LLM Mastery
49. Research Taste: How to Read Papers
Most papers are optimistic about their own contribution, not dishonest. Reading well is a forensic skill, not a deferential one.
Level: Master · Article 49/50 · Karpathy-style LLM course
Read papers in an order that front-loads the evidence, and resist the urge to start at the abstract and read linearly — the abstract is marketing copy written last, optimized to sell the result, and it’s the least reliable part of the document.
Reading order
- What problem, why now? Is this solving something real, or reframing an old problem with new vocabulary?
- The method figure. Can you redraw it from memory after looking once? If not, you don’t understand the method yet, and no amount of prose will fix that.
- The main result table. What’s actually being compared, on what data, against what baselines? Are the baselines current, or convenient?
- The ablations. This is where the paper either earns its claims or doesn’t. More on this below.
- Limitations. Often the most honest section, frequently the shortest, sometimes buried in an appendix precisely because it’s honest.
- Only now, the prose claims and abstract — read them as a hypothesis to check against everything above, not as the conclusion.
This order is backwards from how papers are written, and that’s the point. A paper is composed to persuade: motivation and framing up front, caveats at the end, if at all. Reading it in composition order means you absorb the persuasion before you’ve seen the evidence, and every subsequent number you read gets fit into a narrative you already accepted. Reading figures and tables first, before the framing has had a chance to anchor you, is the single highest-leverage habit change in this article.
Worked example
Say a paper claims: “Our novel gating mechanism is responsible for the improvement.” The ablation table shows:
| Variant | Benchmark score |
|---|---|
| Full method | 74.2 |
| − gating mechanism | 73.9 |
| − data augmentation | 68.1 |
| − larger batch size | 71.5 |
Read this like an engineer reviewing a benchmark, not a fan reading a highlight reel. Removing the “novel” gating mechanism costs 0.3 points. Removing data augmentation costs 6.1 points. If the paper doesn’t report variance across seeds, ask yourself: is 0.3 points distinguishable from noise on this benchmark? For most benchmarks with a few thousand eval examples, a swing of 0.3 points from a different random seed alone is entirely plausible. The evidence in this table says the data augmentation is doing the real work, and the gating mechanism, the thing in the title, might be doing almost nothing. That’s not an accusation of dishonesty; it’s just what happens when a paper’s narrative is written before all the ablations are, and nobody goes back to rewrite the title once an ablation undercuts it.
Failure mode
Two specific ways research taste gets fooled, beyond the ablation-reading issue above:
- Benchmark chasing. A method that specifically improves the exact benchmarks reported, via choices tuned, consciously or not, against those benchmarks during development, will look great in the paper’s table and disappoint the moment you apply it to your own, differently shaped problem. The tell is a paper that reports many benchmarks but no held-out or out-of-distribution evaluation at all.
- Mistaking scale-correlation for the claimed mechanism. If a paper’s proposed technique is only tested at one scale, and “bigger models do better” is a near-universal finding regardless of technique, you cannot distinguish “this technique caused the improvement” from “this technique happened to be tested on the bigger of two models, and bigger models are just better.” Demand, or mentally flag the absence of, a controlled comparison at matched scale.
Exercise
Find one paper, from your own field or the LLM literature, with a public ablation table. Identify the single ablation row that most directly supports the paper’s headline claim. Compute, or estimate if variance isn’t reported, whether the score gap in that row is large relative to typical seed-to-seed variance on that benchmark. Write a half-page verdict: does the ablation table actually earn the title’s claim, or is a different row in the same table doing the real work?
← 48. Build Eval-Driven: A Practical Workflow
50. Capstone: Train Your Own Tiny LLM →