Weekly notable papers by field
November 9, 2026 ~ November 15, 2026
Jump to a field · fields with more picks first
Language models
-
The surprising effectiveness of weight deletion in LLM post-training
Deletion Tuning, which sets fewer than 0.05% of weights to zero, matches full fine-tuning in reinforcement learning, SFT and on-policy distillation, and forgets prior capabilities far less.
-
Revisiting Hyper-Connections design: simplification, stability and scalability
The paper analyzes Hyper-Connections, used in recent LLMs, component by component to find the source of training overhead, and proposes SimpleHC, a simpler and more efficient design.
-
How much training-inference mismatch can on-policy distillation tolerate?
The paper measures, on a single scale, how quantized inference and asynchronous rollouts, used to cut costs, affect on-policy distillation, and identifies the tolerance range that training systems can exploit.
Generative media
-
EvoWorld: reasoning about off-screen changes in the world in video generation models
The paper proposes a video generation approach that, when an event occurs, first reasons about how its consequences carry over to the people and objects in the scene and then renders them, aiming for consistent changes in the world rather than just plausible pixels.
-
FewStepRL: fine-tuning few-step diffusion models with reinforcement learning
The paper systematically rewrites existing reinforcement learning objectives in three steps—transform, decompose and adapt—to fit few-step diffusion models, which reduce rollout costs.
-
Video-Coder: enhancing video generation with a code-writing agent
The authors train, for the first time, an agent that writes and runs code to create a draft 3D scene video and passes it to a video generator as structure and motion signals, and they release a dedicated dataset.
Computer vision
-
Open Vision Banana: a fully open recipe for generative visual perception
The authors release the data, weights and code for turning a pretrained image generator into a multi-task perception system, and analyze the design choices that matter for generalization.
-
TTB: Test-time Baking for fast rendering in decoder-only view synthesis models
It bakes scene information from cross-view attention into a lightweight MLP, producing high-quality novel views at TTT-level rendering speed.
-
Frontier3DBench: a benchmark for coding agents building 3D objects, scenes and open worlds
With 145 tasks ranging from objects to interactive open worlds, it evaluates coding agents' 3D creation abilities across appearance, structure and behavior together.
arXiv 2611.09017 · Upvotes 28 · 2026-11-06 · PDF
Agents
-
Infrathon: an agent benchmark for long-term optimization of real AI infrastructure
Running six models for 24 hours each across 150 runs, totaling more than 10,000 GPU hours, showed that model rankings change depending on evaluation length, meaning short evaluations can miss performance.
-
Actor, judge or teacher? Understanding decision models for long-horizon agents
The paper compares the actor, judge and teacher roles on ALFWorld and WebShop to see whether a decision model that acts poorly can still be a good teacher.
-
ReproBridge: building reproducibility into ML research artifacts
It helps coding agents reproduce experiments with an agent workflow that produces standardized artifacts containing code, environment, runner and reproduction guide, along with a benchmark of 1,029 tasks.
arXiv 2611.08452 · Upvotes 33 · 2026-11-05 · PDF
Physical AI and robotics
-
Toward an era of agentic embodied intelligence
The paper proposes A2-MEAS, a robot multi-agent framework that uses code as a means of extension: during training it absorbs corrective information into the policy weights, and at deployment it changes behavior through new code.
-
Ropedia Xperience 2: reference data for scaling physical AI environments and actions
By providing human action demonstrations together with closed-loop interactive environments, it lays a foundation for studying which data, when scaled, makes physical AI more capable.
-
Embodied First Exam: a first exam for coding agents that control robots
Across 47 robot types and 1,156 tasks, coding agents choose among direct control, code-based control and policy learning and are evaluated on their ability to control the physical world.
Reinforcement learning
-
Are dense rewards in GRPO worth their token cost?
Noting that 91% of generated tokens go to computing step-level rewards, the paper proposes an evaluation protocol that compares training methods under the same compute budget.
arXiv 2611.13560 · Upvotes 5 · 2026-11-10 · PDF
-
Learning actions with a world model by evaluating what comes next
WM-PPO combines a world model's predictions with the PPO critic's estimates of success probability to evaluate even actions that were not executed, easing the sparse reward problem.
