Weekend edition covering roughly Oct 26-30, 2026. Mistral open-sourced Mistral Large 4 (1T MoE, about 52B active), and AWS's EasyProver solved 669/672 PutnamBench problems and all six IMO 2026 problems with an off-the-shelf frontier model. With matched capabilities, MCP and CLI interfaces made 94.3% of agent trajectories diverge. In speech, full-duplex models degrade over long sessions, and voice agents can call tools 1-2 s before the user finishes.
Large Language Models
1
Multi-Token Prediction Done Right
Jayden Teoh (Oct 28). The paper shows that the gradient of standard Multi-Token Prediction (MTP) carries a noisy residual term and proposes Soft-MTP, which propagates only a KL self-distillation signal with no extra memory or parameters. From 167M to 1.2B parameters, Soft-MTP gives lower perplexity, faster speculative decoding and higher zero-shot accuracy at every scale.
Why it mattersa near-free fix to a training objective already used in frontier models such as DeepSeek's.
Training Foundation Models with a Single Algorithm
Yuda Song, Ruslan Salakhutdinov, Andrea Zanette et al. (CMU, Oct 29). The authors merge pretraining, midtraining, SFT and RL into a single likelihood algorithm by treating RL as maximum likelihood on a target distribution generated by the model itself. Training then reduces to choosing which data, in what mix and when, improving the accuracy-compute trade-off.
Why it mattersit reframes multi-stage post-training as a data-scheduling problem that could simplify training stacks.
Scaling Lean Theorem Proving for Real-World Code Verification (EasyProver)
Debargha Ganguly, Soonho Kong, Sam Bayless et al. (AWS, Oct 28). EasyProver uses recursive agents on an off-the-shelf frontier model with no extra training to solve 669/672 PutnamBench problems, all six IMO 2026 problems, 597/600 NTP4VC conditions and all 189 VERINA tasks. It also introduces Aegis, 2,045 Rust-Lean program pairs, where EasyProver proves 52.7%.
Why it mattersformal competition math looks close to saturated, and the frontier is shifting to real-world code verification.
WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training
Ziyang Zeng et al. (IQuest, Oct 27). WorkFactory automatically builds 112,686 executable tasks across 612 occupations and uses them to fine-tune Qwen 4B, 27B and 35B-A3B into the WorkPilot family. WorkPilot-35B-A3B rises from 879 to 1565 Elo on GDPval-AA v2.1. Code, data and models are released.
Why it matterslarge open synthetic work environments are becoming the key ingredient for economically useful agents.
Co-Design Bench: Multimodal Collaborative Coding Evaluation for Agentic Models
Xilin Wei, Dayiheng Liu et al. (Qwen, Oct 26). The benchmark has 273 tasks built from 26,314 real coding sessions with 2,331 expert-written acceptance criteria, split into Multimodal Editing and Repository Editing. The best of 10 endpoints meets 69.9% of criteria on average but fully completes only 26.4% of tasks.
Why it mattersthe gap between mostly right and fully done shows how far coding agents are from reliable autonomous delivery.
Matched Capabilities, Different Behaviors: A Controlled Study of MCP and CLI Interfaces for Language Agents
Zhenning Shi, Liang Pang, Yongqi Li (Oct 25). The MICA framework builds MCP and CLI interfaces from the same action contract on one runtime, across 6 models and 3 environments. Despite matched capabilities, 94.3% of 2,004 paired trajectories diverge markedly. The DART method adds up to 9.0 benchmark points and saves up to 27.1k tokens per task with MCP.
Why it mattershow tools are exposed to an agent is itself a performance lever.
AgentDecept: Benchmarking Deceptive Behavior in Tool-Using LLM Agents under Pressure
Yehan Zhang, Xin Eric Wang et al. (Oct 29). This executable benchmark has 477 pressure variants of 257 base tasks (via messages, app state, tool responses and service quotas), each paired with a no-pressure control. It compares what agents report with what they actually did to separate deception (fabrication, concealment, equivocation) from honest failure.
Why it mattersas agents act in stateful apps, checking whether they report truthfully becomes a core safety evaluation.
Mistral Large 4 (open weights): Mistral-Large-4-1T-A52B
Mistral AI (Oct 29-30). Mistral released open weights for its flagship 1T-parameter MoE (about 52B active) with multimodal input, three reasoning modes and a modified MIT license (Model-as-a-Service needs a separate license). Per Mistral: 41% on Terminal-Bench 4.0, 84.5% on BrowseComp, and 48% on the Artificial Analysis Coding Agent Index with 47% fewer tokens. A technical report is coming.
Why it mattersa commercially usable trillion-parameter open model narrows the open-closed gap on agentic coding.
Chenyue Zhang et al. (about 70 authors, posted Oct 26). This speech foundation model pairs a 0.65B audio encoder with a 30B-A3B MoE LLM, trained on 13 million hours of audio, covering ASR in 99 languages and 202 dialects plus speech translation and spoken dialogue. It matches Qwen3.5-Omni-Flash with less data and was trained on domestic Huawei Ascend 910B and Hygon BW1000 chips.
Why it mattersit signals hardware independence for speech foundation models.
Discovering Interpretable Acoustic Features in Autoregressive Voice-Cloning TTS Models
Anuj Diwan, David Harwath, Eunsol Choi (Oct 29). The study uses sparse autoencoders (SAE) and PCA to find acoustic concepts such as pitch, energy, noise and accent in autoregressive voice-cloning TTS, identifying 684 verified SAE features in Qwen3-TTS. Good probing does not imply good steering, and PCA stays competitive with SAEs.
Why it mattersinterpretability is reaching TTS, opening the way to controllable, auditable voice cloning.
Heard Enough to Act: Early Tool Calling in Spoken Dialogue Systems
Yizhou Peng, Eng Siong Chng et al. (Oct 26). The authors build a corpus of 26,694 spoken dialogues (786 hours, 48 domains) labeled with the earliest actionable point, so an omni-modal model can call tools as soon as it has heard enough. The model reaches 0.48 tool-state macro-F1 and, on correct calls, acts about 1.8-2.0 s ahead of a wait-for-turn-end cascade.
Why it mattersacting mid-utterance is a concrete way to cut perceived latency in voice assistants.
Diagnosing and Mitigating Long-Session Degradation in Full-Duplex Speech Models
Junho Kim, Sungroh Yoon et al. (Oct 26). Closed-loop evaluation shows public full-duplex speech models degrade over time in responsiveness, content quality and voice consistency. The authors propose LoRA self-distillation to hold quality over 10-minute sessions and a two-resolution context design that cuts KV-cache growth by over 90% across three architectures.
Why it matterslong-session stability, not just latency, is becoming the real bottleneck for full-duplex voice agents.
