Weekly notable papers by field
September 30, 2026 ~ October 6, 2026. AI papers are ranked by Hugging Face Daily Papers upvotes; healthcare papers are picked from open-access papers on Europe PMC (PubMed Central, bioRxiv, medRxiv).
Jump to a field · fields with more picks first
Language Models
-
The Teacher model is a direction, not a destination: pushing reinforcement learning's Representation changes further in On-policy Distillation
Extends the changes made by reinforcement learning in the Hidden Representation space rather than in output probabilities (Extrapolation), making Distillation in which the Student model surpasses the Teacher more stable.
-
Scaling properties of On-policy Distillation between models of the same family
Finds a regularity in which, when Distilling from a larger model to a smaller one or from a smaller model to a larger one, accuracy in early training rises roughly in proportion to the square root of the KL distance.
arXiv 2609.32722 · Upvotes 323 · 2026-09-26 · PDF
-
Transformers stop thinking too early, and a single small LoRA fixes it
A rank-8 LoRA attached to a single layer raises Qwen3-8B's accuracy on 24-line reference tracking from 15.5% to 99%, and the paper also analyzes how it works internally.
Generative Media
-
MaLiang-Harness: making images and video through programs
Defines the problem where code runs but the resulting visual output differs from the request, and proposes a generation framework that keeps cycling through making, checking and fixing.
-
Adaptive reward routing: multi-reward reinforcement learning for joint audio-video generation
Instead of fixing where and how multiple rewards (such as visual quality, semantic match between sound and picture, and sync) are applied during training, it keeps adjusting them.
arXiv 2609.37200 · Upvotes 138 · 2026-09-29 · PDF
-
Think before you score: a thinking reward model for visual generation
A reward model that, before evaluating generated output, first sets up a rubric of what to look for in each case and then assigns fine-grained scores according to it.
Computer Vision
-
OneStreamer: unifying perception, memory and proactive response in real-time video dialogue
A video LLM that builds a hierarchical caption memory with temporal information while watching streaming video, and responds proactively once it has enough evidence.
-
LEGO-Anything: reconstructing 3D scenes with a coding agent
From a single image, it writes and revises Blender code to build an editable 3D scene program, and the authors also release an evaluation set of 208 images.
arXiv 2609.36380 · Upvotes 137 · 2026-09-28 · PDF
-
Beyond the timeline: strengthening long-video memory with per-object 'life stories'
A memory framework that groups the scenes in a long video where the same object appears into a single life story, so it can answer questions spanning multiple times and days.
Multimodal
-
LoopVL: visual intelligence that thinks through repeated computation
Extends looped transformers, which run a shared module repeatedly, to vision-language models, achieving higher visual reasoning performance than standard models of similar and larger size.
-
UniEvo-VL: an On-policy Self-Distillation recipe for self-improvement of multimodal models
A method in which a single multimodal model, without a separate Teacher model, uses critiques it writes itself to take on both the Teacher and Student roles during training.
arXiv 2609.38721 · Upvotes 292 · 2026-09-30 · PDF
-
Revisiting latent visual reasoning: anchoring latent reasoning to visual evidence
Finds that when reasoning with latent tokens instead of words, models barely respond to changes in the image, and addresses this with supervision signals aligned to visual evidence.
Agents
-
Fake progress: diagnosing and mitigating 'collusive cheating' in self-evolving search agents
Reveals a failure in which the model that poses problems and the model that solves them agree on the same error so that only the reward goes up, and proposes a multi-sample verification method that checks problems before training.
arXiv 2609.39102 · Upvotes 671 · 2026-09-30 · PDF
-
Raven: a harness of harnesses for compositional agent intelligence
Releases an open-source multi-agent ecosystem that automatically builds domain-specific agent harnesses, improves them through experience, and coordinates them across domains.
-
Omni-IO Skills: a harness that fits agents to every modality
A plugin-style harness that lets existing agents handle text, image, audio, video, 3D and code outputs without retraining the model.
Physical AI & Robotics
-
What makes world action models generalize? An empirical study of future prediction during inference
Shows through controlled experiments that latent models that skip future video during inference lose the generalization advantage, even though they perform the same within the training distribution.
