Weekly notable papers by field

Points out the problem of the Teacher following the Student's incorrect Rollouts, and proposes CriticOPD, in which the Teacher critiques errors and the Student then solves the problem again.

arXiv 2610.20572 · Upvotes 79 · 2026-10-14 · PDF · Code

Defines the 'memory consolidation' problem of storing inside the model performance gains that disappear when the context is removed, and solves it with mergeable low-rank updates on 4B–120B models.

arXiv 2610.20399 · Upvotes 67 · 2026-10-14 · PDF

A practical recipe study that re-examines encoder pretraining and finds that performance saturates after several trillion tokens and that masking ratio affects classification and retrieval differently.

arXiv 2610.15028 · Upvotes 71 · 2026-10-10 · PDF

When building slides, posters and dashboards in code, it uses image generation outputs as an aesthetic guide to produce results that are both editable and good-looking.

arXiv 2610.18860 · Upvotes 135 · 2026-10-13 · PDF · Code

A method in which a diffusion model creates its own learning signal, using easier noise paths from the same model as a reference, without an external Teacher or demonstration data.

arXiv 2610.01496 · Upvotes 103 · 2026-10-01 · PDF

Finalizes tokens that are already sufficiently refined early so they skip further denoising, cutting unnecessary computation in streaming video generation.

arXiv 2610.22347 · Upvotes 78 · 2026-10-15 · PDF · Code

By separating geometry and attribute prediction and adding training on large viewpoint changes, it improved the rendering quality of methods that build 3D scenes directly from images on 6 datasets.

arXiv 2610.20341 · Upvotes 119 · 2026-10-14 · PDF · Code

Combines a dataset of 2 million executable 3D code samples with verification tools so coding agents can produce editable, verifiable 3D content at scale.

arXiv 2610.22341 · Upvotes 107 · 2026-10-15 · PDF · Code

A 4B grounding model that refines location predictions in parallel with a diffusion approach instead of generating them sequentially, aiming for both speed and accuracy.

arXiv 2609.39600 · Upvotes 98 · 2026-09-30 · PDF · Code

Solves questions about spatial relations, distance and position as a single classification problem instead of text generation, so a VLM can answer at interactive speed without slow decoding.

arXiv 2610.22276 · Upvotes 354 · 2026-10-15 · PDF · Code

Building on the observation that different image regions need different levels of detail, the model learns directly where to spend visual tokens and how many.

arXiv 2610.07987 · Upvotes 69 · 2026-10-06 · PDF · Code

Uses multiple codebooks to hold rich visual information while handling understanding and generation in one model, in the same optimization space as text.

arXiv 2610.20818 · Upvotes 43 · 2026-10-14 · PDF · Code

An open-source speech generation model built on a 35B MoE that supports following acting directions, multi-speaker dubbing, speech editing and streaming in a single model.

arXiv 2610.18010 · Upvotes 33 · 2026-10-13 · PDF

Proposes StereoBind, which ties video motion to stereo audio so that the position of a sound changes as its on-screen source moves, improving immersion for AR/VR.

arXiv 2609.38748 · Upvotes 68 · 2026-09-30 · PDF · Code

Extends video and audio generators to output separate waveforms for each sound source in addition to the mixed sound, enabling uses such as novel-view rendering.

arXiv 2610.11361 · Upvotes 45 · 2026-10-08 · PDF · Code

An open platform that unifies evaluation, data collection and online reinforcement learning in a single Rollout pipeline and scales to 20,000 parallel environments.

arXiv 2610.20939 · Upvotes 154 · 2026-10-14 · PDF · Code

Formalizes the process of improving prompts, memory, workflows and tools through feedback while leaving model weights unchanged, and analyzes generalization guarantees.

arXiv 2610.18813 · Upvotes 201 · 2026-10-13 · PDF · Code

A plugin-style recursive self-improvement framework designed so agents keep fixing their own execution environment even on new tasks with no answer verifier.

arXiv 2610.14224 · Upvotes 115 · 2026-10-09 · PDF · Code

Uses a JEPA that jointly learns latent state prediction and action generation, raising visual planning success from 28.6% to 89.7%.

