Top AI Papers: Week of November 2, 2026
This week's papers highlight a strong focus on practical applications and system-level challenges in AI. Themes include improving human-AI collaboration, enhancing the robustness and trustworthiness of AI systems, and developing more efficient methods for data handling and model training, especially for long-context LLMs.
10 picks from every AI/ML paper we summarized this week, each with the reason it made the cut. Selected and annotated by ArXiv TLDR with AI assistance.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.
- #1cs.IR, cs.CY, cs.SI
Recommendation Feedback Loops Shape Urban Income Segregation and Venue Visitation
Why it made the list: Offers a surprising and important finding: recommender systems can exacerbate urban income segregation and cause venue disappearance, highlighting societal impacts.
TL;DR: This paper reveals how recommender systems create feedback loops that increase urban income segregation and cause venue disappearance.
- Simulates long-term effects of recommender systems on urban mobility and social segregation.
- Finds that recommender systems, especially more accurate ones, increase income segregation by 15-21%.
- Shows that retraining algorithms leads to at least 20% of venues losing all visits.
- #2cs.HC, cs.CL
Hybrid Computer Use: A Unified Programmatic, Semantic, and Visual Interface for General Agents
Why it made the list: HCU provides a unified programmatic, semantic, and visual interface for general AI agents, a practical step towards more capable and versatile AI computer use.
TL;DR: HCU is a unified interface for AI agents combining programmatic, semantic, and visual interactions to improve computer-use task performance.
- Introduces Hybrid Computer Use (HCU), integrating code, accessibility (AX), and GUI interactions.
- Enables efficient AX interaction with compact observations, batched actions, and AX diffs.
- Includes an HCU skill to guide agents in selecting and coordinating interaction channels.
- #3cs.AI, cs.CV
ReuScope: Auditing Data Reuse and Evaluation Sensitivity in Capsule Vision 2024
Why it made the list: ReuScope reveals an issue in ML: data reuse in benchmarks can lead to inflated performance, with models showing drops when reused data is removed.
TL;DR: ReuScope audits data reuse and evaluation sensitivity in machine learning models, revealing significant performance drops when reused data is removed.
- Introduces ReuScope, a protocol for auditing data reuse and evaluation sensitivity.
- Identifies byte-identical files and shared source recordings crossing training-validation boundaries.
- Removing reused training data drops public-validation accuracy from 82.5% to 61.2%.
- #4cs.IR, cs.AI, cs.CL
Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses
Why it made the list: Introduces 'knowledge compilation' as a layer for agentic RAG systems, optimizing context and reducing token usage without sacrificing accuracy.
TL;DR: This paper introduces 'knowledge compilation' as a crucial layer for agentic RAG systems, optimizing context and reducing token usage without sacrificing accuracy.
- Proposes 'knowledge compilation' to enrich documents with structured context (e.g., navigation hierarchies, entity graphs).
- Demonstrates significant reduction in query tokens (91%) and latency (from 140s to 30s) using compiled context.
- Introduces RAGFlow, an open-source pipeline and harness for disciplined context management in agentic systems.
- #5cs.IR
Search When Needed, Proceed When Reliable: Metacognitive Reliability Supervision for Efficient Agentic Search
Why it made the list: A metacognitive framework trains LLMs to efficiently decide when to search for information based on the reliability of their intermediate reasoning states, improving agent efficiency.
TL;DR: A metacognitive framework trains LLMs to efficiently decide when to search for information based on the reliability of their intermediate reasoning states.
- Proposes a metacognitive framework for LLMs to supervise search decisions based on intermediate reasoning reliability.
- Estimates current-state reliability to decide if proceeding without search is viable.
- Measures how retrieved evidence changes estimated reliability to score search actions.
- #6cs.LG, cs.CY, cs.DS
Differentially Private Synthetic Data from Sliced Chebyshev Moments
Why it made the list: SPLICE is a new differentially private synthetic data algorithm that offers strong global accuracy guarantees, even in high dimensions, useful for privacy-preserving data sharing.
TL;DR: SPLICE is a new differentially private synthetic data algorithm that offers strong global accuracy guarantees, even in high dimensions.
- Introduces SPLICE, a DP synthetic data algorithm with global distributional guarantees.
- Achieves a Sliced Wasserstein-1 distance of O(1/√n), independent of data dimension.
- Uses noisy sliced Chebyshev polynomial moments for high-dimensional data synthesis.
- #7cs.CR, cs.CL
How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call Traces
Why it made the list: Reveals that LLM agent tool-call traces can leak user privacy, enabling profiling even without conversation access, a privacy risk to address.
TL;DR: This paper reveals that LLM agent tool-call traces can leak significant user privacy, enabling profiling even without conversation access.
- Introduces PrivacyTrace, a benchmark with 1,000 synthetic user profiles and 4,000 agent trajectories.
- Shows LLMs can infer user profiles from tool-call traces with a 26.25% success rate, outperforming guessing.
- Identifies tool arguments and returned results as primary privacy leakage points.
- #8cs.CR, cs.AI
SecPO: Principled Adversarial Training for Prompt Injection Security
Why it made the list: SecPO introduces a principled adversarial training method and loss function to improve LLM security against prompt injection attacks, a practical defense.
TL;DR: SecPO introduces a principled adversarial training method and loss function to significantly improve LLM security against prompt injection attacks.
- Proposes SecPO loss, a theoretically principled approach using an approximated injection-invariant oracle.
- Implements an efficient adversarial training pipeline to synthesize strong prompt injection attacks.
- Reduces attack success rates by 1.5-14.5x (AT) and a further 2.4-4.6x (SecPO loss) over Meta-SecAlign.
- #9cs.AI
Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention
Why it made the list: Latent Sparse Attention (LSA) enables efficient, accurate ultra-long context for LLMs by clustering keys in a learned latent space, improving LLM scalability.
TL;DR: Latent Sparse Attention (LSA) enables efficient, accurate ultra-long context for LLMs by clustering keys in a learned latent space.
- Introduces Latent Sparse Attention (LSA) for subquadratic ultra-long context attention.
- Hierarchically clusters keys in a learned latent space, queries score clusters for efficiency.
- Distilled-LSA (D-LSA) speeds up prefill 22.7x and decode 4.6x on GLM-5.3-FLASH at 10M tokens.
- #10cs.LG, cs.AI, stat.ML
Chronos-2: A foundation model for multivariate time series forecasting
Why it made the list: Chronos-2 is a new foundation model for multivariate time series forecasting, offering zero-shot probabilistic predictions across diverse scientific domains, a useful tool.
TL;DR: Chronos-2 is a new foundation model for multivariate time series forecasting, offering zero-shot probabilistic predictions across diverse scientific domains.
- Models interactions between multivariate time series and heterogeneous covariates for improved accuracy.
- Achieves state-of-the-art accuracy, improving skill scores by 5.2% over strong baselines.
- Generates forecasts for 100 time series in under one second on a single GPU, significantly faster than alternatives.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.