Working with research papers is not just about reading them. It also involves finding relevant work, evaluating its credibility, and deciding whether its ideas are worth your time. Because the volume of published research is overwhelming, you need a practical workflow for filtering papers quickly without overlooking valuable contributions. This guide describes the approach I use to discover, read, assess, and compare research papers.

How to find papers

Top conferences. Browse papers accepted at leading A* and A ranked conferences, such as NeurIPS, ACL, EMNLP, ICML, and ICLR. The program committee has already done much of the hard work by reviewing submissions and filtering out many weak or unconvincing papers, which can save you considerable time. However, peer review is not perfect, and conference cycles introduce a delay: you may not discover a paper until months after the initial work was completed. Popular topics can still produce hundreds of potentially relevant papers.

Frontier research labs. Monitor organizations such as OpenAI, Google DeepMind, Anthropic, Meta, DeepSeek and NVIDIA. Watch their blogs, press releases, repositories, model releases, and technical reports.

Researchers you trust. Follow strong researchers and practitioners working in your areas of interest. For example Andrej Karpathy or Tri Dao.

Telegram channels and research communities. These sources can surface new work very quickly, but their signal to noise ratio is often low. Treat them as discovery tools, not as quality filters.

Hugging Face Papers. Browse its daily, weekly, and monthly paper collections to discover recent and popular work.

Search arXiv directly for papers related to your topic. It provides early access to research, but papers may not have been peer reviewed, so you will need to evaluate their quality yourself.

How to read papers

Keep in mind that academic papers are often written under publication pressure. Authors are incentivized to emphasize novelty and positive results, so they will naturally present their method in the best possible light—even if it failed under certain conditions. I therefore treat reported performance as an upper bound on what I am likely to reproduce in practice.

There is rarely enough time to read every paper from beginning to end. Start by identifying the problem, then jump directly to the method and results. Read the remaining sections only when the paper appears relevant, valuable, or worth implementing:

• Identify the target problem – Start with the abstract and, if useful, ask an LLM to summarize the problem. If it is relevant to you, read the introduction to understand the motivation and verify that the paper actually addresses your needs.
• Inspect the results – Determine whether the improvement is meaningful and whether the comparisons are fair. If the method moves the Pareto frontier—improving quality without disproportionate increases in cost or complexity—it is probably worth investigating further.
• Understand the algorithm – Identify the central idea, its assumptions, and why it is supposed to work. Make sure you can explain it in your own words.
• Examine the implementation – Estimate how difficult the method would be to reproduce, integrate, and run efficiently. Check whether the practical cost is justified by the reported gains.
• Read everything else selectively – A deeper read can reveal limitations, assumptions, and implementation details that are easy to miss. This is especially useful when you plan to invest significant time in the method or use it in a real system.

A few tools can make paper discovery and analysis much faster. They can help you understand the surrounding literature, identify important references, and verify your interpretation of a paper:

• Connected Papers – Creates a visual graph of related research, helping you discover influential papers, prior work, and less obvious connections within a research area.
• Semantic Scholar – Makes papers easier to explore by surfacing summaries, influential citations, related work, and other important information.
• Large language models – Use an LLM to discuss the paper, validate your understanding, explain unfamiliar concepts, and answer questions about the method. Ask it to support its answers with specific passages or page numbers, since it can misinterpret details or invent explanations.

How to identify suspicious papers

A weak or misleading paper does not always look obviously bad. Warning signs often appear in what the authors chose not to evaluate, explain, release, or compare:

• Community scrutiny – Search for the paper on OpenReview. Reviews and comments may reveal questionable assumptions, methodological problems, or missing evidence that you did not notice yourself.
• Shallow evaluation – A limited or poorly designed evaluation may indicate that the authors are hiding weaknesses or have not tested the method rigorously enough.
• Narrow benchmark coverage – Testing on very few datasets or a narrow data distribution makes it unclear whether the method generalizes to other domains. For example, the evaluation may cover only short- or long-context tasks, or only reasoning or non-reasoning models.
• Outdated models – Results obtained only with old models may no longer reflect how the method performs with current architectures and training techniques.
• Small models only – A method that works on small models may behave differently at a larger scale. Without large-model experiments, its scalability remains uncertain.
• No public implementation – The absence of code or a repository makes the results harder to reproduce and verify. This is less concerning when the paper is very recent and the implementation may still be in progress.
• Excessive implementation complexity – If reproducing the method requires an unreasonable amount of engineering, its practical value may be limited. The authors may not have sufficiently considered the algorithm’s efficiency or deployment costs.

How to spot gem

Finding a genuinely valuable paper is rarely about one impressive result. The strongest work usually combines a clear idea, rigorous evidence, credible authorship, and signs that the broader community finds it useful:

• Strong open-source adoption – The method has many GitHub stars or has been integrated into popular frameworks, suggesting broad interest and practical value. For very recent papers, however, implementations and community adoption may still be limited.
• Credible authorship – The authors have a strong publication record, a high h-index, or come from a respected research lab.
• Clear, thorough evaluation – The paper tests the method across multiple benchmarks, model families, and model sizes, with transparent and well-explained results.
• A simple, understandable algorithm – You can fully grasp how the method works, without important details being hidden, omitted, or left unanswered.
• Few fragile assumptions – The idea is grounded rather than speculative and relies on a small number of assumptions, making it more likely to generalize.
• It passes the bullshit test – While reading, you do not feel that the authors are misleading you, cherry-picking results, or using complexity to hide weaknesses.

How to rank papers

Ranking papers becomes difficult when they evaluate different models, datasets, or benchmarks. Instead of searching for one universal score, compare them through shared baselines, performance trade offs, and the consistency of their improvements:

• Transitive comparison – Suppose you need to compare papers A, B, and C, but they do not report results on the same benchmark. If A outperforms B under one comparable setup and B outperforms C under another, you may cautiously infer that A is likely stronger than C. This works best when the shared baselines, metrics, and experimental conditions are reasonably similar.
• Reported results are context dependent – Every result applies to a particular benchmark, model, parameter configuration, and evaluation protocol. A method that wins in one setup may not remain the best under different conditions.
• Look for the Pareto frontier – Prefer methods that offer the best trade off between important dimensions such as quality, speed, memory usage, and implementation complexity. The strongest methods are meaningfully better across several dimensions without making another dimension substantially worse.
• Account for noise – Small differences may come from random variation, hyperparameter tuning, or evaluation choices. Treat an improvement as meaningful only when it is clearly larger than the expected experimental noise.