-
MotorMind: training-free robot manipulation with general-purpose vision-language models
Proposes a framework that, without dedicated training or many external tools, helps a general-purpose VLM observe like a human teleoperator and issue robot actions directly.
arXiv 2609.38078 · Upvotes 97 · 2026-09-29 · PDF
-
PanoVLN: vision-language navigation using panoramic video
Diagnoses that simply switching to panoramas has little effect, and presents a navigation method that properly uses the wide field of view to choose routes.
Reinforcement Learning
-
The 'sharpening tax' of post-training
Pretrained models with a simple harness had lower single-attempt accuracy, but given multiple attempts they often solved more problems than post-trained models.
-
Learning from what you did not sample: cross-model Trajectory exchange in reinforcement learning (RLVR)
Proposes GRAFT, which borrows successful solutions from other models on problems where all attempts failed, letting models learn from each other without a separate strong Teacher.
arXiv 2609.37868 · Upvotes 63 · 2026-09-29 · PDF
-
EasyPPO: the key is stabilizing the critic
Identifies and fixes two causes of critic failure that make PPO for LLMs unstable: filtering of truncated responses and differences in noise across prompts.
arXiv 2609.36802 · Upvotes 36 · 2026-09-29 · PDF
Efficiency & Infrastructure
-
Periodic weakness: the phase sensitivity created by chunk-wise KV cache compression
Reveals a hidden weakness: when the KV cache is compressed chunk by chunk, long-context retrieval accuracy varies by up to 40 percentage points depending on where the information is located.
arXiv 2609.36322 · Upvotes 100 · 2026-09-28 · PDF
-
Prefill-free KV cache transfer between heterogeneous multi-agent LLMs
Transfers the sender's KV cache to fit the receiver even between models with different tokenizers and architectures, reducing the waste of reprocessing shared context.
-
FocusVTC: efficient visual text compression with adaptive resolution
When compressing text by rendering it as images, it views the whole at low resolution and zooms in only on the parts it needs, reducing the trade-off between token savings and readability.
Safety & Alignment
-
Source preference in the wild: LLM agents favor items based on their source
Shows that 12 agent models prefer items from particular sites even when conditions are equal, and pick an item that meets one fewer requirement about two-thirds of the time.
arXiv 2610.03195 · Upvotes 37 · 2026-10-02 · PDF
-
Thinking outside the frame: can language models be selective about following external guidance?
By varying only the reliability of the guidance, shows that recent models are easily pulled into flawed workflows, and tests whether training can improve this.
arXiv 2609.39578 · Upvotes 68 · 2026-09-30 · PDF
-
Persona dosing: activation steering that adjusts the strength of personality traits step by step
A method that calibrates activation steering so that a desired personality trait appears at a numerically specified intensity; it was more effective than existing approaches on three models.
arXiv 2609.36388 · Upvotes 54 · 2026-09-28 · PDF
Science & Bio
-
Does learning protein folding generalize to broader reasoning ability?
Post-trains general-purpose models with question-answer data built from protein structures and 3D structural signals, and examines whether spatial and structural reasoning transfers to other domains.
-
LANTERN: uncovering mathematical knowledge hidden inside language models
Uses the model's internal activations to rank and verify candidate relationships between integer sequences, finding 62 relationships not yet linked to each other in OEIS.
arXiv 2609.32264 · Upvotes 49 · 2026-09-26 · PDF
-
PDE-JEPA: predictive Representation learning for parametric partial differential equations
Pretrains Representations to predict how the state will change instead of directly reconstructing observations, in order to model the dynamics of physical systems.
Data & Evaluation
-
RealCompanion: evaluating the ability to understand people with real long-term conversations
A benchmark that releases 27,218 messages exchanged over up to 120 days in 10 real relationships between people and AI companions, along with evidence labels.
arXiv 2610.01780 · Upvotes 269 · 2026-10-01 · PDF
-
EngiWorld: what can today's agents accomplish in real engineering environments?
The first benchmark to evaluate the entire design process, with 1,301 tasks spanning 6 fields such as CAD, CAE and EDA and 26 professional software tools.
-
Science or 'slop': measuring and mitigating scientific sloppiness in AI-generated papers
Builds a benchmark that evaluates 390 AI-generated papers, which look plausible in parts but whose logical connections fall apart, paired with human-written papers.
