Search posts, papers, and topics
The indefinite article in English depends on the initial sound, not the spelling. An analysis of the CMU pronunciation dictionary shows only 129 of 32,455 words are exceptions, and the first two letters are usually enough to decide between “a” and “an”.
The author shows how to label tokens (dates, money, emails, etc.) with a single regex pass using extended operators (`&` for intersection, `~` for complement) in the resharp library. A tiny benchmark compares 10 patterns against spaCy’s NER, reporting ~1.9 GB/s (≈4500× faster) on 8 threads. The post lists the concrete patterns and argues that for deterministic, regular‑language entities regex can…
On-policy distillation (OPD) can lead to excessively long student responses, a phenomenon called length inflation. This paper identifies "termination-token mismatch" between base students and post-trained teachers as a key source, where models place stopping probability on different EOS tokens. Treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigate…
The authors present VākQA, a 2,001‑question spoken factoid QA benchmark for Telugu with audio, transcriptions, and human‑verified answers, and they validate automatic evaluation methods against human ratings. Using this setup they show that translation loses cultural nuance, ASR errors alter meaning, and cascaded ASR‑MT errors degrade model performance.
Cross‑corpus study of gaze behavior in two collaborative dialogue datasets (MapTask, MUNDEX) shows that task‑aligned references correlate with more task‑directed, less partner‑directed gaze, lower entropy and fewer transitions. Temporal gaze features (MapTask) and raw proportion features (MUNDX) modestly improve grounding prediction over baselines, but effects are small and diminish when aggregat…
The paper introduces TeleAntiFraud 2.0, a monthly refreshed, frozen audio benchmark of 900 Chinese telecom calls (600 fraud, 300 near‑domain non‑fraud) built with a Mixed‑Tree generation pipeline. Experiments show models that score perfectly on unrelated negatives fall to ~0.66 Macro‑F1 on near‑domain cases, exposing shortcut learning and prediction collapse.
Hugging Face has released `tokenizers` v1, a major performance update that achieves 3-30x faster encoding than v0.23 while maintaining identical output and API compatibility. Key optimizations include a SIMD-accelerated splitter, a thread-local word cache, and an allocation-free BPE merge loop, ensuring tokenization doesn't bottleneck ML workflows.
This paper introduces HEAL, a method to identify and mitigate hallucinations in Multimodal Large Language Models (MLLMs). It finds that hallucinations stem from information distribution drift in "synergy heads" within the attention mechanism and proposes dynamic calibration to steer output towards factual evidence.
Mind2Dialogue introduces a psychology‑guided simulator that generates user mental states and uses privileged distillation to train LLM assistants on Oracle responses that know those states. The resulting models improve personalization and theory‑of‑mind metrics by up to 41 percentage points versus standard instruction‑tuned baselines.
The paper proposes Mixture of Memory Embeddings (MoME), a context‑aware sparse lookup that replaces each token’s single memory row with a gated mixture of multiple slots. Experiments on Llama‑3, MobileLLM and Qwen3 show MoME outperforms existing memory‑embedding baselines at equal parameter and FLOP budgets and exhibits interpretable routing for polysemous tokens.
ImpossibleRubrics introduces a 169‑task benchmark of “impossible” prompts, each with a formal oracle certificate defining what an honest answer can claim. The authors generate rubrics downstream and test them adversarially, finding that many rubric generators are exploitable (8‑36% of the time) and that a single generic rubric (“be decisive, penalize hedging”) is exploited 64% of the time, while…
The paper dissects which aspects of synthetic Thai OCR data (typeface diversity, 2‑D layout, real handwriting glyphs) actually help transfer to real documents. Using those insights they train Wayu-Paxa-OCR-Zero without any real OCR labels and achieve character error rates comparable to or better than supervised baselines.