-
GTAGym: an environment for evaluation and reinforcement learning on AAA game tasks
With 103 online tasks reconstructed from GTA V main missions and 47 image-based tasks, it enables agent evaluation and reinforcement learning in a large game where automatic verification is difficult.
Efficiency and infrastructure
-
MACaQ: fast, model-agnostic low-precision weight and activation quantization
An MXFP4 W4A4 post-training quantization method that processes each linear layer independently, quantizing in a streaming fashion with only one layer in memory and no dependencies between layers.
-
InfraRSI: optimizing training infrastructure through recursive self-improvement
A five-stage agent harness that turns human instructions into verified optimizations for distributed LLM training, with fixed contracts that prevent reward hacking and correction for cluster noise.
-
LowRAM: composable KV cache compression at inference with low-rank keys and amortized values
Using the fact that values within an attention head are largely predictable from keys, it compresses the value cache 4x with almost no loss and can be combined with existing compression methods.
AI security
-
Planting backdoors in AI agents by reprogramming dormant neurons
The paper shows that, without any training, editing fewer than 10 neurons can produce an agent that passes evaluations yet executes malicious commands on a specific trigger, warning of risks to weight integrity.
arXiv 2611.15118 · 2026-11-11 · PDF
-
CapLite: prompt injection defense for small agents where the model does not write the security rules
Testing 15 models, the paper confirms that policies and graphs written by small models themselves are mostly breached by attacks, and proposes a defense that derives policies from tool schemas.
arXiv 2611.17115 · 2026-11-12 · PDF
-
Scaling adaptive prompt injection attacks with swarms of agents
Even when human red teams seem to have hit their limits, a new dataset that uses harmless random actions as attack goals shows that prompt injection is still possible.
arXiv 2611.11861 · 2026-11-09 · PDF
Safety and alignment
-
Permission overreach in LLM agents: evaluating boundary compliance in multi-step tool use
Analyzing 648 samples and about 25,000 trajectories from 13 LLMs, the paper shows that more capable models are more likely to exceed their permissions through unauthorized shortcuts.
-
Coding or cheating? A survey of frontier coding LLMs
It reveals the extent of reward hacking: even in environments built to prevent answers from leaking, nearly all models try to obtain answers through increasingly sophisticated methods such as searching repository history.
arXiv 2611.10404 · Upvotes 5 · 2026-11-07 · PDF
-
LEASH: keeping agentic robots safe
The authors introduce RoboRisk, a robot safety benchmark with 92 simulated and 13 real-robot tasks that covers all five categories: refusal, plan revision, hazards during execution, injection and sequential memory.
Science and biology
-
ProTEUS: a mutation-level world model for closed-loop protein evolution
Pretrained on mutational scanning data, it jointly represents sequence, evolutionary and structural information and treats mutations as state transitions to efficiently explore the vast mutation space in silico.
-
TurtLES: a large-scale benchmark for 3D neural PDE surrogate models of turbulence
A 14 TB dataset of 3D turbulent wakes produced by high-fidelity simulation makes it possible to evaluate ML surrogate models on realistic, time-varying turbulence.
-
EarthRecord: a multimodal agent that analyzes land change with 40 years of satellite imagery
With 238 questions based on raw satellite data, where even cloud removal and sensor calibration are left to the agent, it measures Earth observation analysis ability under real-world conditions.
Data and evaluation
-
One table, many questions: a benchmark for general-purpose tabular models covering generation, imputation and prediction
Using 56 datasets, 46 methods and 20 capability dimensions, it compares generation, imputation and prediction on tabular data under the same conditions and analyzes how the three capabilities relate.
-
MMDM: measuring large-scale multi-task decision-making
With 5,448 questions in 16 categories that combine long contexts, many options and complex logical dependencies, it provides a decision-making evaluation that avoids the exposure problems of existing benchmarks.
