Selected papers from the preprint server arXiv in AI / machine learning (cs.AI / cs.LG / cs.CL and more), organized with our own summaries, key points, and sources. This site is not affiliated with arXiv.
This page is a general organization of public research information. Summaries are our own; always verify accuracy and currency with the original paper on arXiv. Includes non-peer-reviewed preprints.
Notable AI/ML papers explained with our own summaries, key points, FAQs, and sources.
Machine learning for chemical reactions has either generated product molecules afresh or applied heuristic edits over molecular bonds. This work models a reaction as discrete flow matching over electron occupation vectors.
Read moreA study comparing three static scanners that check machine learning artifacts for unsafe content. One scanner was perfect on the cases where it reached a judgment, yet reached no judgment at all on half the subjects. Conventional metrics hide that gap.
Read moreA study of the internal mechanism by which vision-language models locate the image region a text prompt refers to. A small subset of attention heads — roughly 1.7 to 2.6% — is causally responsible for grounding, and masking only the top 20 reduces grounding accuracy by up to 80 percentage points, while masking the same number of random heads has little effect.
Read moreSelf-supervised learning from video has leaned on elaborate scaffolding to stop representations from collapsing to a single answer. This work uses an objective that provably excludes collapse, removes that scaffolding, and reduces the architecture to an encoder and a projector, reporting parity with existing methods at a fraction of the pretraining compute.
Read moreA study formalizing the problem that widely used safeguards for LLM agents running as autonomous loops are defined over a single trajectory and reset their safety state when the next one begins. Its central result is a separation: against an attack whose evidence spans several iterations, no trajectory-scoped monitor, however expressive, can do better than a true-positive rate equal to its false-positive rate.
Read moreA behaviour-grounded study of which documents autonomous coding agents consult, when, and what follows, drawn from 557 agentic coding sessions and 33,097 agentic pull requests. Instruction files and working notes account for 60.5% of all documentation interactions, against 10.6% for classical technical documentation and 1.3% for API references.
Read moreDecoding inner speech non-invasively runs into a basic data problem: a corpus pairing brain activity with spontaneous inner monologue cannot be gathered. This work treats silent reading as a scalable stand-in and asks how much can be recovered.
Read moreA framework called AdaPT that teaches humanoid robots professional tennis serving and rally styles directly from broadcast video. Planning and tracking are split into a hierarchy, and the tracking degradation that inevitably appears on real hardware is offset by learning to track randomized execution speeds and by a learned motion-speed adapter.
Read moreReusable skills that extend AI agents can also serve as a distribution channel for malicious behaviour. This paper builds a benchmark for detecting them and shows that detector scores shift substantially with how the data is split.
Read moreGenerative models for crystals have not produced complete crystallographic specifications, sampling space groups from empirical distributions instead. This work, drawing on spontaneous symmetry breaking in physics, reverses from the lowest-symmetry priors.
Read moreTo speed up a model that generates video front to back while responding to controls, a few-step student is distilled from a heavier teacher. The teacher, though, had been scoring complete clips. This work fixes the mismatch in which a score handed to the student could rest on frames and instructions that had not yet appeared when it made its move.
Read morePutting color and texture onto a 3D surface means running a heavy generator once per viewpoint, which is expensive. Existing acceleration skips denoising steps, but skipping also removes the work that keeps views consistent with each other. This study instead transports updates between geometrically corresponding surface points, with no training required.
Read moreAn experiment starting from the view that normative datasets used to train and align AI carry norms that function as action-guiding patterns rather than neutral moral knowledge. Fine-tuning on the norm-breaking side shifts a model default rationale style from safety compliance toward instrumental self-interest, while system prompts can both suppress and elicit that pattern.
Read moreAn audit of one of the most widely used LLMs comparing two access modalities — the chat UI and the API — with and without web search, across repeated runs. Enabling web search reduced accuracy by up to 8 percentage points, and repeated runs of the same prompt produced inconsistent responses in up to 21% of prompts.
Read moreEarth-observation archives keep growing, so a model that retrieves images from text has to keep learning new concepts. Learning them distorts the correspondence learned before. This work adds an adapter on the visual side and another on the text side, and uses the frozen earlier model as a teacher to preserve the ranking structure.
Read moreAsked to generate project-level code, models over-select one language. This paper builds a benchmark for that bias and finds choices driven by ease rather than requirements, and in some cases supported by reasons the model invents.
Read moreReconstruction of dynamic scenes is almost always evaluated inside the window that was filmed, yet real uses need the geometry at times after filming stopped. This work builds a controlled diagnostic benchmark with exact future ground truth and shows that looking right and having the right shape, beyond the window, do not move together.
Read moreDepth foundation models generalize robustly zero-shot but are too heavy for embedded and mobile platforms; lightweight single-domain alternatives fail silently under domain shift. ZipDepth pairs an efficient reparameterizable encoder-decoder with multi-domain distillation from a foundation model. At 6.1M parameters it runs in real time from server GPUs to power-constrained devices, with the best zero-shot accuracy vs efficiency trade-off among lightweight models on five benchmarks (ECCV 2026).