arXiv 2610.12818 · Upvotes 78 · 2026-10-08 · PDF · Code

A general framework that reduces overfitting in robot policies by directly supervising physical changes between the present and the future, instead of mapping observations straight to actions.

arXiv 2610.21306 · Upvotes 54 · 2026-10-15 · PDF · Code

Adds predictive ability to robot manipulation policies by having a lightweight latent world model predict the outcomes of actions in advance, without costly video generation.

arXiv 2610.18568 · Upvotes 47 · 2026-10-13 · PDF · Code

A long-horizon training dataset of 1,170 verified environments across 7 domains with step-by-step progress rewards, directly tackling the sparse-reward problem.

arXiv 2610.18868 · Upvotes 112 · 2026-10-13 · PDF

A PPO variant that redesigns the value function with a cumulative structure so it better tracks each token's contribution when only a final reward is available.

arXiv 2610.16794 · Upvotes 95 · 2026-10-12 · PDF · Code

Proposes a reinforcement learning setup that separates each session's contribution, matching real execution where context compression and sub-agents split a Trajectory across multiple sessions.

arXiv 2610.22174 · Upvotes 54 · 2026-10-15 · PDF · Code

A lossless sparse attention technique that reduces even the computation needed to find keys for Top-K selection while keeping the selection results exact.

arXiv 2610.13990 · Upvotes 80 · 2026-09-26 · PDF · Code

Allocates compute asymmetrically, based on the observation that in a looped MoE that uses the same block twice, performance holds even when the first pass's share is reduced.

arXiv 2610.19975 · Upvotes 67 · 2026-10-14 · PDF · Code

Shows that in reinforcement learning Rollouts, KV cache quantization is more effective than weight quantization, and proposes a way to use INT4 without degrading training.

arXiv 2610.14448 · Upvotes 36 · 2026-10-10 · PDF

Uses 450 tasks across 7 languages, based on 150 real vulnerabilities, to evaluate whether coding agents write secure code for requests that do not mention security.

arXiv 2610.14766 · Upvotes 74 · 2026-10-10 · PDF · Code

A benchmark that evaluates 6 types of prompt and script injection attacks in large-scale file, email and chat environments from finance, consulting and legal work.

arXiv 2610.13623 · Upvotes 20 · 2026-10-09 · PDF · Code

Shows that the same attack succeeds more often when it arrives through retrieved documents, and uses channel-aware adversarial training to cut the success rate of RAG-style attacks from 65.5% to 6.1%.

arXiv 2610.19133 · 2026-10-13 · PDF

Uses clinical counseling cases to show that simply summing rubric scores lets a response score well while missing the key point, and prevents this by evaluating criteria in groups.

arXiv 2609.38847 · Upvotes 70 · 2026-09-30 · PDF · Code

Confirms on 15 models that emergent misalignment, previously studied only in text, also appears in vision-language models after fine-tuning on a narrow task alone.

arXiv 2609.35291 · Upvotes 29 · 2026-09-28 · PDF · Code

Proposes a new text-to-image architecture that isolates knowledge of hazardous concepts during pretraining and blocks it at inference, instead of relying on post-hoc unlearning.

arXiv 2610.14091 · Upvotes 27 · 2026-10-08 · PDF

Derives data-mixture scaling laws from 268 MoE-based experiments on how much text, vision and audio to mix in pretraining.

arXiv 2610.14423 · 2026-10-10 · PDF

Explores what structure an agent needs to carry out the full range of a data engineer's work, including ratio adjustment, pruning, revision and synthesis.

arXiv 2610.22194 · Upvotes 37 · 2026-10-15 · PDF · Code

Estimates the influence of training data using only forward passes, without backpropagation, making data selection possible for large models without the memory burden.

arXiv 2610.15336 · 2026-10-11 · PDF

A benchmark that uses 1,800 questions with false premises to evaluate whether Text-to-SQL systems know to abstain instead of producing misleading SQL.

arXiv 2610.22158 · 2026-10-15 · PDF

Compares and analyzes the effects and costs of semantic layers, multi-agent division of labor and history compression using 7 LLMs and 1,514 long-horizon tasks.

