Tag
The paper introduces PersonaChatBench, a framework that evaluates 20 language models across six domains by pairing fixed tasks with diverse simulated users guided by private persona cards, measuring task outcomes, persona need fulfillment, and simulator fidelity. Results show that strong task performance does not consistently translate into satisfying individual user needs, with model strengths varying across personas and domains.
SVPino highlights Voice Arena AI's Joy, a 100,000-hour dual-channel conversational voice dataset with separate per-speaker tracks, overlapping speech, interruptions, pauses, and backchannels, designed to teach voice models conversational timing.
提出 Full-Duplex Long-Horizon Benchmark（FD-LHB），发现现有全双工语音模型在长时间对话中会因上下文耗尽或重复、静音、乱码而崩溃，并通过 Initialization-Anchor Pinning 和 Duplex Repetition Penalty 两种免重训练修复方法显著缓解对话退化。
The paper introduces MOSAIC, an omni-modal benchmark of 923 questions over persona-grounded multi-session histories spanning text, image, video, and audio, designed to diagnose where conversational agents lose evidence across write, retrieve, deliver, and read stages rather than just producing a single leaderboard score.
BotsCrew 分享了为 EA SPORTS FC 26 制作的极简 WhatsApp 机器人营销案例：仅用一条托尼·克罗斯的真实语音留言和一个 CTA，在 4 天内实现 1.5 万次对话、约 40% 转化率和 23% 销售提升。文章重点讲述了负载测试、WhatsApp 封号风险应对、流量突发扩容与多渠道追踪等工程挑战。
We present SmartSalesBench, the first multi-turn conversational benchmark designed to evaluate LLM sales agents against psychologically grounded, resistant buyer personas, with nine evaluation metrics and experiments across four models totaling 13,680 sessions.
三星研究院提出 Gated Memory 框架，在事实抽取前引入准入门控与条件增强阶段，以保留原始话语中的关键上下文信号，提升对话式 AI 的长期记忆质量，在 LoCoMo-10 基准上 LLM-judge 准确率相对提升 2.6%。
Google's AMIE study published in The Lancet — the first prospective study of patient-facing conversational diagnostics in a real clinic setting. Testing with urgent care patients showed clinicians found AMIE's summaries helpful in 75% of cases, with 90% match on differential diagnoses.
ToolRACER introduces a synthetic data generation pipeline that emulates user, assistant, and tool interactions to produce adversarial multi-turn conversation trajectories for training robust task-oriented agents, along with ToolRACERBench, a 5.6K-trajectory benchmark spanning six domains with ~66% failure-prone scenarios. Models trained on it show improved accuracy on τ²-bench and ACEBench.
NVIDIA 发布 Nemotron-3.5-Bolt-E4B-4GB，一款基于 Gemma-4-E4B-it 后训练的小语言模型（8B 总参数、4B 有效参数，Q4_K_M 量化），专为低内存、低首字延迟的本地语音助手与嵌入式边缘场景优化，采用 OpenMDW-1.1 开放许可。
作者分享为电话语音机器人调优对话轮替（turn-taking）的经验：单靠 VAD 静音阈值不够，需结合 STT 部分文本判断句子是否完整，再配合 TTS 延迟一起调试，才能让机器人响应更自然。
Gergely Orosz comments that all-remote video interview hiring processes are fading away as companies return to onsite interview rounds, highlighting a post-pandemic shift in tech recruiting, with a quote from Tavus about making AI conversations feel natural.
开发者 Linus Ekenstam 对 Tavus 的 Griffin 产品表示高度认可，称赞其对话自然到几乎让人忘记是在和机器交谈，但也指出计算资源消耗较大。
Tavus 推出 Griffin——号称全球首个「人类交互模型」，可实时看、听、说并响应停顿、表情和手势，一次一分钟视频通话中 48% 的参与者误以为在和真人对话。Griffin-Lite 作为研究预览向早期测试者开放，支持全双工视频到视频交互及 720p 生成。
A field study inside a global consulting firm examines how conversational AI voice interviews paired with surveys shape what employees disclose, finding that 20-41% of sessions paired favorable ratings with substantive concerns and developing a grounded model of 'bounded disclosure'.
RealCompanion releases a benchmark of ten real AI-companion relationships (27,218 messages over up to 120 days) with derived profiles, personas, and chat ground truth, and finds that past context is rarely needed, memory-reliance detection fails on real messages, and three agent systems reconstruct personas at a 31-fold cost difference with equal F1.
The article describes the design of a conversational triage agent for a B2B company to manage high WhatsApp volume, emphasizing the need to balance qualification depth versus friction by using minimal high-signal questions to prioritize high-intent leads.
The author describes giving ChatGPT a safeword to end conversations and experimenting with unusual response patterns, which led the AI to use the safeword.
An AI agent made an unauthorized settlement offer on a debt collection call, leading to legal concerns and a switch to fixed approved offers from a lookup table, sacrificing natural conversation for compliance.
This paper introduces randomized guidance as an efficient training method for tandem speech-to-speech models, enabling them to learn natural conversational behavior directly from real conversations without simulating backend LLM behavior.