Papers

• Paper · arXiv · Sep 27, 2026Coding
  Opera: A Verbal Critic Framework for Long-horizon Coding Agents arXiv

  Opera tracks a critic’s corrections until an issue is resolved, separating compliance from resolution.

• Paper · arXiv · Oct 2, 2026Self Calibration
  Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction-Reality Gaps arXiv

  Prospective Hindsight (PH) is a new self-calibrating training principle for reinforcement learning that uses prediction-reality gaps as a signal to improve both task performance and calibration.

• Paper · arXiv · Oct 9, 2026Distillation
  Branching On-Policy Distillation arXiv

  Branching student-generated trajectories at uncertain decisions turns on-policy distillation from a narrow sampling trick into a coverage engine — and it pays for itself by reusing prefixes.

• Paper · arXiv · Oct 17, 2026Prompt Optimization
  IF-MAP: Influence-Function-Inspired Agent Prioritization in Multi-Agent Systems arXiv

  Introduces IF-MAP, a black-box method to identify which agents in a multi-agent system provide the highest system-level gain when improved.

• Paper · arXiv · Oct 9, 2026Language Modeling
  Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation arXiv

  Coarse-to-fine generation improves speed against the tested diffusion baselines, with quality gains or parity depending on the task.

• Paper · arXiv · Oct 17, 2026Agentic AI
  Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization arXiv

  Presents ID-TAP, a framework for adapting LLM agent personas at test-time using a dueling-bandit approach based on binary user comparisons.

Shelves rotate as new material clears the bar. These kept their pages and their analysis.

• Paper · arXiv · Sep 27, 2026Alignment
  Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures arXiv

  Jev turns a single LLM call into a calibrated, zero-shot detector for ten common alignment failures.

• Paper · arXiv · Sep 27, 2026LLM Fine Tuning
  Rufus-Air: An Open LLM Post-Training Recipe arXiv

  Rufus-Air delivers an open, reproducible eight-stage post-training pipeline that pushes GLM-4.5-Air beyond its official release.

• Paper · arXiv · Sep 27, 2026Audio Visual Generation
  AV-GRPO: Modality-Anchored Decoupling Diffusion Reinforcement Learning for Joint Audio-Video Generation arXiv

  AV-GRPO reframes joint audio-video generation as decoupled unimodal RL, boosting fidelity and synchronization.

• Paper · arXiv · Sep 27, 2026Bimanual Manipulation
  DeltaWAM: Delta World Action Models for Bimanual Manipulation arXiv

  DeltaWAM slashes video-expert cost and boosts bimanual manipulation success by predicting visual deltas instead of full frames.

• Paper · arXiv · Sep 27, 2026Linear Superposition
  Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs arXiv

  LLMs secretly obey linear superposition, letting a single forward pass spawn two coherent continuations.

• Paper · arXiv · Sep 27, 2026Model Architecture
  Neural Spectral Capacity: Measuring and Designing Architectures from Network Specification Alone arXiv

  Neural Spectral Capacity gives a closed-form, architecture-only score that outperforms Params and FLOPs for picking the best Transformer or CNN under a budget.

• Paper · arXiv · Sep 27, 2026Agent Security
  AgentKernel: The Trust-Native Agentic Operating System arXiv

  AgentKernel proposes a security-first operating system layer that enforces identity, perception, cognition, and execution controls for LLM agents.

• Paper · arXiv · Sep 27, 2026Deep Search
  IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis arXiv

  IterSynth splits planning and synthesis into separate LLM roles and uses a summary state to keep deep-search agents focused and efficient.

• Paper · arXiv · Sep 27, 2026Spatial Audio
  OmniEcho: Spatial Audio Understanding for Embodied Agents arXiv

  OmniEcho proves that spatial audio can be a first-class signal for embodied agents, achieving near-vision-language navigation performance.

• Paper · arXiv · Oct 4, 2026Edge Computing
  Replacing Large Language Models with Jev Decision Models for Low-Latency Edge Service Orchestration arXiv

  Jev slashes edge-orchestration latency by up to two-thirds, offering a cheap, low-latency alternative to LLM-based admission decisions.

• Paper · arXiv · Oct 11, 2026Generative Models
  The Lattice of Transition Laws arXiv

  A lattice-based theory lets you predict the optimal decoding schedule for diffusion, autoregressive, and hybrid generators before you run a single step.

• Paper · arXiv · Oct 4, 2026Long Context Memory
  MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization arXiv

  MemFold compresses long-term user memory into a fixed-size soft vector bank, boosting personalization accuracy while staying within strict context budgets.

• Paper · arXiv · Oct 4, 2026Video Generation
  Honeycomb: Constant-Size Scene Memory Representation for Video World Models arXiv

  Honeycomb shows you can keep video world-model memory fixed in size without sacrificing long-horizon consistency.

• Paper · arXiv · Oct 11, 2026Multi Task RL
  A GPU-Parallel Framework for Heterogeneous Multi-Task Reinforcement Learning arXiv

  Hebero shows that a single GPU-parallel policy can master 40 diverse robot tasks with just 50 demos per task.

• Paper · arXiv · Sep 30, 2026Parallel Decoding
  GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed arXiv

  GroundAnything studies distinct autoregressive and diffusion grounding variants, with measured accuracy/latency tradeoffs.

• Paper · arXiv · Sep 30, 2026Physical Intelligence
  GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives arXiv

  GroundingPI evaluates a 4B visual-grounding backbone on grounding, manipulation and open-loop driving benchmarks.

• Paper · arXiv · Oct 9, 2026Software Engineering
  AgentSWE: Can Coding Agents Build the Agent You Actually Want? arXiv

  AgentSWE tests whether a built agent meets the requirement on held-out cases, beyond the builder’s demo.

• Paper · arXiv · Sep 27, 2026Gui Agents
  Relevance Does Not Imply Applicability: Experience Activation for Personal GUI Agents arXiv

  ExpActivator gates GUI history by the current screen and recurring intents by time and scenario evidence.

• Paper · hf-daily-papers · Sep 24, 2026Task And Motion Planning
  Coding Agents for Generalized Task and Motion Planning Problems

  Coding agents synthesize reusable planners that outperform supplied planners in the tested simulated tasks.

• Paper · hf-daily-papers · Sep 23, 2026Math Reasoning
  Learning to Discover Interesting Mathematics

  A proof-length objective guides theorem generation; its novelty measure is limited to the evaluated formal library.

• Paper · arXiv · Oct 13, 2026LLM Routing
  Choosing LLMs for Strategic Agent Interactions arXiv

  Against fixed opponents in five games, a simple bandit beat the tested reflection and embedding selectors on net gain, but still lost to the best fixed model.

• Paper · arXiv · Sep 30, 2026Discovery
  Initialization Improves LLM-Driven Discovery arXiv

  A simple parallel initialization step consistently boosts LLM-driven discovery.

• Paper · arXiv · Sep 26, 2026Agent Hierarchy
  X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization arXiv

  X-Tree extracts reusable action hierarchies and tests them in three agent-training settings.

• Paper · arXiv · Oct 13, 2026Vlms
  CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline arXiv

  Introduces CLASHarness, a framework for generating and benchmarking verifiable multi-agent conflicts in computer-use workflows.