Papers

• 📄
  Clean Engineering, Unstable Measurement: A Preregistered Reliability Failure of Black-Box LLM Observers on Shared Endpoints arXiv cs.AI

  This preregistered study is a bombshell: LLM-as-judge evaluations on shared endpoints show Spearman correlation of only 0.400 on repeat attempts, fund

• 📄
  Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training arXiv

  This paper tackles a deceptively simple question: when should an LLM actually reuse its past experience during autonomous post-training, and when is t

• 📄
  Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration arXiv

  Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration fixes overconfidence in models without ruining their

• 📄
  CorporateBench: Large-Scale Q&A Benchmarking with Temporal Knowledge Bases arXiv cs.AI

  CorporateBench introduces the first large-scale, human-validated Q&A benchmark for enterprise LLMs, tackling the ‘synthetic data’ problem with 230K re

• 📄
  PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents arXiv

  A self-improving agent framework that learns from mistakes in real-time, enabling long-horizon tasks without human intervention.

• 📄
  ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize arXiv cs.AI

  ESPO kills prompt bloat by diagnosing error structures before optimizing — a systematic alternative to evolutionary prompt tinkering that actually kee

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

• 📄
  SWE-Prime: Fewer Trajectories, Better Performance arXiv cs.AI

  SWE-Prime proves that not all successful agent trajectories are equal—and filtering them leads to better fine-tuning outcomes.

• 📄
  LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics arXiv cs.AI

  LeVJEPA redefines video pretraining by learning representations from *temporal coherence* without collapse—no heuristics, no masking, just pure self-s

• 📄
  Successive Capacity Growth: Task-Complexity-Driven Width and Depth Expansion for Vision Transformer Encoders in JEPA World Models arXiv cs.AI

  Successive Capacity Growth (SCG) turns Vision Transformers into dynamic learners that grow their brains—width and depth—as tasks get harder.

• 📄
  Stochastic Estimation of Transduced Language Models arXiv cs.CL

  This paper shows how to compute probabilities for transduced language models (TLMs) without drowning in an exponential sea of source strings—turning a

• 📄
  Boosting LLM Exploration via Weak-Model Guidance in RLVR arXiv cs.CL

  This paper turns up the entropy in RLVR by letting a weak model guide exploration—keeping LLM reasoning diverse even when rewards are punishingly stri

• 📄
  CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes arXiv

  CritICL turns failure into fuel: it uses small model mistakes at inference time to supercharge large language models, bridging the gap between weak an

• 📄
  Beyond F1: Evaluating Coverage and Failure Recovery in AI Model Security Scanners arXiv cs.AI

  This paper tears down the myth that AI security scanners are as reliable as we think—exposing critical blind spots in coverage and failure recovery th

• 📄
  WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution arXiv cs.AI

  WikiSkill turns agent experience into a Wikipedia-style knowledge graph that lets AI agents evolve skills like humans—by reusing and refining what the

• 📄
  VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement arXiv

  This paper turns physical reasoning into a *live* debugging session for AI agents—letting them test, break, and refine their own world models in real

• 📄
  Select, Compress, Reinvest: A Controlled Study of Visual-Token Allocation in Long-Video MLLMs arXiv

  A token‑budget playbook that lets multimodal LLMs actually watch hours‑long videos without blowing up the GPU.

• 📄
  Locked at the Entrance, Open Inside: Where RLVR Narrows the Solution Space arXiv

  This paper exposes a blind spot in LLM evaluation: current benchmarks miss the *solution structure* gap that makes models fail in the real world.

• 📄
  Using Grounded Theory for Agent Behavior Analysis at Scale arXiv

  A scalable grounded‑theory pipeline that turns massive logs of LLM‑agent actions into actionable behavioral taxonomies.

• 📄
  Compile by Training: Turning Natural-Language Specifications into Local Neural Functions arXiv

  Compile by Training turns natural-language specs into neural functions—bridging the gap between human intent and machine execution with a single train

• 📄
  PACE: Towards Surfacing Hidden Conflicts in User Requests arXiv

  PACE introduces a systematic framework for detecting hidden contradictions in user requests — a critical capability for any agent or dialogue system t

• 📄
  RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution arXiv cs.AI

  RedEvoAgent turns red-teaming LLMs into adaptive, self-improving attackers that evolve jailbreak tactics through experience—critical for hardening age