A message-intervention study of role-specialized QA pipelines found that faithful rationales passed from a reasoner to a verifier add almost no answer accuracy over no rationale, while corrupted ratio…
Split-conformal claim filtering raises the share of multi-hop RAG responses whose retained claims are fully supported to 95.80%-97.20% at a 95% target, up from 55.60%-76.03% without filtering, accordi…
A new arXiv paper (arXiv:2609.27035v1) introduces Reinforcement Learning with Decomposed Subtasks (RLDS), a method that replaces the scalar GRPO advantage with Subtask-Decomposed Advantage Estimation …
An open-source Python library called ChunkRank is being submitted to COLING 2027, and its developer is seeking an arXiv cs.CL endorsement and an ARR service contributor from the Hugging Face community…
A new arXiv paper (2609.18126v1) formulates agentic AI deployment as a workflow portfolio problem in which a firm jointly chooses run size and allocation across workflow types, reporting that portfoli…
Researchers introduced Mosaic, a training-free framework that formulates GraphRAG retrieval as a per-query control problem, with an LLM analyzer converting query-specific evidence requirements into a …
A case study of an LLM coding agent implementing a multi-component data system found five defects introduced during a single session, with retrieval-filtering trade-offs evaluated on the HotpotQA benc…
Researchers introduced MASkills, a continual learning framework that optimizes multi-agent LLM systems through agent skills, enabling skill libraries to evolve via refinement, induction, consolidation…
Amazon AGI researchers found that reinforcement learning fine-tuning with Group Relative Policy Optimization (GRPO) improves long-context model performance by up to 20 points on HotpotQA benchmarks an…
A new benchmark called RENDER, introduced in an arXiv paper (arXiv:2608.23568v1), shows that the format of reader-facing memory artifacts can swing LLM memory evaluation scores by up to 72.6 points. O…
Researchers propose FCPRAG, a fusion-controlled parametric retrieval-augmented generation framework that adds a lightweight controller for sample-level adapter fusion, improving multi-passage LoRA inj…
Researchers introduced TRIAD, a three-stage automated dataset generation approach for evaluating retrieval-augmented generation (RAG) systems on domain-specific knowledge bases, according to a paper o…
A new arXiv paper argues for ingest-time semantic compilation over query-time interpretation in RAG, reporting 85.2% correctness on compiled claims versus 72.5% for chunk-based RAG. Another study intr…
A new arXiv preprint (2608.19535v1) proposes telemetry-informed adaptive compression for retrieval-augmented generation (RAG) on edge devices, based on experiments on the NVIDIA Jetson AGX Thor with L…
Researchers propose IterCOMP, a training-free prompt compression framework that iteratively integrates evidence for multi-hop question answering, achieving substantial improvements in Exact Match and …
A new arXiv paper (2608.10694v1) proposes a cost-aware cross-tier transfer method for evolutionary optimization of LLM prompts and agentic programs, reducing search cost by 5.6-14x (up to 25-54x with …
A new paper on arXiv (2608.08237v1) proposes SAGE, a learned SLO-aware adaptive retrieval policy for production RAG systems that dynamically selects the number of passages per query. On Natural Questi…
A 2026 study by Daeyoung Roh and Donghee Han of 12,000 paired agent trajectories across HotpotQA, 2WikiMultiHopQA, and MuSiQue found that agentic RAG failures often stem from procedural errors: on MuS…
A study labeling 7,440 answers from four open-weight LLMs found that more than half of the hallucination labels were incorrect, and correcting them reordered the results. The author, Priyanshi Jain, r…
Researchers propose PI-Mem (Parallel-Iterative Memory), a mechanism that processes all chunks in parallel and iteratively refines a shared memory, enabling long-context reasoning up to 3.6 million tok…