AI Agent Papers2026
Automates distillation of heterogeneous traces from a target person or role into versioned, inspectable skill packages for LLM agents — producing separate capability and bounded-behavior tracks that support natural-language corrections, rollback, and cross-host installation. Ships with an open system and a skills gallery.
Uses search-agent reading traces and tiered distractors to train LLMs for long-context, multi-hop reasoning, and introduces a rubric reward that supervises entity-level steps (applied only to correct finals). Improves evidence-grounded reasoning and resists reward hacking across 4B–30B models.
Proposes TrOPD, a method that restricts token-level on-policy distillation to regions where teacher supervision is reliable to stabilize training under teacher–student distribution mismatch. Adds outlier handling (clipping, masking, forward-KL) and off-policy guidance; shows consistent gains on math reasoning, code generation and general benchmarks.
Studies small trainable adapters (PEFT) used as persistent personal models on top of large foundation models, analyzing three scaling axes—Scale Up, Scale Down, Scale Out—and introducing MinT, an infrastructure for adapter identity, provenance, evaluation, and serving.
Analyzes how single-domain RL fine-tuning on LLMs induces cross-domain interference and shows this damage concentrates in a low-dimensional shared conflict subspace; proposes a local perturbation theory and short domain "refresh" procedures that selectively recover earlier domains with minimal collateral loss.
Studies when and how to combine visual future rollouts from world models with abstract reasoning in multimodal LLMs. Proposes PF-OPSD — a teacher-student distillation that uses ground-truth future videos during training — and evaluates on two human-verified benchmarks, improving accuracy ≈10% while improving robustness to noisy rollouts.
Learns fine-grained preferences over sub-trajectories to identify and penalize redundant steps in long chain-of-thoughts, letting models "fold" reasoning chains into concise paths; reports ~56% token reduction on DeepSeek-R1-Distill-Qwen-7B while keeping accuracy.
Generates repository-specific LoRA adapters via a hypernetwork to inject repo-level knowledge into code LMs with zero inference-time token overhead. Provides a Static snapshot mode and an Evo mode that updates adapters per commit; evaluated on the 604-repo RepoPeftBench.
Benchmark for evaluating proactive LLM mediators in realistic, multi-domain conflict scenarios by constructing cases from real disputes, probing five socio-cognitive adaptation axes, and using a topic-localized evaluator that achieves 0.82 alignment with human experts.
AI Agent Papers2026
Dynamic interactive benchmark that tests whether LLM agents can adaptively plan and re-plan when world and user constraints are progressively revealed. Built on 307 household tasks with a multi-turn protocol that exposes hidden constraints only after plan violations, emphasizing iterative revision and constraint inference.
Removes the subspace of frequent, uninformative tokens that LLMs inject into text embeddings via the model's unembedding matrix. EmbedFilter is a lightweight linear transform that refines LLM-derived embeddings to improve zero‑shot semantic retrieval, enable dimensionality reduction, and speed up indexing; code on GitHub.
Analyzes the parameter-space geometry of on-policy distillation (OPD) for LLM training, showing OPD updates affect fewer weights, avoid principal directions, and rapidly lock into a low-dimensional update subspace. Compares OPD with supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) and studies implications for optimization and objective mixing.