Research Daily

One research paper at a time, taken apart slowly. Each lesson builds the ideas from first principles, draws them, argues with the authors, and ends with a quiz so the ideas stay with you.

New here? Read how this site works first. It takes two minutes.

Latest papers

Added 20 Oct 2026 Ming-3.0-tts-tiny Ming-3.0-tts-tiny (open-weight model release, no paper) A speech model that never picks a token: each 80 ms of audio is 64 real numbers, drawn by a small flow model from a 7.9B mixture-of-experts language model's hidden state. Stretch. 23 pages, quick tour 8 min SpeechTTSGenerative modelsMultilingual Finished Added 20 Oct 2026 DeepInfer DeepInfer: High-performance, large-scale inference framework for DeepSeek models (open-source release, no paper) DeepSeek open-sourced the engine that serves 155 trillion tokens a day. Here is how one request travels through it, and which tricks you could borrow. Stretch. 23 pages, quick tour 10 min LLMsSystemsEfficiency Finished Added 20 Oct 2026 ByteRobust Robust LLM Training Infrastructure at ByteDance A 9,600-GPU training run breaks every few hours. How do you keep it training 97% of the time? Comfortable. 25 pages, quick tour 10 min LLMsSystemsRobustnessEfficiency Finished Added 15 Oct 2026 Mind the Accent Gap Mind the Accent Gap: British Accent Robustness in Speech-Driven Financial Voice Assistants A Scottish customer asks a banking assistant about spending at Tesco, and the transcript reads "Hi Monte Santesco". How often does that happen, and which ASR prevents it? Warm-up. 19 pages, quick tour 8 min SpeechASRBenchmarksRobustness Finished Added 15 Oct 2026 Llama 4 Herd The Llama 4 Herd: Architecture, Training, Evaluation, and Deployment Notes Two open models both use 17B parameters per token, yet one needs almost four times the memory. A withdrawn, unofficial survey of Llama 4 explains why. Comfortable. 22 pages, quick tour 10 min LLMsMultimodalEfficiencyDistillation Finished Added 14 Oct 2026 OmniVoice OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models One 0.8B model clones a voice in 646 languages, and half of them had about ten hours of training audio or less. Stretch. 26 pages, quick tour 8 min SpeechTTSMultilingualGenerative models Finished Added 14 Oct 2026 AudioLBM Audio Super-Resolution with Latent Bridge Models Most upsamplers paint the missing treble starting from random noise. AudioLBM starts from the muffled clip itself and walks it to 48 kHz, then on to 192 kHz. Stretch. 24 pages, quick tour 8 min SpeechAudioGenerative modelsDiffusion Finished Added 14 Oct 2026 AudioSR AudioSR: Versatile Audio Super-resolution at Scale One diffusion model takes speech, music or sound effects recorded at anywhere from 4 to 32 kHz and fills in the missing treble up to 48 kHz. Comfortable. 23 pages, quick tour 8 min SpeechAudioGenerative modelsDiffusion Finished Added 11 Oct 2026 Breeze TTS 2 Breeze TTS 2 (open-weight model release, no paper) A streaming TTS model that tops the open-weight leaderboard ships with code and weights but no paper. The code still tells you exactly how it works. Comfortable. 22 pages, quick tour 8 min SpeechTTSGenerative modelsEfficiency Finished Added 4 Oct 2026 OPD Scaling Laws Scaling Properties of Same-Family On-Policy Distillation A 0.5B math model that scores 39.8% can teach a 14B model to 77.7%. Yet a bigger teacher can make a small student worse, and of two teachers with near-equal scores, the bigger one teaches worse. This paper finds the laws behind all three. Stretch. 25 pages, quick tour 10 min LLMsDistillationScaling lawsReinforcement learning Finished Added 2 Oct 2026 ICL for Robots In-Context Learning for Robots: Methods and Applications A robot that copies your packing demo and one that ignores it can finish with the exact same tray. This 100-page survey asks how robots learn from a demo without retraining, and how you'd ever tell. Stretch. 25 pages, quick tour 8 min RoboticsIn-context learningAgentsMultimodal Finished Added 30 Sep 2026 Duplex-MPE Duplex-MPE: Benchmarking Multi-Party Interaction in Full-Duplex Dialogue One voice model has speech after 99.55% of the requests put to it. Sounds great, until you learn it also talks through 20,355 of the 20,360 moments when it should have kept quiet. Comfortable. 22 pages, quick tour 8 min SpeechBenchmarksEvaluationDatasets Finished Added 30 Sep 2026 SchrodingerRepo Schrödinger's Code Repository: Have LLMs Learned SWE-bench or Memorized It? Rename the files and functions in Django without changing what the code does, and coding agents solve 6 to 14 points fewer SWE-bench tasks. Comfortable. 25 pages, quick tour 8 min LLMsAgentsCodeBenchmarks Finished Added 30 Sep 2026 Disaggregated Quantization Disaggregated Quantization: Specializing LLM Prefill and Decode Take a 1-bit Qwen3.8-27B that scores 29% on MMLU-Pro, leave its weights untouched, and add a separately trained 4-bit prompt reader streamed from the SSD: it scores 61.5% and handles long prompts up to 1.78× faster. Stretch. 25 pages, quick tour 8 min LLMsEfficiencySystems Finished

Browse every paper or explore by tag.