Single-paragraph summaries optimized for social sharing. Insight over information.
Structure
- Context: What's the problem?
- Insight: What did they realize that others missed?
- Solution: How does insight → method? (should feel natural)
- Evidence: Concrete comparison showing it works
Then: Implication line + 📎 arxiv link
Key Rules
- Explain like reader is smart but unfamiliar with the domain
- Use concrete examples/analogies (e.g., "trash can role" >> "focus on a specific token")
- Show cause-and-effect chains explicitly
- Compare/contrast with alternatives ("X failed while Y succeeded")
- Bold 2-4 key concepts
- Match user's language (Korean/English)
Example
Input: arXiv 2601.15380
Output:
Transformer's attention determines "which tokens to look at and how much," and this paper reinterprets softmax attention as the solution to an optimization problem called **Entropic Optimal Transport (EOT)**. The insight this perspective provides is: hidden within attention computation is an implicit uniform prior that "all positions are equally important." Why is this a problem? In LLMs, there is the attention sink phenomenon where the first token receives enormous attention regardless of meaning. Since softmax must output probabilities that sum to 1, when a query has no token it should properly attend to, it needs somewhere to "dump" attention; under a uniform prior, implementing this requires the first token's key vector to carry structural information like "I am a trash can"—the representational capacity of the key, which should originally express only semantic content, is wasted. Since the EOT interpretation revealed this problem, the solution is also natural: change the prior from uniform to learnable. The GOAT proposed in this paper separates "the default importance of each position" into a separate learnable term, so that the key vector handles purely meaning and the prior handles positional information. In experiments, while existing methods failed sharply when exceeding the training length, GOAT maintained information retrieval performance even in long contexts.
Implication: The EOT perspective reveals the hidden assumptions of attention and opens up the design freedom to change those assumptions—attention sink is a byproduct of the uniform prior, and it is solved by explicitly modeling the prior.
📎 https://arxiv.org/abs/2601.15380
Avoid
- Jargon without intuition
- Findings without comparison to alternatives
- Method description without motivation (only "this is how it was done" without "why it was done this way")
Multiple Papers
When summarizing multiple papers:
- Lead with the unifying theme/problem
- Contrast what each paper realized differently
- Synthesize implications across papers
Language
Match the user's language (Korean/English). Maintain the same insight-first structure regardless of language.