Weekly notable papers by field

The authors propose a Distillation method in which the Student learns to reproduce only the ordering of candidate tokens instead of the Teacher's token probabilities, making it less sensitive to the Teacher's confidence and temperature settings.

arXiv 2611.02112 · Upvotes 35 · 2026-11-01 · PDF

Using Qwen3, the authors show that OPD combining several Teachers looks almost perfect within the training distribution but does not fully transfer Extrapolation generalization on math, coding and agent tasks.

arXiv 2611.05609 · Upvotes 24 · 2026-11-03 · PDF

A 4-billion-parameter model that repeats the same block any number of times so that a single checkpoint can fit multiple inference compute budgets. It was pretrained on 6 trillion tokens and supports a 256K context.

arXiv 2611.03998 · Upvotes 12 · 2026-11-02 · PDF

Drawing on experience building Kling, the authors argue that video generation quality depends on multimodal understanding across the whole pipeline, from data and conditioning to Alignment and evaluation, and describe how they use a shared foundation model, Kling-MLLM.

arXiv 2611.04916 · Upvotes 75 · 2026-11-03 · PDF

To reduce the visual drift that builds up when edited outputs are reused for further edits, the authors propose a framework that trains only the final steps while keeping existing generation and editing abilities.

arXiv 2611.06741 · Upvotes 40 · 2026-11-04 · PDF · Code

To stay consistent even when an object that left the frame comes back, a few-step video generation model is trained on the difference between the Teacher's predictions with and without past context.

arXiv 2611.02782 · Upvotes 19 · 2026-11-02 · PDF · Code

For the task of finding a target image from a reference image and a modification text, the authors separate the step of first understanding the intended change from the step of building the retrieval embedding, making better use of the MLLM's compositional understanding.

arXiv 2611.05193 · Upvotes 48 · 2026-11-03 · PDF · Code

The method lifts the outputs of a pretrained 2D tracker and a 3D geometry model into 3D, then efficiently corrects errors in occluded segments, avoiding the shortage of 4D training data.

arXiv 2610.42136 · Upvotes 28 · 2026-10-30 · PDF · Code

A recurrent Transformer that keeps solving from its current estimate, so it can extend a map without recomputing views it has already processed, or refine the geometry as more observations arrive.

arXiv 2611.02857 · Upvotes 18 · 2026-11-02 · PDF · Code

To resolve the mismatch between reasoning in discrete tokens and image generation in continuous space, the authors propose LARC, which reasons in continuous space, so reasoning can be used for image generation.

arXiv 2610.42149 · Upvotes 45 · 2026-10-30 · PDF · Code

The first large-scale multimodal dialogue dataset and benchmark that can only be solved by understanding, maintaining and updating a dialogue state that changes each turn. It fills a gap left by evaluations that focused mostly on single turns.

arXiv 2611.05280 · Upvotes 32 · 2026-11-03 · PDF

The authors built the first benchmark that measures the ability to recall earlier observations and statements during multi-turn conversations about long videos.

arXiv 2610.40131 · Upvotes 31 · 2026-10-29 · PDF · Code

Using only two light rounds of SFT and reinforcement learning, without separate mid-training, DPO or OPD, the authors built a fully open model with broad agent capabilities.

arXiv 2611.03233 · Upvotes 49 · 2026-11-02 · PDF · Code

The method extracts dependencies between tool calls as rules from real MCP service execution logs and synthesizes environments from them, producing realistic tool combinations and tasks that span multiple environments.

arXiv 2611.05036 · Upvotes 41 · 2026-11-03 · PDF · Code

Sampling many independent Trajectories for the same issue leads to growing redundancy. The authors propose a data collection method that reduces this by using earlier collection history as a kind of advice.

arXiv 2611.02074 · Upvotes 40 · 2026-11-01 · PDF

Using a distributed evaluation system that links robots at multiple sites, the authors evaluated 5 VLA policies and 3 AI agents on 4 types of robots, including humanoids, through blind pairwise comparisons.

arXiv 2611.02497 · Upvotes 50 · 2026-11-02 · PDF · Code

The authors released a unified harness for controlling robots with open-weight models. The default Qwen3.8-Flash-Next scored 31.39 on RoboDojo, beating the published GPT-6-based result of 28.97.

arXiv 2611.01560 · Upvotes 48 · 2026-11-01 · PDF · Code

The first fully open recipe for pretraining on 50,000 hours of first-person video annotated with hand motions and descriptions. The authors found a log-linear scaling law between hours of pretraining data and robot success rate.

arXiv 2611.06817 · Upvotes 32 · 2026-11-04 · PDF

The authors separate two sources of mismatch, numerical differences in the inference engine and stale weights from asynchronous sampling, and analyze which properties determine the correction method needed.

arXiv 2610.40945 · Upvotes 11 · 2026-10-29 · PDF · Code

For emotional expression videos, which have no ground truth, fixed rewards miss the intended emotion and invite reward hacking. The authors address this with a self-evolving reward calibration method.

