FlowLM: A New Paradigm for Converting Diffusion Models to Few-Step Generation

Diffusion-based language models can generate high-quality text, but they require as many as 2,000 steps of iterative sampling, which has long been a practicality hurdle. To address this, FlowLM proposes a method for efficiently converting existing diffusion language models into "flow matching."

By realigning curved sampling trajectories into straight flows, it achieves quality on par with or better than 2,000-step diffusion sampling in just a few steps. Moreover, the model reaches performance saturation with only half the training epochs needed for training from scratch, confirming that it effectively leverages the knowledge of pretrained models.

From the perspective of accelerating text generation, this can be regarded as a research advance that takes a solid step toward practical application.

"Unlearnable Problems" Also Exist in RLVR

Reinforcement learning with verifiable rewards (RLVR) is widely regarded as an indispensable technique for improving language model performance, but a study featured on Hugging Face Daily Papers has pointed out a significant limitation: even in settings where correct answers are occasionally sampled, "the majority of hard problems remain unlearnable."

The cause is attributed to "flawed internal representations" that reward-based training cannot repair. This is an important correction to the view that RLVR is a cure-all, and it suggests the need for approaches that improve the model's representations themselves.

"Follow the Instructions" vs. "Complete the Pattern" — The Internal Conflict Within LLMs

A paper titled "Do as I Say, Not as I Do" empirically examined the clash between "instruction following" and "pattern completion" in LLMs.

The study investigated, across 13 models and 16 types of instructions, the phenomenon in which models get pulled toward adversarial patterns embedded in the conversational context despite explicit user instructions such as "answer in a specific language" or "output a specific token." Instruction-following rates varied widely across models, ranging from 1% to 99%, and did not correlate with standard capability benchmarks.

Particularly interesting is the finding that while models can predict their own behavior with 83.5% accuracy, they tend to "underestimate" their own susceptibility to patterns. Chain-of-Thought reasoning improves robustness, but it can also produce a divergence (dissociation) between correct reasoning and incorrect outputs.

Mix-Quant: Accelerating Long-Context Inference for Agent LLMs

LLM agents handle complex tasks through tool use, memory retrieval, and multi-turn interactions, but the overhead on the input side is substantial, making the prefilling stage a frequent inference bottleneck.

Mix-Quant tackles this problem by proposing "phase-aware quantization." It is a quantization strategy optimized for each stage: low-precision quantization for speed during input processing (prefilling), and high precision maintained during output generation (decoding). As a practical approach that delivers substantial speedups in agents' long-context, multi-turn inference without sacrificing accuracy, it is attracting attention.

Decoding the "Coupling" Mechanism of Full-Duplex Speech Dialogue with Moshi

In human conversation, "neural coupling" — the synchronization of speakers' brain activity — is a well-known phenomenon. A study has been published that verifies this phenomenon in AI speech dialogue models.

The researchers simulated full-duplex dialogue between two instances of the Moshi model, manipulating channel noise and other conditions to analyze the coordination of their internal representations. This work brings us one step closer to elucidating interaction dynamics that resemble human conversation — something turn-based systems cannot achieve.

Full-duplex speech dialogue could become a core technology for next-generation AI assistants, making this an important study for deepening our fundamental understanding of it.

---