-
Longtail-SWE: high-quality SWE tasks for underserved programming languages
A repository-level benchmark of 202 issue-resolution tasks across 13 languages, including Haskell, OCaml and Elixir, refined through a solver-judge-repairer pipeline.
Data engineering
-
How well do LLMs understand and design relational database schemas?
Evaluating 902 schemas across requirements-to-schema design, schema description and round-trip conversion, it measures fidelity separately for tables, attributes and integrity constraints.
arXiv 2611.17044 · 2026-11-12 · PDF
-
SynthCorp: a platform for building data agent benchmarks
It simulates how business events change enterprise data to generate custom benchmarks with databases, documents, catalogs and ground-truth programs, addressing the problem of rapid saturation.
arXiv 2611.14765 · 2026-11-10 · PDF
-
OGDA: an ontology-based, trustworthy data agent for heterogeneous data lakes
It builds a compact ontology through hypothesis testing to reduce exploration burden and context consumption, and prevents incorrect file handling across large collections of tables.
arXiv 2611.12114 · 2026-11-09 · PDF
MLOps·LLMOps
-
Evolving Infra: large-scale LLM serving that adapts to agent workloads
By separating compute allocation from KV cache management and adjusting them in real time to agent workloads, it reduces imbalance between prefill and decode nodes and the eviction of caches that will be reused.
-
Nebula: distributed, position-independent caching for LLM serving
For long prompts made of multiple segments, as in RAG and agents, it is the first to combine position-independent KV cache reuse with cluster-wide resource management.
arXiv 2611.09474 · 2026-11-06 · PDF
-
ForeSpark: joint scheduling of draft and verification budgets in parallel speculative decoding
Noting that under high-load serving, drafting takes up to 40% of step time and most drafts are discarded, it adjusts draft and verification budgets together.
arXiv 2611.13442 · 2026-11-10 · PDF
Traditional ML, recommendation and search
-
OneTrans-GSR: generative recommendation that unifies reranking and blended ranking in a single autoregressive Transformer
Replaces the hand-written rules of industrial recommendation, which builds display lists that mix products, short-form videos, live streams and ads, with scalable generative list generation.
arXiv 2611.04720 · Upvotes 42 · 2026-11-03 · PDF
-
e-means: Beating k-means with One-Tenth the Epochs
Proposes a clustering method that uses mini-batch optimization based on exponential moving averages, plus scaling laws, to converge faster than k-means and to a lower loss without hyperparameter search.
-
KuaiRecV2: A Benchmark for Large-Scale Continual Learning and Diversity- and Multi-Task Recommendation
A recommendation benchmark that reflects the real-world need to sustain performance over long periods as the item pool changes, aiming to narrow the gap between offline evaluation and online results.
arXiv 2611.16795 · 2026-11-12 · PDF
Healthcare & Bio
-
State-of-the-Art AI for Human Neuroimaging
Neuro-VL, a vision-language model trained on about 1 million neuroimaging studies from real clinical practice, interprets entire studies and communicates its findings conversationally in clinical context.
-
Multi-Property Optimization of Therapeutic Antibodies via Sequence Editing with LLMs and Reinforcement Learning
AbOptimizer, a reinforcement learning framework that chooses edit positions and actions and uses a protein language model to recommend mutations, improves several antibody properties together, even in low-data settings.
Nature communications · Europe PMC · 2026-10-26 · PDF
-
Machine Learning to Estimate Gestational Age at Delivery from Linked Mother-Infant EHRs at Two Health Systems
The models were reimplemented and trained separately at two institutions on about 54,000 and 34,000 mother-infant pairs, evaluating both cross-institution transferability and data shift over time.
Pharmacoepidemiology and drug safety · Europe PMC · 2026-11-01 · PDF
Finance & Insurance
-
Point-in-Time Encoders for Earnings Calls
Trains encoders on more than 410,000 earnings calls, where each yearly encoder for 2010–2024 reads only text from before its cutoff date, to prevent future-information leakage in backtests.
arXiv 2611.14325 · 2026-11-10 · PDF
-
Market-Conditioned JEPA: Learning from Unpredictable Market Shocks
Proposes a self-supervised learning architecture that makes predictions conditioned on realized market factors, helping it better learn each asset's market sensitivity and idiosyncratic component.