arXiv 2611.06518 · Upvotes 26 · 2026-11-04 · PDF · Code

Using an entropy-based auxiliary reward to distinguish reasoning processes that correctness alone cannot tell apart, the method raised the self-improvement and transfer performance of the small model Qwen3-1.7B by 2.08 percentage points.

arXiv 2610.41296 · Upvotes 9 · 2026-10-30 · PDF

The authors identify three reasons why speculative decoding with parallel Drafters does not achieve its full possible speedup in practice, and propose a method that optimizes automatically for each deployment environment.

arXiv 2611.02501 · Upvotes 20 · 2026-11-02 · PDF · Code

A set of kernels that use MXFP8 and NVFP4 low-precision arithmetic for attention training while addressing quantization cost, leakage of future information through shared scales, and forward-backward mismatch.

arXiv 2611.03344 · Upvotes 19 · 2026-11-02 · PDF · Code

A Diffusion Drafter that works without the target model's Hidden State. It drafts the next block during verification, so drafting and verification can run at the same time.

arXiv 2611.06425 · Upvotes 19 · 2026-11-04 · PDF · Code

Based on real fraud cases, the authors built 608 tasks using forged multimodal content to evaluate how well personal AI agents withstand social engineering attacks.

arXiv 2611.08281 · 2026-11-05 · PDF

The authors define, in terms of permissions, indirect injections that manipulate only the outcome without changing the task itself, and evaluate them with 90 verified cases across three tool-use environments.

arXiv 2611.07640 · 2026-11-05 · PDF

The method neutralizes backdoors hidden in third-party LoRA adapters with a lightweight fix that corrects the directional distortion of the poisoned adapter, leaving the base model unchanged.

arXiv 2611.03752 · 2026-11-02 · PDF

Unlike existing approaches that look only at the command text, it runs each shell command in advance in a snapshot sandbox, decides whether to approve it based on its actual effects, and rolls back blocked commands.

arXiv 2610.40753 · Upvotes 15 · 2026-10-29 · PDF · Code

Every tool call an agent makes is checked before execution against a policy written in Lean4. Rejected calls can even be repaired, with proof-based corrections, into a close alternative that preserves the user's intent.

arXiv 2611.02332 · Upvotes 12 · 2026-11-02 · PDF · Code

The method automatically finds datasets that look ordinary but trigger unexpected behavior outside the training domain, extending research that had relied on a small number of hand-crafted cases.

arXiv 2611.06324 · 2026-11-04 · PDF

The authors propose an agent that systematically optimizes its workflow and model architecture and builds up memory across tasks, continually improving its molecular search strategy itself through experience.

arXiv 2610.39714 · Upvotes 18 · 2026-10-29 · PDF

Comparing more than 90,000 cases of how well LLM research proposals, written using only 2025 literature, matched actual research in 2026, the authors show that LLMs fall far short of the human research community.

arXiv 2611.03424 · Upvotes 47 · 2026-11-02 · PDF · Code

It evaluates LLMs' ability to solve research-level problems using a collection of 12,151 open research problems and a benchmark of 200 problems, including math problems solved in 2026, formalized in Lean 4.

arXiv 2610.37739 · Upvotes 6 · 2026-10-27 · PDF

The authors show that the reason existing methods that pick hard examples underperform random selection in LLM fine-tuning lies in coverage, and confirm this with a simple clustering-based selection method.

arXiv 2611.07323 · 2026-11-05 · PDF

The authors distinguish whether curation raises the achievable performance or reaches the same performance faster, and show that raising the performance ceiling requires a fixable gap between the training data and the target.

arXiv 2610.36983 · 2026-10-27 · PDF

The authors point out that in existing Gradient-based data selection, attribution scores swing widely when only wording or format changes slightly, and propose CIGA, which focuses on content.

arXiv 2610.36755 · 2026-10-27 · PDF

In LLM-based semantic filtering, which filters records using natural-language conditions, the method minimizes cost over the full set of per-stage threshold combinations, reducing calls to expensive models.

arXiv 2611.08367 · 2026-11-05 · PDF

The authors built executable DBMS environments at scale that reproduce a wide range of real anomalies, so database agents can be trained and evaluated on diagnosis, recovery and performance optimization.

arXiv 2611.07339 · 2026-11-05 · PDF

Models silently settle computation choices that a request leaves open, such as counting rows versus distinct entities. The authors expose this with SpecBridge, which actually executes sets of requests that differ only in that choice and checks the results.

arXiv 2611.04318 · 2026-11-03 · PDF

A serving memory management approach that removes the fixed boundary between memory reserved for model weights and memory for attention state, and flexibly reallocates memory as the workload changes.

arXiv 2610.40305 · 2026-10-29 · PDF

Shows that first-come-first-served and prefill-first policies designed for autoregressive models significantly reduce speed when serving Diffusion LLMs, and proposes a scheduling engine tailored to them.

arXiv 2610.39852 · 2026-10-29 · PDF

Addresses the reliability problem of agentic AutoML, where intermediate errors propagate silently, by having skills grounded in expert knowledge evolve with each run and pass verification at every stage.

