Weekly notable papers by field

A decision model that thinks more deeply by running its layers repeatedly instead of generating text, allowing a single checkpoint to spend different amounts of compute on easy and hard questions.

arXiv 2610.20951 · Upvotes 106 · 2026-10-14 · PDF · Code

A parametric memory architecture that injects learned memory directly into intermediate hidden states, aiming to reduce both forgetting during continued pretraining and the cost of real-time retrieval.

arXiv 2610.25100 · Upvotes 50 · 2026-10-18 · PDF

The authors scale Lafa, a continuous flow language model that generates in parallel block by block, up to 8B. It trains on little data by building on existing autoregressive models, and prefix caching makes inference fast.

arXiv 2610.28559 · Upvotes 23 · 2026-10-20 · PDF · Code

An open image generation system that combines a 5B DiT pretrained from scratch with agentic, intent-aware prompt enhancement. It renders long text in two languages and offers precise layout control.

arXiv 2610.23374 · Upvotes 103 · 2026-10-16 · PDF · Code

An interactive video generation strategy that appends noisy latents to extend the generation horizon and caches predictions for reuse, aiming for both the quality of a long horizon and the responsiveness of a short one.

arXiv 2610.20556 · Upvotes 85 · 2026-10-14 · PDF · Code

To reduce the error accumulation that occurs when diffusion models take their own predictions as input, the authors make a correction step that fixes the base model's predictions into a separate training task.

arXiv 2610.14136 · Upvotes 50 · 2026-10-09 · PDF · Code

Applies Shapley values to channels at each location to build a full location × channel attribution matrix, explaining at once where and which channels a model looks at in multispectral images.

arXiv 2610.21114 · Upvotes 680 · 2026-10-15 · PDF · Code

The authors propose VPM, a world model that takes multi-camera video and language instructions and jointly predicts future video from arbitrary viewpoints and 3D trajectories of scene points.

arXiv 2610.28594 · Upvotes 52 · 2026-10-20 · PDF · Code

A feed-forward 3DGS framework that drops uniform per-pixel Gaussian placement and instead places Gaussians adaptively based on content, reducing redundancy and better preserving planar structures.

arXiv 2610.13177 · Upvotes 34 · 2026-10-09 · PDF · Code

Shows across 22 benchmarks that increasing the inference budget alone yields only small gains, and proposes an open inference harness that combines diverse tools with persistent multimodal memory.

arXiv 2610.26860 · Upvotes 120 · 2026-10-19 · PDF · Code

For code generation where the rendered result matters, such as charts, web pages and animations, it tackles the data quantity and quality problems holding back open models through data synthesis based on rubric decomposition.

arXiv 2610.20538 · Upvotes 71 · 2026-10-14 · PDF · Code

A 4B multimodal model based on Qwen3.5 predicts masks from its own visual representations without a separate segmentation encoder, and reasons while pointing to regions.

arXiv 2610.23279 · Upvotes 49 · 2026-10-16 · PDF · Code

A runtime for long-horizon agents that organizes context around reusable work objects instead of rereading the interaction history every time.

arXiv 2610.18784 · Upvotes 124 · 2026-10-13 · PDF · Code

Synthesizes executable environments and tasks at scale, supporting GUIs, MCP services and CLIs together, to train computer-use agents without the security risks of real systems.

arXiv 2610.26828 · Upvotes 100 · 2026-10-19 · PDF · Code

A mobile agent released with 35 apps that mimic commercial apps along with all training data and recipes, providing about 12,000 trajectories for supervised learning and verifiable reinforcement learning data.

arXiv 2610.14157 · Upvotes 60 · 2026-10-09 · PDF · Code

Deployed robots pool their execution experience to improve a shared policy, scaling learning across the whole robot fleet without relying on expert corrections.

arXiv 2610.25736 · Upvotes 137 · 2026-10-19 · PDF

Learns from first-person internet videos how the view changes when moving toward a goal and uses this to generate robot actions, sidestepping costly real-robot data and the sim-to-real gap.

arXiv 2610.25690 · Upvotes 91 · 2026-10-19 · PDF · Code

A framework for directly post-training robot planners that previously relied on closed models, supporting different harnesses such as VLA skill calls and control code generation.