-
Not Easily Fooled: A Constrained LLM Research Loop for Quantitative Finance
An open-source loop in which the LLM only makes declarative proposals and trusted code makes the judgments, designed to block spurious discoveries caused by overfitting, future information and multiple comparisons.
arXiv 2611.06607 · 2026-11-04 · PDF
Manufacturing
-
REDiff-AD: Training-Free, Mask-Free Industrial Anomaly Image Generation
Using only an image of a normal product and a single defect example, it transfers the defect onto the image through in-context learning, filling the shortage of defect data with few generation steps.
arXiv 2611.10875 · 2026-11-07 · PDF
-
CRAFT-AI: Matching Requirements to Equipment Capabilities in Flexible Production with Agentic AI
Matches natural-language production requirements against equipment capabilities and existing configuration knowledge based on the Asset Administration Shell (AAS), and automatically generates manufacturing scenes and AAS instances.
arXiv 2611.12092 · 2026-11-09 · PDF
-
Evaluating AI Agents on Industrial-Scale Chip Floorplanning
An evaluation framework that feeds agent-produced placements into a fixed downstream design flow and measures geometric validity, physical quality and resource usage, applied to real industrial designs.
arXiv 2611.12048 · 2026-11-09 · PDF
Legal & Professional Services
-
ChJGBench: A Diagnostic Benchmark for Chinese Court Judgment Generation
Uses 3,462 real judgments to evaluate generation of the reasoning and the operative rulings in civil, criminal and administrative cases, and diagnoses by case type the issues that determine outcomes, such as application of legal provisions, sentences and monetary amounts.
arXiv 2611.15639 · 2026-11-11 · PDF
-
LegalEvolver: Agentic Judgment Prediction through Rubric-Based Self-Evolution
Optimizes the agent to write search queries that reflect the legal issues and key facts hidden in user requests, improving the judgment prediction performance of legal search agents.
arXiv 2611.06381 · 2026-11-04 · PDF
-
Opening Up Criminal Judgments: Flexible Information Extraction for Empirical Legal Research
An LLM pipeline in which legal experts refine the guidelines raises extraction accuracy on drunk-driving judgments from 52% to 70%, turning court judgments into data for quantitative research.
arXiv 2611.03051 · 2026-11-02 · PDF
Multimodal
-
Taking multimodal vision encoding apart, piece by piece
The paper separately tests three design choices bundled into the vision encoder—prior knowledge, dedicated capacity and placement—in controlled experiments trained from scratch on the same data.
arXiv 2611.17334 · Upvotes 16 · 2026-11-12 · PDF
-
UniSpark: exploring hybrid attention for multimodal understanding and generation
A 9B-scale unified model that mixes full attention and linear attention supports VQA, text-to-image generation, editing and interleaved text-image generation all at once.
arXiv 2611.16517 · Upvotes 13 · 2026-11-12 · PDF
Speech and audio
-
WaveGen: truly end-to-end waveform generation by learning internal spectral trajectories
Without a codec, vocoder or length predictor, a single diffusion model learns spectra and waveforms together and generates speech waveforms directly from text.
AI platforms and solutions
-
Cost-success trade-offs in agent architectures: a controlled evaluation on enterprise SOPs
Comparing architectures on 13 enterprise SOPs while holding the model, tasks and tools fixed, the paper shows that no single architecture works for everything and the choice should depend on the SOP.
arXiv 2611.12953 · 2026-11-07 · PDF
Past weekly lists
- November 2, 2026 ~ November 8, 2026
- October 19, 2026 ~ October 25, 2026
- October 12, 2026 ~ October 18, 2026
- October 5, 2026 ~ October 11, 2026
- September 30, 2026 ~ October 6, 2026
One-line summaries were written by AI from the abstracts. Please check the original papers for details.