The Flow has not summarised this story yet — read it at arXiv Computation and Language.
Hy‑MultiTurn is a Chinese benchmark designed to evaluate deep multi‑turn dialogue understanding over long interactions. It introduces six controlled evaluation modes—constraint memory, precise execution, constraint synthesis, object localization, action suppression, and reference resolution—across 209 tasks ranging from 12 to 76 turns, incorporating dialogue length, irrelevant distractions, and colloquial phrasing. Testing 22 state‑of‑the‑art models shows the benchmark is highly challenging, with even the best model meeting all criteria only 41.1% of the time and no model excelling in every mode.
arXiv:2607. 19361v1 Announce Type: cross Abstract: Most safety guardrails for large language models (LLMs) evaluate each prompt-response pair in isolation, which misses failures that arise only over a dialogue as benign turns compose into harm.
The paper introduces LOGIC (Logit‑Space Integration for Contextual Biasing), a new framework that injects contextual entity information directly into the decoding layer of Speech Large Language Models, bypassing the limitations of prompt‑based methods. LOGIC operates with constant‑time complexity regardless of the size of the entity list, and experiments with the Phi‑4‑MM model across 11 multilingual locales show an average 9% relative reduction in Entity WER while adding only a 0.30% increase in False Alarm Rate.
The paper introduces a decoupled data approach for the Neural Finite State Machine (NFSM) framework to improve full‑duplex dialogue. It serializes real human‑human spoken dialogues into FSM tapes using a rule‑based event‑guided transformation, while shaping semantics through human‑agent text dialogues. A Source‑Aware Calibrated (SAC) loss is proposed to balance state‑transition token distribution and align each data source with its strongest supervisory signal, leading to better turn‑taking performance without sacrificing semantic quality.
arXiv:2607. 22610v1 Announce Type: new Abstract: When a language model produces a response in a multi-turn conversation, which tokens from prior turns shaped that answer, and how did those dependencies propagate across prior turns?
The paper introduces SCALE-QA, a new QA benchmark that tests conversational memory in flat, unsegmented multi‑topic threads by requiring agents to infer which earlier episode supports a later task decision. The dataset contains 3,000 audited questions across ten domains, uses deterministic four‑way multiple‑choice grading, and includes a runtime builder for reproducibility. The authors also propose Temporal‑Semantic Interleaved Memory Reconstruction (TSIM), a hierarchical memory stack that segments turns into coherent episodes and indexes them with deterministic summaries and cluster‑routing views, achieving significant accuracy gains over strong RAG baselines and long‑context LLMs.