language-models
64 posts tagged language-models.
- LLMON puts markup closer to the model
- AutoScaffold trains models with hints, then removes the hints
- B3 speeds up diffusion language models by overlapping the slow part
- NarrativeBench tests whether AI search shows the stories it leaves out
- A World Map of LLM Answers Is an Eval, Not a Belief System
- Tencent’s Qwen DSpine hint is about inference cost, not model size
- MAEMs turn residual stream directions into testable text
- Forecasting an LLM’s answer before token one
- LoopVQ and the Hidden Memory Cost of Looped Transformers
- TextSeal2 keeps watermarks alive under speculative decoding
- Mistral Large 4 lands where builders actually test models
- MuonGS relaxes weight-norm control without giving up training speed
- What Gemini 4 Carbon Actually Tells Us (and What It Doesn't)
- dSpark and the Order Problem in Diffusion Language Models
- AdaBPE makes tokenization care about LaTeX
- AI prose gets hard to read when it removes friction
- Speech-only conversation models move the hard part to turn-taking
- When Deferring to a Human Beats Letting the LLM Guess
- Adaptive MTP makes speculative decoding less wasteful
- Beam’s 501B claim is really about control, not size
- Indic pretraining is becoming a data pipeline problem
- Time-series models need two views, not one prettier chart
- PerSo splits persuasion into reading people and choosing the next move
- Back-translation beats more data in low-resource translation
- Temperature Should Read the Room
- Meta’s Muse Wonder looks competitive, not crowned
- The Same Attention Head Does Different Jobs Depending on the Task
- Wikipedia Game results show why AI search still needs guardrails
- Legal AI citations are not proof of legal reasoning
- LLMs belong in the feature factory, not always in production
- Tokenizer compression is not the same as multilingual quality
- Scoring LLMs When There Is No Right Answer: A Lesson From Survey Economics
- IdeaAnchor Trains Models on the Shape of Research Ideas
- Pivot-SD trains the denoising choices that steer diffusion LMs
- Local LLM runners are becoming product choices, not hobby projects
- What HC-DLM Fixes About Diffusion Language Models
- The multilingual loophole in LLM unlearning
- Wild AI Web Text Is Starting to Poison Pretraining
- NoPE attention is not position-blind if local layers carry recency
- Not All Quantization Errors Are Equal: STEPQuant and the Memory Cost of Linear Attention
- Telescopic Language Models: one training run that serves every compute budget
- Models may refuse based on who they think you are
- What LLMs Miss About Haitian Creole, and Why Low-Resource Culture Breaks Evals
- Audio models do not automatically share speech and text concepts
- MISVO steers frozen language models at inference time
- When a model’s rejection reason changes its next choice
- Authorship verification works better as similarity, not classification
- Translation Fine-Tuning Breaks the Controls General Evals Miss
- Speaker-Centered Memory: Why Group Chats Break Your AI Agent
- Membership Inference Gets an Entropy Correction: Reading the ETD Paper
- Softmax attention needs an off switch
- ComPO and the Case Against Optimizing the Loss You Wrote Down
- What Actually Makes a Tokeniser Good: Search Beats Objective
- Teaching Models to Say 'I Don't Know' With a Prompt, Not a Retrain
- A thin local AI signal still tells builders what to watch
- When a Language Model's Odds Don't Add Up
- ConvergeFlow's Bet: Kill the Decoder, Trust the Flow
- Mimir v1 tests whether small open models can win on clean data
- Entity matching needs architecture tests, not bigger-model reflexes
- Maskability Index makes prompt choice less vibes-based
- Language model embeddings do not want to collapse
- Co-LMLM Puts Facts in a Database, Not the Weights
- DiaLLM separates dialect understanding from dialect writing
- Safety evals need to test messy language, not just bad behavior