arXiv 2611.04180 · 2026-11-03 · PDF

Points out the limitation of reranking evaluators that reduce a user's continuous browsing process to independent per-item click predictions, and proposes an approach that models the browsing journey itself.

arXiv 2611.07591 · 2026-11-05 · PDF

Identifies a user-independent item-average component as the reason recommendations concentrate on a few items when LLM recommenders are trained with reinforcement learning on collaborative filtering scores, and proposes two reward corrections.

arXiv 2611.07588 · 2026-11-05 · PDF

Studies how the value of diversity, uncertainty and information-gain criteria changes as labels accumulate, in tabular foundation models that receive labels only as context.

arXiv 2611.07886 · 2026-11-05 · PDF

Built a gestational age estimation model from electronic health records of about 89,000 linked mother-newborn pairs at two institutions, and validated it through an independent reimplementation at another institution.

Pharmacoepidemiology and drug safety · Europe PMC · 2026-11-01 · PDF

Identified glycopeptide biomarkers in peritoneal dialysis patients, built a prediction model combining them with clinical indicators, and validated it in a prospective cohort and a multicenter external cohort.

Clinical and translational medicine · Europe PMC · 2026-11-01 · PDF

A model that classifies viral keratitis subtypes by fusing slit-lamp images taken under three illumination methods, validated across multiple centers with data from two external institutions.

Frontiers in cell and developmental biology · Europe PMC · 2026-10-19 · PDF

For trading research, where backtest improvements easily turn out to be artifacts of overfitting or look-ahead leakage, releases an open-source research loop in which the model only makes proposals and trusted code makes the judgments.

arXiv 2611.06607 · 2026-11-04 · PDF

Built the OLIVIA protocol to test the claim that LLM interpretations of news predict stock returns without look-ahead information, memorization or selection bias, and applied it to S&P 500 news.

arXiv 2611.03985 · 2026-11-02 · PDF

Analyzed more than 15,000 answer pairs based on SEC filings, finding among other things that hallucinated answers are more than twice as long as factual ones, and evaluated detection performance by hallucination type.

arXiv 2611.02258 · 2026-11-01 · PDF

Proposes MUSTARD, a manufacturing quality inspection framework that adds spatiotemporal reasoning to MLLMs to catch dynamic defects, such as irregular motion, that are invisible in static images.

arXiv 2611.08080 · 2026-11-05 · PDF

Evaluated an approach that pretrains a single encoder with self-supervised learning on battery, semiconductor CMP, CNC machining and bearing data, then reuses it for predictive maintenance and manufacturing tasks with few labels.

arXiv 2611.02489 · 2026-11-02 · PDF

Reproducing the industrial anomaly detection method INP-Former revealed a limitation: image AUROC of 99.7 on MVTec AD dropped to 63.4 on MVTec AD 2.

arXiv 2611.01230 · 2026-10-31 · PDF

Addresses the limitation of legal search agents that generate search queries directly from user requests and miss hidden legal issues, using a query optimization framework for judgment prediction.

arXiv 2611.06381 · 2026-11-04 · PDF

An LLM pipeline in which legal experts iteratively refine guidelines to structure criminal judgments; extraction accuracy on drunk-driving cases rose from 52% to 70%.

arXiv 2611.03051 · 2026-11-02 · PDF

Tackles the difficulty of linking long factual narratives to short statutory provisions with a decompose-and-refine pipeline in which an LLM splits the narrative by issue and generates search queries.

arXiv 2610.38336 · 2026-10-28 · PDF

The method uses idle compute during listening to decode ahead by assuming the next segment is silent, based on the observation that results do not change in 95.3% of segments even when real audio arrives.

arXiv 2611.06982 · Upvotes 8 · 2026-11-04 · PDF

Streaming video-to-audio generation, held back by real-time constraints, had lagged far behind offline models. The method tackles bottlenecks in both the visual input stage and the audio generation stage, producing sound with imperceptible latency.

arXiv 2611.05988 · Upvotes 7 · 2026-11-04 · PDF · Code

Four open decision models that, given a question and answer options, directly output a probability for each option without generating tokens, reducing inconsistent answers, output parsing, cost and latency problems in production services.

arXiv 2611.07854 · Upvotes 22 · 2026-11-05 · PDF · Code

In a real enterprise Text-to-SQL case, agents were tasked with uncovering customer requirements and carrying out follow-up training; the top model found more than 80% of the hidden requirements.

arXiv 2611.03551 · 2026-11-02 · PDF

In a Transformer that processes radio-wave LiDAR data, which is robust to fog and rain, the method reduces low-signal background tokens and uses augmentation that preserves physical properties, addressing both compute cost and generalization.

arXiv 2611.07613 · Upvotes 5 · 2026-11-05 · PDF · Code

Compared central command, auction-based consensus and independent agents in a simplified perimeter defense simulation, showing that the consensus team allowed fewer intrusions (15%) than central command (32%).

arXiv 2610.41044 · 2026-10-29 · PDF