Large Language Models
1
Beyond Passive Imitation: On-Policy Mid-Training Distillation for Language Models
OPMD brings on-policy distillation into mid-training: the student samples its own trajectories and gets dense reverse-KL rewards from a frozen teacher. On MATH-500 and Minerva it lifts Qwen3-1.7B-Base, while plain next-token training on the same 1B math tokens degrades it, with no inference overhead.
Why it mattersOn-policy distillation has mostly been a post-training tool; this moves it into mid-training, where plain next-token prediction can actively hurt some base models. If it holds at scale, domain continual pre-training may shift from "more tokens" to teacher-guided learning on the student's own trajectories.
arXiv ↗PDF ↓
2
Evolving Infra: Agent-Adaptive LLM Serving at Scale
An agent-aware serving system that lets nodes switch between prefill and decode and evicts KV cache using agent context signals. With GLM-5.2 and Kimi K2.6 on up to 256 GB300 GPUs, it cuts mean end-to-end latency up to 4.1x and P95 TTFT up to 14x versus NVIDIA Dynamo.
Why it mattersAgentic workloads reuse context very differently from chat, and serving stacks tuned for chat leave a lot of performance on the table. Coming from the Stoica and Song Han groups at production GPU scale, this signals that agent-aware serving is becoming its own systems category.
arXiv ↗PDF ↓
AI Agents
3
Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
When each user has their own agent sharing a resource, peer-to-peer teams trail a single coordinator by about 10–45 points across four environments; GPT and Qwen agents cancel others' jobs without asking 3.9–7.1 times per episode. A team-lead prompt and a read-before-commit guard recover much of the gap; MAMUBench is released.
Why it mattersAs every user gets their own agent, shared resources become a coordination problem that today's models handle poorly and sometimes unsafely. MAMUBench gives an open way to measure this, and the simple mitigations show that system architecture matters as much as model choice.
Anthropic Alignment blog ↗arXiv ↗PDF ↓Code (MAMUBench) ↗
4
SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering
A full-precision teacher steers a quantized agent to fix failed steps using execution feedback, then the fixes are distilled into the student. Under W4A16 it beats the best quantized baseline by 1.67–2.18 points across five agent benchmarks and stays within 1 point of BF16, using 67% fewer GPU hours than QeRL on τ²-Bench.
Why it mattersMost quantization work is evaluated on static benchmarks, but agents fail in compounding, multi-step ways. A closed-loop, execution-aware recipe that nearly closes the gap to BF16 makes cheap on-device and high-volume agents more realistic.
arXiv ↗PDF ↓
5
GTAGym: Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning
An evaluation and RL environment built on GTA V, with 103 story-derived online tasks and 47 image-based shooting tasks checked from engine state. The best of six frontier LLMs reaches 70.9% on GTAVBench-AGI, and LoRA RL on generated tasks lifts Qwen3.8-27B from 15% to 23% on native GTA V tasks.
Why it mattersCommercial AAA games are a rich, hard-to-game testbed for long-horizon visual control, and pairing a benchmark with a procedural task generator makes it usable for training, not just evaluation. The remaining weak spots (aerial navigation, interception) show where computer-use agents still lack fast perception–action loops.
arXiv ↗PDF ↓
6
Dependable Long-Horizon LLM Agents: A Trajectory-Centered Empirical Study of Where Errors Enter and Which Controls Reach Them
A trajectory-level study of 95 failed runs from six multi-agent systems: faults enter early (median 18% into a run), 72% of failures already contained contradicting evidence that no agent acted on, and 45% of hand-crafted Who&When failures were really provider content-filter crashes.
Why it mattersIt shifts agent reliability work from labelling failures to locating where controls could have intervened, and shows that much of the needed evidence is already in the trajectory. The harness-crash finding is also a warning that some "agent failures" in public datasets are infrastructure artifacts.
arXiv ↗PDF ↓
Model Releases
7
Gemini Decision 1 (Gemini Decisions API)
A model that scores candidate options in one forward pass instead of generating text, returning calibrated probabilities. It handles up to 1,000 yes/no, multiple-choice or rubric questions per request over multimodal input up to 256K tokens, in 150–250 ms, at $0.10 per 1M input tokens with free output (Google-reported).
Why it mattersMany agent steps (tool selection, routing, policy checks, grading) are bounded choices that do not need free-form generation. A calibrated, single-pass decision model priced on input only could become a standard building block for agent harnesses and LLM-as-judge pipelines.
Google blog ↗
8
ERNIE 5.5 Flash Preview
The first post-trained model on ERNIE 5.5 Flash Base: a 280B-parameter MoE with 14B active and a 1M-token context, aimed at coding, 3D modeling and knowledge work. Available via Qianfan and OpenRouter; Baidu says weights will be open-sourced under Apache 2.0 soon, and benchmark results are shown only as images.
Why it mattersA 14B-active, 1M-context MoE promised under Apache 2.0 would add another strong, cheap-to-serve option to the open model ecosystem. Until the weights and checkable benchmark numbers ship, treat it as a preview rather than a settled result.
ERNIE blog ↗
Speech
9
Walrus Technical Report
An 8B speech and audio model initialized from Gemma 4 that covers ASR, speech translation, TTS and audio understanding across 100 languages. A three-stage data curriculum over 600M+ samples lets it match or beat the ~4x larger Qwen3-Omni on most benchmarks.
Why it mattersIt is a clean argument that data curation and curriculum, not novel architecture, drive multi-task speech quality. Matching a ~4x larger omni model with an off-the-shelf backbone is useful guidance for teams building speech LLMs on modest budgets; no weights are linked yet.
arXiv ↗PDF ↓
10
Emilia2: A Large-Scale Dataset for Controllable Long-Form Speech Generation
A 229K-hour multilingual in-the-wild corpus for long-form and multi-speaker speech generation, with style instructions for 4.77M examples, nonverbal tags, and a 23-hour test set with recordings up to 30 minutes. A 0.6B model trained on it posts the lowest CER/WER on TTSD-Eval among systems compared.
Why it mattersLong-form, multi-speaker generation (podcasts, audiobooks, dialogue) is where current TTS breaks down, and open data at this length has been scarce. A large instruction- and nonverbal-tagged corpus plus a 30-minute test set gives the field a shared target; the download link and license are not yet stated on arXiv.
arXiv ↗PDF ↓Demo ↗
Voice Agents
11
Introducing Live 3: Built for the rhythm of conversation
Fish Audio's full-duplex voice model listens while it speaks and moves conversational timing into the speech model itself, trained on real conversations, with a separate reasoning layer for instructions and tools. It can decide to respond in as little as 186 ms; English only, no quantitative benchmarks published.
Why it mattersIt reflects the industry shift from cascaded voice pipelines to speech-native turn-taking with a separate reasoning layer for tools. Without public benchmarks, the 186 ms claim is best checked against open full-duplex turn-taking evaluations.
Fish Audio blog ↗
12
Earlybird: Predictive Streaming Speech Recognition in 29M Parameters
A 29M-parameter (16.8 MB at 4-bit) streaming recognizer trained on audio cut mid-word to predict the current and next words. The median word on LibriSpeech is predicted 90 ms before it ends and 69% of audiobook words are predicted early, versus 3–25% for other models; weights are on Hugging Face.
Why it mattersFor voice agents, every 100 ms of endpointing latency matters, and predicting words before they end lets the downstream LLM start working early. A tiny, calibrated, open model makes this easy to try on-device.
arXiv ↗PDF ↓Hugging Face (weights) ↗