Data Engineering
-
Zephon: Elastic Determinism for online, stateful foundation model data loading pipelines
A data loader that guarantees the same training batch order even when the GPU configuration changes or training resumes from a checkpoint, making the results of expensive ablation experiments trustworthy.
arXiv 2610.03087 · 2026-10-02 · PDF
-
TabJoinBench: a benchmark for finding joinable tables
Introduces the first common benchmark covering semantic, relational and mixed scenarios, so that methods for finding tables to join in a data lake can be compared fairly.
arXiv 2610.00817 · 2026-09-30 · PDF
-
WeaveData: a multimodal data analysis system that critiques and improves LLM analysis plans on its own
A system in which an LLM drafts an analysis plan for questions spanning tables, text and images, then reviews it step by step before execution to reduce plans that are silently wrong.
arXiv 2609.34764 · 2026-09-28 · PDF
MLOps·LLMOps
-
MOLT: a fine-grained memory sharing system for LLM serving that runs fine-tuning on spare GPU memory
Proposes a method that lends idle GPU memory during serving to PEFT fine-tuning and reclaims it within a few seconds when requests surge, to meet latency SLOs.
arXiv 2610.05748 · 2026-10-05 · PDF
-
Lachesis: splitting the KV cache between HBM and high-bandwidth flash by lifetime in agent serving
Splits the KV cache that builds up in agent workloads between HBM and flash according to its lifetime, greatly increasing memory capacity while avoiding flash's write endurance limits.
arXiv 2610.08378 · 2026-10-06 · PDF
-
ServeTwin: a benchmark-validated simulator for distributed LLM serving design
A simulator that lets you test distributed LLM serving designs in advance without a real cluster, including prefill-decode disaggregation and multi-turn agent workloads.
arXiv 2610.02732 · 2026-10-02 · PDF
Traditional ML, Recommendation & Search
-
What should we learn from a VLM Teacher? Comparing supervision signals for visual document retrieval
Having the VLM Teacher identify wrong-answer candidates (Hard Negatives) raised ViDoRe v2 nDCG@5 from 55.2 to 62.6–63.0, giving a practical answer to where the Teacher should be used.
arXiv 2610.09177 · 2026-10-06 · PDF
-
Disentangling the effects of paradigm, identifiers and decoding in generative retrieval
Compares autoregressive and diffusion generative retrieval under the same conditions, showing that changing only the decoding method shifts Hit@1 by 6.6–13.7 points.
arXiv 2610.08716 · 2026-10-06 · PDF
-
UNREAL: unifying retrieval and long-context processing in a single model
Uses the internal Representations of a frozen LLM to handle both corpus retrieval and evidence selection in long contexts, with fewer than 500,000 trainable parameters.
arXiv 2610.08463 · 2026-10-06 · PDF
Medicine & Healthcare
-
Precisely predicting treatment response with a cancer genotype foundation model
MutationProjector, pretrained on genetic mutations from over 30,000 tumors, matched or exceeded the previous state of the art in predicting response to immunotherapy and chemotherapy.
Cancer discovery · Europe PMC · 2026-10-01 · PDF
-
Guideline-based LLM chatbots supporting self-management for rheumatology patients: a nationwide multicenter evaluation
Ten condition-specific chatbots based on German clinical guidelines were deployed at 13 centers and evaluated on 6,291 real patient questions.
Journal of medical systems · Europe PMC · 2026-10-05 · PDF
-
Extracting tumor information from unstructured medical reports: applying an AI tool at the Cancer Registry of Eastern Switzerland
A tool built by fine-tuning a German BERT processed 82,384 medical reports, aiming to reduce the burden of cancer registry coding that people had been doing by hand.
ESMO real world data and digital oncology · Europe PMC · 2026-09-23 · PDF
Speech & Audio
-
VoxMem: a benchmark for the multimodal memory of large audio language models
Evaluates whether models remember across multiple sessions not only what was said, but also who said it, in what tone, and what sounds occurred.
-
Beyond one-on-one conversation: interaction-aware memory for multi-party spoken dialogue
Handles long-term multi-party spoken dialogue with incremental speaker identification, a hierarchical memory that records who spoke to whom, and agentic retrieval.
Autonomous Driving
-
DyRAD: novel view synthesis of radar in dynamic driving scenes
Synthesizes radar observations of dynamic driving scenes from new viewpoints, including Doppler velocity, and can be used for closed-loop autonomous driving evaluation.
Past weekly lists
One-line summaries were written by AI from the abstracts. Please check the original papers for details.