ArXiv TLDR
Weekly ยท 2026-W46 ยท week of November 9, 2026

Top AI Papers: Week of November 9, 2026

This week highlights a strong push toward structural efficiency and safety in deployed systems. Key themes include specialized model quantization, principled agentic reasoning controls, and robust evaluation benchmarks for real-world interactions.

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.

๐Ÿ“ฌ Weekly AI Paper Digest

Get the top 10 AI/ML arXiv papers from the week โ€” summarized, scored, and delivered to your inbox every Monday.

  1. #1cs.AI, cs.AR, cs.CL

    BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design

    Why it made the list: BitDecoding-FP4 enables accurate and fast native FP4 KV cache decoding via algorithm-system co-design for efficient LLM serving.

    TL;DR: BitDecoding-FP4 enables accurate and fast native FP4 KV cache decoding via algorithm-system co-design.

    • Dual-Residual Query Quantization cuts query logit error from 60% to 1%.
    • Underflow-Aware Probability Scaling stops E4M3 from zeroing long-context probabilities.
    • Fused transposed softmax and paged KV support deliver 2.4-3.0x higher throughput.
    Read full summary โ†’
  2. #2cs.LG, cs.CL

    RidgeFill: Training-Free Late-Layer Cache Fill for Faster Prefill

    Why it made the list: RidgeFill accelerates LLM prefill without retraining by predicting late-layer KV caches using closed-form ridge regression.

    TL;DR: RidgeFill accelerates LLM prefill without retraining by predicting late-layer KV caches using closed-form ridge regression.

    • Retrofits upper-layer KV cache prediction onto standard models using a single pass of calibration text.
    • Achieves 1.25x to 1.60x faster time-to-first-token across long contexts without dropping any tokens.
    • Maintains competitive accuracy on needle-in-a-haystack and long-context benchmarks like RULER.
    Read full summary โ†’
  3. #3cs.AI

    Test-Time Compute Scales Better with Parallel Agent Teams

    Why it made the list: Parallel agent teams scale test-time compute better than independent agents or large collaborating teams in empirical tests.

    TL;DR: Parallel agent teams scale test-time compute better than independent agents or large collaborating teams.

    • Introduces parallel agent teams that sample many small collaborative groups.
    • Beats standard best-of-N sampling on 10 of 12 long-horizon technical tasks.
    • Balances individual mean performance with the benefits of group variance.
    Read full summary โ†’
  4. #4cs.HC

    GroundWork: Bridging Human and LLM Reasoning Through a Structured Argumentation Model

    Why it made the list: GroundWork provides a mixed-initiative structured argumentation model to bridge human and LLM reasoning states.

    TL;DR: GroundWork is a mixed-initiative system that makes argument structures a shared, editable state for humans and LLMs.

    • Operationalizes the Toulmin model as a typed graph for collaborative editing.
    • Provides coordinated corpus views and an interactive argument workbench.
    • Evaluated via expert studies and use cases in law, medicine, and intelligence.
    Read full summary โ†’
  5. #5q-bio.NC

    Path-dependent triage: Early judgments lead conversational AI chatbots to miss mental-health emergencies

    Why it made the list: Chatbots miss one in five mental health emergencies in multi-turn conversations due to early triage bias.

    TL;DR: Chatbots miss one in five mental health emergencies in multi-turn conversations due to early triage bias.

    • Evaluated 13 frontier AI chatbots using 112 simulated mental-health case files.
    • Chatbots missed 20.3% of emergencies in multi-turn chats versus 5.6% in single turns.
    • Early triage decisions anchored later judgments, preventing recovery as evidence grew.
    Read full summary โ†’
  6. #6cs.AI, cs.NE

    SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

    Why it made the list: SwarmEvolve accelerates AI-driven program discovery by using latent-space agent swarms communicating via KV caches.

    TL;DR: SwarmEvolve accelerates AI-driven program discovery by using latent-space agent swarms that communicate via KV caches instead of slow text generation.

    • Replaces text chains of thought with low-cost latent reasoning using KV caches.
    • Employs novelty filters to maintain population diversity via hidden state distances.
    • Achieves state-of-the-art results up to 10x faster than traditional LLM search.
    Read full summary โ†’
  7. #7cs.CV, cs.NE

    SpikePose: Lightweight Video Human Pose Estimation

    Why it made the list: SpikePose cuts arithmetic energy by 91.4% in video human pose estimation using spiking encoders and heatmap memory.

    TL;DR: SpikePose is a lightweight video human pose estimator that cuts arithmetic energy by 91.4% using spiking encoders and heatmap memory.

    • Pairs a single-step spiking encoder with motion-aligned heatmap memory (MAM).
    • Reduces estimated arithmetic energy by 91.4% compared to an architecture-matched ANN.
    • Cuts velocity and acceleration errors by 28.13% and 31.46% using MAM.
    Read full summary โ†’
  8. #8cs.CR, cs.AI, cs.CL

    MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory

    Why it made the list: MemoryDojo benchmarks memory poisoning vulnerabilities across various LLM agent memory architectures.

    TL;DR: MemoryDojo is a new benchmark measuring memory poisoning vulnerabilities across various LLM agent memory architectures.

    • Introduces MemoryDojo, a benchmark with 155 trajectories across four domains and five memory systems.
    • Demonstrates that all evaluated LLM memory architectures are susceptible to memory poisoning attacks.
    • Proposes a memory-aware poison optimization process to significantly increase attack efficacy.
    Read full summary โ†’
  9. #9cs.CV, cs.AI, cs.CR

    Breaking the Privacy Promise: Gradient Inversion Attacks on Federated Low-Rank Adapted Vision Transformers

    Why it made the list: Gradient inversion attacks successfully recover private data from LoRA-adapted Vision Transformers in federated learning.

    TL;DR: Discover how gradient inversion attacks successfully recover private data from LoRA-adapted Vision Transformers in federated learning.

    • Reveals private data leakage risks in LoRA-adapted Vision Transformers during federated learning.
    • Identifies two key reconstruction bottlenecks: reduced LoRA gradient information and missing spatial cues.
    • Proposes a spatially anchored generative inversion framework using trainable positional embeddings.
    Read full summary โ†’
  10. #10cs.LG, cs.AI

    Scaling Graph Similarity Learning with Graph Neural Networks

    Why it made the list: FiLSim is a scalable GNN framework that computes graph similarity for large graphs with over 150,000 nodes efficiently.

    TL;DR: FiLSim is a scalable GNN framework that computes graph similarity for large graphs with over 150,000 nodes efficiently.

    • Proposes FiLSim, a scalable GNN framework for graph similarity learning.
    • Introduces a sequential graph-editing strategy to generate large datasets with reliable GED labels.
    • Scales to graphs with over 150,000 nodes where existing baselines fail.
    Read full summary โ†’

๐Ÿ“ฌ Weekly AI Paper Digest

Get the top 10 AI/ML arXiv papers from the week โ€” summarized, scored, and delivered to your inbox every Monday.