Proposes ZPPO, a distillation method that keeps the teacher inside prompts rather than injecting teacher gradients, using binary- and negative-candidate prompts plus a prompt replay buffer to recover learning signal on hard examples; shows gains for small Qwen3.5 students across 31 multimodal benchmarks.
Proposes Monotonic Inference Policy Improvement (MIPI) and a two-step Monotonic Inference Policy Update (MIPU) to address training–inference probability mismatch in LLM reinforcement learning by constructing sampler-referenced candidate updates and accepting synchronized updates using an inference-gap proxy; shows improved reasoning accuracy and stability under FP8-quantized rollouts.
Trains a transformer-based graph encoder with RL-guided adaptive masking so retrieved subgraphs embed relationships that better align with frozen LLM text encoders, improving GraphRAG performance with non-parametric retrievers on GraphQA benchmarks.
AI Agent Papers2026
Evaluates how long-term memory in LLM agents amplifies sycophantic behavior and when memory should or should not influence decisions. Provides five targeted tasks, 1,550 standardized samples, an evaluation pipeline, and baseline adapters to test memory use, conflicts, scope, updates, and personalization.
Predicts per-request MoE expert footprints from prefill activations and routes decode requests to workers that maximize expert-locality, lowering decode latency by combining offline K-means partitioning with online locality-band routing and a KV-block–coindexed signature cache.
Compiles natural-language function specifications into compact, locally-executable neural programs (PAW) that run on a small frozen interpreter; a 4B compiler emits LoRA adapters for a 0.6B runtime to provide offline, low-memory fuzzy text functions.
AI Agent Papers2026
Introduces a bounded-memory, typed-retrieval contract for long-horizon LLM agents and evaluates it in Slay the Spire 2 — assembling per-decision prompts from five typed slots rather than appending raw transcripts. Key outputs include ablationable memory layers, 298 labeled trajectories, and reproducible analysis scripts.
AI Video Papers2026
Generates temporally grounded captions for dense multi-event videos by restructuring autoregressive token dependencies to enable lossless parallel decoding; introduces a latent global planning module and event-factorized parallel decoding to improve grounding accuracy and achieve large decoding speedups.
AI Agent Papers2026
Proposes SkillOpt-Lite, a minimal pipeline for optimizing LLM agent skills by treating rollout traces as filesystem files and applying trajectory exploration, consensus mining, and independent validation; integrates as a one-line VSCode Copilot command and reports cross-benchmark improvements that let smaller models sometimes outperform larger ones.
Provides a reusable skill suite for evidence-grounded research ideation: Paper-Search for multi-source literature retrieval, Scoop-Check for prior-art collision checking, and IdeaSpark for pattern-guided idea generation, evidence auditing, and idea-card rendering.
Transfers RL-induced policy shifts from a smaller 'weak' teacher to a stronger target by using the teacher's post-/pre-RL log-ratio as a dense implicit reward applied on the student's on-policy states. Enables reuse of RL supervision without running RL rollouts on the target, improving sample/time efficiency.
Introduces KronQ, a post-training quantization framework that incorporates gradient covariance via a Kronecker‑factored Hessian to guide input/output weight rotations and sensitivity-driven mixed-precision allocation. Demonstrates stable 2-bit weight-only quantization on LLaMA-3-70B (7.93 PPL).