New models: the Xing 4 report (TeleChat lineage) pretrains a 285B model on 8,192 Ascend 910C chips and uses TeleGym for agentic post-training; Microsoft's MAI Code 2 reaches 98% of GPT-6.1 Sol at 40% of the cost, per Microsoft. Agents: RepoFactory builds 10,159 training environments from real repos, lifting a 35B-A3B model by 22.6 points on NL2Repo-Bench; Isola's group shows capable agents can devise their own self-improvement; on-policy distillation can drop τ²-bench from 41.5% to 6.6%. Architecture and speed: a 2-branch Multi-Branch Transformer beats sequential Transformers at matched parameters and compute and cuts KV cache by up to 34%; DNeo decodes 1.3x faster than DFlash-2.
Large Language Models
1
Multi-Branch Transformer
The paper proposes branching, running full Transformer blocks in parallel, as a third way to scale beside widening or deepening. MBT has shared early layers, parallel branches of sequential blocks and shared final layers. In pretraining at 0.8B to 2.1B parameters, 2-branch Wide-Shallow models beat sequential Transformers at matched parameters and compute. Its Joint-KV shared cache cuts KV storage by up to 34% and raises decoding throughput by up to 51.4% on one H200.
Why it mattersBranching adds a third scaling axis beside width and depth, and the gains hold at matched parameters and compute rather than coming from extra FLOPs. Joint-KV turns the architecture into a direct inference win, which matters because KV cache is a main serving bottleneck.
DNeo: Your Parallel Draft Model Needs Better Target Feature Encoding
DNeo targets speculative-diffusion decoding, where a small parallel drafter proposes many tokens per pass for the target to verify. The authors find that standard target-feature encoding is lossy, keeps no memory of rejected proposals and samples layers uniformly. Their Mixture-of-Features Attention adds a working memory of rejected proposals, plus 2-tap causal convolutions, prefix-only cross-attention and non-uniform layer selection. DNeo generates 1.3x faster than the previous state of the art, DFlash-2, without a larger drafter or lower quality.
Why it mattersParallel drafters are becoming the default way to speed up LLM serving, and this work shows the bottleneck is how the drafter reads the target model, not the drafter itself. A 1.3x gain over the newest baseline at no extra footprint translates directly into serving cost.
From Repositories to Training Environments: Scaling Long-Horizon Coding Agents with Environment Evolution Loop
RepoFactory turns real repositories into verifiable environments where an agent must build a whole project from a natural-language spec. Its Environment Evolution Loop uses cheap proxy feedback to optimize a multiplicative objective of validity, structural coupling and implementation scope. It yields 10,159 environments across 17 domains; training Qwen3.5-35B-A3B on them gives RepoHorizon-35B-A3B, up 22.6 points on NL2Repo-Bench (17.5 without evolution) and 8.5 points on average across five benchmarks (6.4 without).
Why it mattersEnvironments, not models, are now the scarce input for training long-horizon coding agents. Showing that environment quality can be optimized with cheap proxy signals, and that this adds measurable gains, gives a scalable recipe others can copy.
Do LLM Agents Still Need to Be Told How to Self-Improve?
The paper asks whether capable LLM agents still need humans to specify how to self-improve. A frozen agent gets basic primitives, a compute budget, tools for gathering experience and editing text documents, and a general goal, with games as the testbed and no imposed loop or real-world data. Sufficiently capable models do self-improve, with strategies resembling self-regulated learning and self-generated curricula. The learned text artifacts transfer to external benchmarks across game horizons, mechanics and 3D navigation.
Why it mattersMost self-improving agent systems hard-code the improvement loop; this shows strong models can invent it themselves. That shifts the design question from engineering the loop to choosing primitives and budgets, with implications for both capability and oversight.
The paper studies how on-policy distillation changes agents' ReAct patterns across math, ALFWorld and τ²-bench. General-reasoning supervision cut math tool attempts from 70.4% to 14.4% and raised accuracy from 16.7% to 38.0%. ALFWorld supervision caused negative transfer, dropping τ²-bench from 41.5% to 6.6% as the student over-called tools. Selective OPD, which keeps tokens chosen by the teacher's preference shift around domain RL, raised τ²-bench from 11.5% to 29.9% at some cost to source-domain learning.