arXiv 2610.28589 · Upvotes 48 · 2026-10-20 · PDF · Code

A reinforcement learning technique for long-horizon agents that assigns rewards per turn and reduces policy staleness, without a critic, step-level labels or an external teacher.

arXiv 2610.23181 · Upvotes 57 · 2026-10-16 · PDF

Proposes 'harness scaling', which extends training from a single fixed harness to diverse execution configurations, as a new scaling axis for agent reinforcement learning.

arXiv 2610.29458 · Upvotes 27 · 2026-10-21 · PDF · Code

Defines multi-level process rewards spanning tool calls, task steps and final outcomes in Lean4 to strengthen agent reinforcement learning for SWE and terminal tasks.

arXiv 2610.22448 · Upvotes 24 · 2026-10-15 · PDF · Code

In agent tasks, the draft model's acceptance length drops by up to 80%. The authors propose speculative decoding that does not use the long past context responsible for this at all.

arXiv 2610.23787 · Upvotes 35 · 2026-10-16 · PDF · Code

Replaces MoE experts kept separately in each layer with Meta Experts shared across layers, and compares from several angles how well this reduces parameter redundancy.

arXiv 2610.27598 · Upvotes 31 · 2026-10-20 · PDF

Uses a quantized draft model to predict upcoming experts several steps ahead, reducing the PCIe bottleneck in moving MoE experts offloaded to host memory.

arXiv 2511.14102 · Upvotes 27 · 2025-11-18 · PDF

Arguing that GUI agent safety must be judged on screen transitions rather than single screens, the authors propose a world model that predicts the next screen in the observation encoding space to catch attacks that reuse legitimate screens.

arXiv 2610.11942 · Upvotes 188 · 2026-10-08 · PDF · Code

Measures the 'mosaic effect', in which combining multiple documents reveals sensitive conclusions. Even the best model exposed a forbidden combination in 7.08% of its answers to benign queries.

arXiv 2610.30418 · Upvotes 23 · 2026-10-21 · PDF

Implements, through two-stage fine-tuning, a speech model backdoor in which a specific speaker's natural voice itself acts as the trigger, without tampering with inputs at inference time.

arXiv 2610.27210 · 2026-10-19 · PDF

A benchmark of 50 expert-verified physics research tasks that evaluates whether AI agents can design investigations themselves and reach conclusions grounded in evidence.

arXiv 2610.21767 · Upvotes 54 · 2026-10-15 · PDF · Code

An autonomous research agent that continues research across sessions, producing 120 record-breaking results, including narrowing the bounds on an n-queens constant by about 84 times.

arXiv 2610.24944 · Upvotes 24 · 2026-10-18 · PDF · Code

A long-running testbed that measures how much agents improve on the best published models across 20 scientific ML benchmarks in 11 fields. The agents surpassed the previous best results on every benchmark.

arXiv 2610.28763 · Upvotes 29 · 2026-10-20 · PDF · Code

A benchmark of 48 tasks and 40,033 items that diagnoses decision models used for routing, guardrails and output verification using contrast groups and reliability criteria.

arXiv 2610.28522 · Upvotes 74 · 2026-10-20 · PDF · Code

Uses the open-source framework LASAGNA to test 450 RAG configurations on 8 benchmarks, checking whether the effects of individual techniques hold when they are combined or scaled up.

arXiv 2610.29722 · Upvotes 54 · 2026-10-21 · PDF · Code

Points out that similarity-only data selection is biased toward data clustered around development samples, and rebalances it with a von Mises-Fisher kernel.

arXiv 2610.27142 · 2026-10-19 · PDF

Combines fuzzy, transformation and semantic joins with LLMs but calls the LLM only selectively, maximizing recall while meeting a precision target within a fixed cost.

arXiv 2610.25178 · 2026-10-18 · PDF

Documents table relationships in a data lake offline with little overhead, reducing the cost and latency agents incur from exploring tables on every query.

arXiv 2610.25131 · 2026-10-18 · PDF

A training-free multi-agent framework that runs multiple interpretations of an ambiguous question as SQL to check whether the results actually differ, and generates clarifying questions based on the differences in the data.

arXiv 2610.24248 · 2026-10-16 · PDF

It uses execution-based evaluation to test how well agents can fix problems across the full lifecycle of real ML repositories, including data pipelines, serving, monitoring and CI/CD.

