HACKER NEWS → PAPERS → IDEAS
Hacker Paper
Follow a paper the community keeps coming back to. Find the question behind it.
Read the guide and its limits, return to the cited pages, then connect the findings with talks and theme readings.
9papers9guides4connected readings
How are these papers selected?
Following HackerTube’s recurring-mention threshold: at least 3 distinct HN accounts across 3 UTC months, in linked comments with at least 100 characters of their own prose. Mentions may recommend, cite, or criticize a paper; these signals are counted separately. They are not a scientific quality score.
HN evidence in this edition is frozen at 2026-11-12. Guides are written from the full papers; read the originals at the linked scholarly sources.
9 papers
2024 · ICLR 2024Guide available
Grégoire Delétang, Anian Ruoss, Paul-Ambroise Duquenne et al.
The paper connects autoregressive prediction with lossless coding: probabilities assigned to the next symbol can be converted into code lengths through arithmetic coding. The authors compare language models with classical compressors on preprocessed text, image, and audio samples, distinguishing rates for the encoded data alone from rates that also include model parameters. This helps examine training objectives and evaluation conventions; it does not establish that good compression makes a reliable assistant.
23 HN accounts mentioned it1 explicitly recommended itLast on HN 2026-11-02
2024 · arXiv preprintGuide available
Shibo Hao, Sainbayar Sukhbaatar, DiJia Su et al.
Coconut lets a model perform intermediate reasoning without first translating each state into words: it feeds the last hidden state directly into the next input embedding and later resumes text output. Training begins with demonstrations containing textual steps and progressively replaces some steps with continuous states. Experiments on math questions and synthetic logic tasks show particular trade-offs between accuracy and reasoning cost. The method provides another computational representation; it does not establish that hidden states amount to human thought.
16 HN accounts mentioned it1 explicitly recommended itLast on HN 2026-09-06
2024 · arXiv preprintGuide available
Jacob Pfau, William Merrill, Samuel R. Bowman
When a model writes more intermediate text, do gains come from the expressed steps or from additional opportunities to compute? The authors train small transformers on two synthetic summation tasks, inserting meaningless symbols such as repeated dots between a question and its answer. Under particular training conditions, these symbols also improve answers, showing that useful computation can occur in intermediate representations whose visible tokens reveal no reasoning. Training is crucial: demonstrations with a parallelizable decomposition help, whereas filler-only training or a different, sequential decomposition does not deliver the same benefit.
15 HN accounts mentioned it2 explicitly recommended itLast on HN 2026-11-11
2025 · arXiv preprintGuide available
Cameron R. Jones, Benjamin K. Bergen
How is appearing human measured? Judges simultaneously chat by text with a human and an AI for up to five minutes, then choose which is human. The study evaluates four types of system in undergraduate and Prolific populations, comparing prompts with and without an assigned persona. Persona-prompted GPT-4.5 is selected as human in 73% of pooled games, compared with 36% without that persona. The useful question is what this demonstrates about short conversational humanlikeness, and how far it remains from general intelligence or long-term reliability.
14 HN accounts mentioned it0 explicitly recommended itLast on HN 2026-10-03
2024 · NeurIPS 2024Guide available
Andy Arditi, Oscar Obeso, Aaquib Syed et al.
The paper moves refusal from a question about response style to experiments on internal representations. Across 13 open chat models with 1.8B–72B parameters, it identifies a model-specific one-dimensional direction causally involved in refusal. This provides a concrete object for studying how post-training shapes behavior. Its results require separating refusal wording, harmfulness of completions, and general-task performance; they do not establish one semantically clear safety switch shared by all models.
12 HN accounts mentioned it1 explicitly recommended itLast on HN 2026-11-03
2025 · arXiv preprintGuide available
Om Dobariya, Akhil Kumar
Does a multiple-choice question get a different answer when its tone changes? The authors use Deep Research to generate 50 base questions and answer keys, rewrite each into five tone variants, and primarily evaluate ChatGPT-4o. Ruder variants score higher on this set. This is a small wording experiment: the rewrites also change words and length, so it does not isolate an emotional response by the model or establish abusive prompting as a general recommendation.
9 HN accounts mentioned it1 explicitly recommended itLast on HN 2026-10-11
2025 · arXiv preprintGuide available
Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan et al.
When an agent reads email, web pages or files, external text can redirect its next action. CaMeL separates planning from a trusted user request from parsing untrusted content. An interpreter then tracks each value’s provenance and permitted readers and checks security policies before tool calls. This places part of the protection in the execution system around the model, rather than only asking the model to recognize malicious instructions. AgentDojo experiments show the value of this design alongside costs in task completion, policy maintenance, user confirmation and side channels.
8 HN accounts mentioned it3 explicitly recommended itLast on HN 2026-10-01
2024 · arXiv preprintGuide available
Matthias Gerstgrasser, Rylan Schaeffer, Apratim Dey et al.
Must training each generation on the previous model’s output inevitably make models worse? This paper separates replacing old data with newly generated data from retaining the initial dataset and accumulating outputs across generations. It compares these processes in language, molecular-conformation, and face-generation experiments, then analyzes error in a particular linear-regression framework. Accumulation substantially changes degradation, but avoiding collapse does not mean that every quality measure remains unchanged forever.
6 HN accounts mentioned it0 explicitly recommended itLast on HN 2026-09-13
2025 · arXiv preprintGuide available
Subhash Kantamneni, Max Tegmark
When a model writes 3+63=66, what happens internally? This paper studies addition in GPT-J, Pythia-6.9B, and Llama-3.1-8B. It fits integer representations with a generalized helix containing linear and periodic components, then uses activation patching, ablations, and component analysis to test their computational role. This is stronger evidence than an attractive projection plot, but the precise computation turning the two input helices into the answer helix remains unidentified.
4 HN accounts mentioned it0 explicitly recommended itLast on HN 2026-10-02
No matching papers. Try another title, author, or keyword.
Hacker