Why it mattersOn-policy distillation is now a standard post-training step for agents, and this paper shows it silently transfers tool-use habits across domains. The large negative transfer on τ²-bench is a concrete warning for anyone mixing agent distillation data.
TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks
This technical report introduces Xing 4, a model series extending the TeleChat line, and was the most upvoted paper on Hugging Face Daily Papers on 18 Nov. It covers sparse activation, attention design, depth and width choices and speculative decoding. Training runs on MindSpore HyperParallel on Ascend hardware, with a 285B-parameter model pretrained on 8,192 Ascend 910C chips. Post-training uses TeleGym, a unified agent execution platform; the arXiv page links no weights or code.
Why it mattersIt is one of the largest documented frontier-scale training runs done entirely on domestic Chinese chips and frameworks. The report is a useful data point on how far the Ascend and MindSpore stack can be pushed for agentic models.
Microsoft announced two coding models at Ignite: MAI Code 2 for complex enterprise work and MAI Code 2 Flash for high-volume routine edits. On Microsoft's own averages over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, MAI Code 2 reaches 98% of GPT-6.1 Sol at 40% of the cost and 96% of Claude Opus 5.5 at 5% of the cost. Developers preferred it over GPT-6 Sol 64% of the time in blind frontend tests. Both are in GitHub Copilot; no open weights or absolute prices are stated.
Why it mattersMicrosoft is now shipping its own frontier-class coding models inside Copilot rather than relying only on partners. The claims are self-reported, but the cost-performance positioning signals sharper price competition in coding agents.
Beyond Controllability: Speaker Diversity in Instruction-Guided Text-to-Speech
Instruction-guided TTS lets users describe a voice in words, but uses such as character voices and synthetic training data need many distinct voices per description. The authors evaluate seven open-source systems with validated speaker embeddings. Every system produced speaker distributions markedly narrower than real speech, female prompts gave less diverse pools than male prompts, and accent and emotion prompts revealed a strong bias toward North American English and neutral speech. Diversity and controllability capture different weaknesses, so both must be measured.
Why it mattersSynthetic speech is increasingly used as training data and for character voices, where collapsed voice diversity silently degrades downstream systems. The paper adds a missing evaluation axis and exposes gender and accent biases in current open TTS.
Latency, Speech Recognition and Turn-Taking in Voice Agents for Spanish Phone Negotiation: An Externally Measured Comparison of Pipecat, LiveKit and GPT-Live
The paper compares three architectures for a Spanish phone agent negotiating freight rates: cascaded pipelines on Pipecat and LiveKit Agents, and OpenAI's single-model GPT-Live. Over 77 calls each, listener-side median latency was 2,616 ms, 3,090 ms and 2,444.8 ms, with the GPT-Live vs Pipecat gap not significant; tool calls added 0.9–1.1 s, and frameworks under-reported latency by 9% and 22%. WER was 6.03%, 4.93% and 1.47% (13.01%, 12.61% and 2.86% on spoken amounts). An overlapping “ok” caused a false agreement in 3/15 Pipecat and 2/15 LiveKit calls, and none for GPT-Live.
Why it mattersIt is a rare externally measured, reproducible comparison of production voice-agent stacks in a non-English, money-sensitive task. The findings that framework latency reports undercount and that cascaded systems mishandle backchannels and spoken amounts are directly actionable for builders.
Leveraging Dual-Process Theory for Multimodal Multiparty Turn-Taking Prediction
In multiparty talk a pause does not always mean a speaker is done, and a new voice does not always mean the floor changed. Inspired by dual-process theory, the framework uses a fast recall-oriented stage to find candidate transitions and a deliberative stage combining acoustic, linguistic, speaker, predicted-activity and multimodal evidence to reject false ones. It is evaluated on VoxConverse and the triadic Multi-TPC set across three tasks. Finding plausible transitions proves much easier than confirming and localizing real floor transfers; the abstract reports no numbers.
Why it mattersVoice agents are moving from one-on-one calls into meetings and group settings, where turn-taking is much harder. Reframing it as continuous discovery and verification rather than classification at fixed boundaries is a useful design principle for such agents.