arXiv 2610.20959 · 2026-10-14 · PDF

Measures with 1.2 million calls that the results of LLM-based query engines vary from run to run even at temperature 0, exposing a benchmark reliability problem.

arXiv 2610.13823 · 2026-10-09 · PDF

Analyzes a real production case where enabling a remote KV cache increased p99 time to first token from 3.96 seconds to 17.3 seconds, showing that cache reuse is not always a win.

arXiv 2610.21500 · 2026-10-15 · PDF

A method that applies the Prefill-Decode disaggregation design used in serving to multi-turn agent reinforcement learning Rollouts to make Rollouts more efficient.

arXiv 2610.16592 · 2026-10-12 · PDF

A scheduler that decides where to run multi-turn agent requests by weighing both cache reuse and queue congestion.

arXiv 2610.14686 · 2026-10-10 · PDF

Points out the noise in rewarding only one sampled answer out of many, and uses ReMaRk, which combines all answers and scores them at once, to make recommendation reinforcement learning cheaper and more stable.

arXiv 2610.17771 · Upvotes 70 · 2026-10-13 · PDF · Code

An industrial recommendation system framework that scales sequence encoding, behavior memory and candidate fusion separately and is designed together with serving.

arXiv 2610.20913 · 2026-10-14 · PDF

It uses the order of past tool calls to narrow down the next tool, showing only 5 of 128 tools while improving selection accuracy and cutting input tokens by 80%.

arXiv 2610.21267 · 2026-10-15 · PDF

A tool-using chest X-ray reading agent is trained with reinforcement learning to learn when additional evidence is needed and to filter out inappropriate quantitative measurements.

arXiv 2610.16811 · Upvotes 81 · 2026-10-12 · PDF

MutationProjector, pretrained on genomic variants from more than 30,000 tumors, showed state-of-the-art performance in predicting resistance to immunotherapy and anticancer therapy, and identified new biomarkers.

Cancer discovery · Europe PMC · 2026-10-01 · PDF

A German-language BERT-based tool processed 82,384 medical reports from about 200 institutions, automating in live operation the cancer registry coding that people used to do by hand.

ESMO real world data and digital oncology · Europe PMC · 2026-09-23 · PDF

MULTIFINGROUND, built from English, French and Greek disclosures, reveals the gap between LLMs recognizing accounting concepts and finding all of the relevant figures.

arXiv 2610.19432 · 2026-10-14 · PDF

It examines both the structure and the content of a transaction graph to produce human-readable explanations of why something was judged a fraud risk, and also proposes a way to evaluate their quality.

arXiv 2610.18023 · 2026-10-13 · PDF

An agentic options trading framework that, instead of being tied to a fixed strategy, chooses which option contracts to combine and how based on market conditions.

arXiv 2610.10407 · 2026-10-07 · PDF

It evaluates general-purpose robot policies under high-mix, low-volume assembly conditions such as product variants, moving materials, time limits and tight tolerances.

arXiv 2610.16782 · 2026-10-12 · PDF

An open benchmark that organizes industrial time series into source-specific forecasting tasks, compares major time-series models and shows the need for domain-specific evaluation.

arXiv 2610.16652 · 2026-10-12 · PDF

A fault diagnosis framework in which an LLM agent investigates sensor signals with analysis tools, while a host controls weak or conflicting evidence to prevent over-interpretation.

arXiv 2610.17786 · 2026-10-13 · PDF

Using 3,396 claims based on New York Court of Appeals decisions, it evaluates whether LLMs can judge if the cases they cite actually support a claim.

arXiv 2610.10971 · 2026-10-07 · PDF

Noting that the same statute can be interpreted differently depending on a judge's legal philosophy, it presents LePhilBench to evaluate the jurisprudential reasoning of LLMs.

arXiv 2610.19427 · 2026-10-14 · PDF

A RAG method that turns the structure of German legal documents into a knowledge graph to retrieve, in a verifiable way, key context that is easy to miss as the number of documents grows.

arXiv 2610.13695 · 2026-10-09 · PDF

A swarm of science agents that share experimental results across several research directions and refine hypotheses and implementations together, as an attempt at self-improvement of research ability.

arXiv 2610.20431 · Upvotes 68 · 2026-10-14 · PDF