arXiv 2610.23506 · 2026-10-16 · PDF

By replaying real production requests, the paper analyzes why turning on a remote KV cache tier raised p99 time to first token for 16k-token requests from 3.96 seconds to 17.3 seconds.

arXiv 2610.21500 · 2026-10-15 · PDF

In multi-turn conversations and agent workloads, KV cache writes can shorten SSD lifetime from years to months. The paper reduces this problem while preserving the cache hit rate.

arXiv 2610.30791 · 2026-10-22 · PDF

Based on the finding that items in a user's history need very different amounts of computation, the paper replaces the common practice in sequential recommendation of spending the same computation on every item with per-item depth allocation.

arXiv 2610.30128 · Upvotes 41 · 2026-10-21 · PDF · Code

LinkedIn proposes intent-based people recommendations that reflect complementary relationships, such as a job seeker wanting to connect with recruiters rather than with similar peers.

arXiv 2610.30941 · 2026-10-22 · PDF

Using seven BEIR datasets, the paper analyzes the losses that hybrid search, which combines BM25 and dense retrieval, incurs on individual queries, and treats "not fusing" as a legitimate option.

arXiv 2610.30612 · 2026-10-22 · PDF

MutationProjector, pretrained on genomic mutations from about 30,000 tumors, matched or exceeded state-of-the-art performance in predicting resistance to immunotherapy and chemotherapy, and also identified new biomarkers.

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

A public deep learning model was prospectively validated in shadow mode on 826 consecutive patients at a tertiary hospital in Nepal, confirming its practical applicability in resource-limited healthcare settings.

BMJ open · Europe PMC · 2026-10-01 · PDF

An LLM automatically extracts ER, PR and HER2 status and other data from 532 TCGA breast cancer pathology reports to classify triple-negative breast cancer.

JMIR formative research · Europe PMC · 2026-10-23 · PDF

LLMs carry out the full quant research process, from forming hypotheses to finalizing strategies, on real China A-share data, and are evaluated on undisclosed 2024 performance.

arXiv 2610.30113 · 2026-10-21 · PDF

A benchmark of "information injection" attacks that manipulate the news an agent trusts, evaluating about 10,000 trading decisions across 5 injection channels and 7 event types.

arXiv 2610.26438 · 2026-10-19 · PDF

For 26 options trading agents that have actually been released, the study is the first to measure which safeguards they have and the worst-case losses that can arise along the order path.

arXiv 2610.24289 · 2026-10-16 · PDF

To handle time delays between process stages that change from moment to moment, the method directly estimates the delay distribution and uses it to pass information between stages to predict final quality.

arXiv 2610.29220 · 2026-10-21 · PDF

Physics simulation and polarization rendering generate large numbers of accurately labeled carbon-fiber SMC images, addressing the shortage of training data for inline quality inspection models.

arXiv 2610.28050 · 2026-10-20 · PDF

To run predictive maintenance models for rotating machinery on a range of industrial edge hardware, the study compares how well four compression techniques work in practice.

arXiv 2610.30228 · 2026-10-21 · PDF

The study removes the correct holding from some CaseHOLD questions and compares whether decision models and generative LLMs properly abstain when no correct answer is available.

arXiv 2610.29961 · 2026-10-21 · PDF

With 916 items asking about monetary amounts, rates and age thresholds in Norwegian statutes as of specific dates, it evaluates LLMs' accuracy on legal figures by classifying answers as correct, outdated or wrong.

arXiv 2610.29207 · 2026-10-21 · PDF

The study analyzes whether internal representations corresponding to all 26 CATO legal factors, used in comparative case-based reasoning, form inside LLMs.

arXiv 2610.28799 · 2026-10-20 · PDF

A family of speech recognition models supporting 30 languages and 16 Chinese dialects, extended to speaker-attributed transcription and general audio understanding.

arXiv 2610.28711 · Upvotes 37 · 2026-10-20 · PDF

A voice agent technique that runs reasoning periodically and executes tools in advance while the user is speaking, reducing the wait after the user finishes.

arXiv 2610.26233 · Upvotes 24 · 2026-10-19 · PDF

Tests whether a text refusal direction extracted from the LLM backbone also works for images and video, exploring a way to improve safety without harmful multimodal data, which is hard to collect.

arXiv 2606.31876 · Upvotes 24 · 2026-06-30 · PDF · Code