC, GRPO, RoPE, etc.)
• State each one's limitation in one clause, not a paragraph
• If there are two competing approaches, contrast them (e.g., "X is fast but rigid; Y is flexible but expensive")
• Skip generic motivation ("LLMs are important..."). Start from the technical baseline

2. Key problem & insight

• One sentence identifying the specific failure mode the paper targets
• One sentence stating the core insight (the "aha") that makes the method possible
• This is the block where a reader decides whether to keep reading — make it concrete

3. Proposed method

• Lead with the method name and a count of its components ("CacheSlide with three core components")
• For each component:
  • Bold the component name (use the paper's actual name, not a paraphrase — grep the paper if needed)
  • Start with a one-line high-level idea (what it does, why it exists)
  • Follow with the mechanism in 1-3 lines: the key variables, the decision rule, or the equation
  • When the mechanism has discrete steps (e.g., pretraining/inference phases, top-k selection + blending), use a sub-bullet list — but keep each sub-bullet ≤1 line
• Do NOT describe every algorithm line. Pick the parts that distinguish the method from the baseline
• If the paper has equations that are central (e.g., a loss function, a novel advantage formulation, a position encoding rule), include them with $\LaTeX$ — but only the ones a reader needs to understand the idea

4. Results

• Report the headline numbers with units (latency X×, accuracy +N points, throughput Y×)
• Name the baselines the paper beats (e.g., "vs. vLLM prefix caching", "vs. DAPO", "over GRPO")
• Skip ablations unless the ablation itself is the main contribution

Style rules

• Explain the insight. Don't directly copy text from the paper or combining multiple sentences, understand and explain the insight of the paper.
• Be concise, not terse. A reader should be able to understand the mechanism, not just memorize its name. But cut any sentence that only restates what a well-named component already implies.
• Use the paper's actual terminology. If the paper calls it "WCA (Weighted Correction Attention)", do not paraphrase it as "selective token correction." Open the paper and check.
• Avoid marketing language. No "novel", "cutting-edge", "state-of-the-art" unless quoting a benchmark result. No "seamlessly", "elegantly", "robustly".
• Prefer verbs over nouns. "Ranks tokens by deviation and blends the top-k" > "performs a ranking-based blending mechanism".
• Equations inline, not blocks. Use $...$ for short formulas. Only use $$...$$ if the formula is longer than a line and central to the method.
• No emojis. Don't add them unless the surrounding file already uses them.

Length guidance

• Background + Key problem: 2-4 bullets total
• Proposed method: 1 bullet per component, each with an idea line + ≤3 mechanism sub-bullets
• Results: 1 bullet
• Total target: 8-15 lines for a typical systems/ML paper. Go longer only if the method has >3 components or equations are essential.

Process

1. Fetch the paper. Use WebFetch on the arXiv HTML or the PDF. If WebFetch returns binary PDF metadata, try WebFetch on the HTML mirror (arxiv.org/html/<id>) or download via curl and use Read with pages:
2. Find the real component names. Skim the method section and the algorithm pseudocode. Do not invent names.
3. Identify the baseline being beaten. This tells you what goes in the Background block.
4. Draft all four blocks. Then re-read and cut anything that's restating the obvious.
5. Check against the file's existing style. If the target markdown file uses a different bullet convention, match it.

Anti-patterns

• Listing every contribution in the abstract. Papers oversell. Pick the 2-3 that actually matter.
• Copy-pasting the abstract's framing. Abstracts are written for acceptance; summaries are written for recall. Re-frame.
• "The authors propose..." Drop it. Start with the technique name.
• Deep-diving one component while ignoring the others. Balance coverage across the components — readers need the whole shape of the method.
• Claiming results without numbers. "Outperforms baselines" is useless. "3.1-4.3× lower latency vs. vLLM" is useful.

Example (reference)

- CacheSlide: Unlocking Cross Position-Aware KV Cache Reuse for Accelerating LLM Serving [[FAST'26](https://www.usenix.org/system/files/fast26-liu-yang.pdf)]
  - Background:
    - PDC (Position-Dependent Caching): KV tied to absolute positions, reuse only on exact prefix matches
    - PIC (Position-Independent Caching): strips position encoding, reusable anywhere but loses attention fidelity
    - RoPE: high positional sensitivity — any shift invalidates cached keys
    - CoPE: content-gated position encoding, less sensitive to shifts
  - Key problem & insight: agent prompts have reusable segments that maintain consistent *relative* ordering despite absolute shifts (RPDC pattern). PDC/PIC don't exploit this; CoPE can, if positions are locked to a learned template rather than recomputed from live context
  - Proposed method — CacheSlide with three components:
    1. **CCPE (Chunked Contextual Position Encoding)**: pretrain a template $e^*$ of the most frequent CoPE encoding per task; at inference, reuse chunks get positions from $e^*[i]$ (pinned), recompute chunks get live CoPE
    2. **WCA (Weighted Correction Attention)**: token-level gate on top of CCPE — rank tokens in a reuse chunk by $d_i = \|K^{\text{new}}_i - K^{\text{cache}}_i\|$, top-k (~5-17%) get blended $K_i \leftarrow \alpha K^{\text{new}}_i + (1-\alpha) K^{\text{cache}}_i$, rest use cache as-is, applied every $\tau$ layers
    3. **SLIDE (KV cache manager)**: make WCA's I/O pattern SSD-friendly — relocate updated tokens to fresh pages (sequential writes), spill clean pages first, reclaim scratch pages during decode
  - Results: 3.1-4.3× latency reduction, 3.5-5.8× throughput improvement over state-of-the-art baselines

This example shows all four blocks at the target length — use it as a template, not a script.