The paper introduces RefineCut, an open‑weight planner that edits a typed video timeline by applying structured patches for clip selection, trimming, ordering, transitions, music alignment, and duration. A deterministic verifier checks each patch against an explicit constraint ledger, and the planner is trained via verifier‑replayed distillation and a second evolutionary stage (RefineCut‑Evo) that uses the verifier and a task rubric to generate high‑margin preference pairs. On the RefineCut‑Bench dataset, the 8‑billion‑parameter planner improves from a Video‑Editing Score of 0.620 to 0.924, matching or exceeding its frontier teachers in a closed verifier loop, and the gains transfer to other large models such as Llama‑3.1‑8B and GLM‑4‑9B.
arXiv:2610.02826v1 Announce Type: new Abstract: Successful trajectories on difficult tasks provide valuable supervision for model improvement, but specialized harnesses introduce interventions that m...
arXiv:2608. 02694v1 Announce Type: cross Abstract: Long-horizon video editing agents receive final-product feedback only after many interdependent decisions.
WorldGuide is a closed‑loop video world model that, given an initial image and a task goal, predicts an atomic action, generates the corresponding video clip, and uses that clip to decide the next action or terminate the task. It trains a Planner to predict the next action or completion from visual progress and an Executor to realize those actions, while a hierarchical visual memory tracks state over long horizons. The authors introduce WorldGuide‑Bench, a dataset of ~59K step‑annotated videos across 245 tasks, and show that WorldGuide outperforms the recent MiniMax‑H3 model on both WorldGuide‑Bench and VideoCraft‑Bench, highlighting the benefit of coupling planning with learned execution for goal‑directed procedural video generation.
The study demonstrates that governance records—structured logs linking task contracts, model attempts, verifier decisions, and outputs—can serve as effective supervision for bounded AI models. Using a verifier-selected self‑training approach, the authors show that a Qwen3‑14B model trained on plans accepted by an independent VAL verifier achieved significant gains in plan acceptance across numerous PlanBench replanning cases, outperforming other selection strategies. The results highlight the feasibility of one‑shot execution and cumulative learning without relying on oracle targets or stronger teachers.
The paper investigates how fine‑grained credit assignment affects learning in multimodal agentic reinforcement learning, using video editing as a testbed. Four movie‑trailer editing tasks—Content Selection, Narrative Sequencing, Rough Cut, and Fine Cut—are introduced to represent real‑world editing workflows with varying complexity and horizon lengths. The study finds that decaying process rewards significantly boost performance on the long‑horizon Fine Cut task, while offering modest gains on shorter tasks; however, fixed process‑reward weighting can lead to reward‑hacking and degraded outcomes, and on‑policy self‑distillation provides limited benefit.
StateWise is a method that trains a single large language model to maintain and update its own constraint state during multi‑turn dialogue. By fine‑tuning the model to summarize conversations into a typed state with tombstones for revoked constraints and then self‑training on trajectories that pass exact round‑trip checks, it reduces relapse and improves conversation success. The approach outperforms a compiled‑ledger baseline on relapse without requiring inference‑time scaffolding, and it is evaluated on two backbones with held‑out constraint families and unseen phrasings.
Translating natural-language planning intent into verified plans is a longstanding challenge: people communicate goals in language, while classical planners require formal PDDL specifications. Recent agentic frameworks bridge this gap by orchestrating a pool of specialized repair agents inside a verifier-checked refinement loop, but the orchestrator at the centre is itself a prompted frontier LLM, paying a frontier-LLM API call at every refinement step.
arXiv:2606. 04970v1 Announce Type: cross Abstract: We envision a proactive multi-modal assistant system which gives users real-time step-by-step guidance on a procedural task, autonomously deciding \textit{when} to interrupt, and \textit{how} to coach.
PROOF-Gen is a method that improves distillation of tool‑calling models by recovering successful trajectories from teacher failures. It uses per‑scenario prompt optimization to generate corrective guidance that steers the teacher to a passing trajectory, then removes this guidance before training so the student learns from clean demonstrations. On τ2‑bench, PROOF-Gen recovers 93% of failed scenarios, boosting Qwen3‑4B‑Instruct‑2507’s Pass^1 from 0.132 to 0.529 and improving Gemma 4 E4B‑it by 7.2pp on BFCL v4 multi‑turn, while also raising deployed on‑device model performance by up to 5.0pp across response‑quality metrics.
ACE is a self‑correcting agentic canvas editor that operates on a hierarchical scene‑graph rather than flat document formats, enabling reliable multi‑slide presentation automation. It pairs a presentation‑specialized action space of 98 tools with CARE, a content‑aware router that reduces input tokens by about 89%, and a ground‑truth‑free instruction‑following judge that feeds natural‑language critiques back into the agent for self‑correction. In benchmarks, ACE outperforms a comparable agentic HTML pipeline on instruction following (4.23 vs. 3.81), runs 1.75× faster, costs 44% less, and is preferred by 58.7% of blind raters, with 81% favoring the self‑corrected output.
The paper proposes a reward‑density principle for language‑model post‑training: sparse sequence‑level rewards are best for exploring and discovering new behaviors in a strong teacher, while dense token‑level teacher supervision is optimal for compressing that behavior into a smaller deployment model. The authors validate this rule with a four‑stage workflow—teacher RL, forward‑KL warmup, on‑policy distillation, and optional post‑bridge student RL—showing that an RL‑improved 8B teacher distilled through a dense bridge outperforms direct GRPO on a 1.7B student across math and code benchmarks. Ablation studies confirm each stage’s contribution, and the approach generalizes to Llama and Qwen models, offering a practical guideline for allocating scarce labeled data.
arXiv:2609.38142v1 Announce Type: new Abstract: A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advi...