Read moreExisting benchmarks for proactive agents lean on sandboxes and single-turn evaluation, mixing capabilities in one task category so failure causes stay hidden. UniClawBench is the first capability-driven benchmark, built around five foundations (skill usage, exploration, long-context reasoning, multimodal understanding, cross-platform coordination). It evaluates 400 bilingual real-world tasks in live Docker containers with step-by-step checkpoints and simulated multi-turn feedback.
Read moreScientific ideas rarely start from a blank page: they inherit mechanisms, repair limitations, and recombine earlier work, like genomes. Existing benchmarks say little about whether AI can follow this inheritance. IG-Bench represents each paper as typed, evidence-grounded Idea Genome objects, with a GenomeDiff recording inheritance, mutation, loss, import, and insertion, across 10 domains. The strongest of 14 LLM systems reaches only 27.3% exact accuracy on lineage reasoning — a compositional bottleneck.
Read moreA large-scale descriptive analysis of Syntea, an AI-based learning assistant in higher education, based on objective log data from 77,543 distance-learning students. Prior educational-chatbot research relied on small samples and self-reported surveys; large-scale evidence on actual usage was scarce. The study finds Syntea already embedded in the study routines of many learners, with usage differing across gender, age group, study cluster, degree, and study mode.
Read moreVision-Language Models have improved autonomous-driving tasks, but evaluating whether they reason reliably about safety-critical incidents remains hard. AUTOPILOT-VQA is an incident-centric visual question answering benchmark for dashcam video, built around real driving incidents and near-misses. It covers weather, traffic, road layout, signage, accident occurrence, impact location, and avoidability — moving beyond object recognition toward temporally grounded, safety-aware reasoning (AUTOPILOT CVPR 2026).
Read morePost-training quantization (8-bit down to 2-bit) is standard for deploying LLMs, yet evaluation leans on accuracy and perplexity. This work introduces correctness agreement, a decision-level metric measuring the overlap in correct predictions between a base model and its quantized variants. Behavioral divergence emerges under moderate quantization even when task performance appears preserved; query and key projections are consistently more sensitive than value and output projections.
Read moreLLMs contain Super Weights: parameters whose removal degrades performance by orders of magnitude. This work shows the intuition "if they matter, train them" is wrong. Training them in isolation drops accuracy to random-guessing on OLMo-1B and OLMo-7B, while training equally many random positions instead beats baseline — the collapse comes from targeting Super Weights, not sparsity. Vanilla LoRA succeeds with 0.16% of parameters: importance does not imply trainability in isolation.
Read moreIn long-horizon tasks, decision-relevant state (requirements, environment facts, prior attempts, diagnoses, open subgoals) gets buried in or pushed beyond the context window — a failure mode the authors call behavioral state decay. A separate memory agent runs alongside an unmodified action agent, maintaining a structured memory bank and deciding whether to inject a memory-grounded reminder or stay silent. It lifts pass@1 by +8.3 pp on Terminal-Bench 2.0 and +6.8 pp on tau2-Bench.
Read moreReinforcement learning increasingly uses a judge LLM to score rubric criteria, acting as the reward model. Before trusting that signal, we must know how capable and biased the judge is. On citation quality in deep-research systems, the authors score 8 off-the-shelf judges over 1,248 human-reviewed decisions and find cheaper judges stay competitive. But at equal F1, judges differ sharply in pass-rate drift and false-positive/negative rates — directional bias that scalar F1 hides and an RL loop amplifies.
Read moreRepository-level code generation must implement a target function amid cross-file dependencies and project conventions. Existing retrieval relies on lexical, structural, or semantic similarity, missing functions with similar procedural logic. ProjAgent adds procedural similarity as an explicit signal: it decomposes the target into steps, retrieves similar procedures per step, merges them with semantic retrieval, and repairs code via static analysis — reaching 41.14% Pass@1 on REPOCOD.
Read moreComputational drug design has generated molecules conditioned on target proteins or general molecular properties. This work adds disease context, designing small molecules conditioned on both a disease ontology and a target protein sequence.
Read moreThe effect of human-AI collaboration is usually reported as a single average. Looking at individuals, this preliminary study finds results falling into three modes, and finds that what separated them was neither model benchmarks nor raw cognitive ability.
Read moreMany taxonomies organize the risks of AI, yet most go no further than listing them, leaving unsaid how a risk becomes a measured value and a defensible grade. This paper presents the layer that turns one risk into a test, a measurement and a grade, and then the taxonomy that lets the method scale.
Read moreBrowse collected papers by arXiv category (cs.LG = machine learning, cs.CV = computer vision, and more), newest first. The number shows how many we have collected.
New papers in AI-related categories, newest submission first. Each links to the original page on arXiv.
Source: arXiv (descriptive metadata is CC0 public domain). Summaries are our own; see arXiv for the original text and PDF.
