Text encoding changes byte-level language model accuracy by language
Aether uses script-specific encodings with UTF-8 fallback; Chinese and Korean gain, while Georgian and Tamil do not.
Stories per day · 30 days
4 RSS
Relatedagentdojo1agentic-ai1agent-security1asr1audio-language-model1belebele1benchmark1byte-level1catastrophic-forgetting1central-asian-languages1cer1classification1closed-book-qa1cultural-evaluation1culturalriddles1diffusiongemma1diffusion-language-models1distillation1
46 stories, newest first · one card per story
Stories tagged #multilingual
Otter adds transcription that follows language switches in meetings
Thirteen languages are live now and 14 more are coming, for 31 languages after rollout.
Saidzoda fine-tune lifts Belebele recall for four Central Asian languages
Belebele recall@1 rises 14.4 to 26.0 points, to 75.2%–77.7%, while English and Russian recall fall.
TOPSD limits teacher supervision to verified prefixes in distillation
It beats conventional on-policy self-distillation on math reasoning in ten languages, especially lower-resource ones.
Language-subspace interventions fail to fully capture language information in LLMs
Language-subspace interventions improve multilingual accuracy by only 1.60 points but sacrifice response-language fidelity in Qwen2.5-7B.
NVIDIA announces Nemotron 3.5 Transcribe and Granary v2
The post title describes promptable, adaptable multilingual transcription; the feed had no further details.
Component translation does not reliably change LLM refusal rates
On 190 harmful behaviours, only an exploratory Mistral Small 4 Spanish swap stayed significant after correction.
Homoglyph watermarks still favor scripts used by higher-resource languages
After copy-paste strips invisible characters, detection needs about 2 English or Persian code points but 93 Lao.
Speech LLM ASR expansion forgets source languages unless merges are selective
SSU-Wanda+TIES best preserved source languages in single-target expansion; recursive multi-target updates were highly order-sensitive.
Foyer study: logistic regression beats LLMs on inbox routing
On about 200 labels, logistic regression led every field, while the most accurate system ignored 18% of emails needing action.
Tenable detectors now score threats natively in 10 languages
A single model replaces translate-then-detect and is about 20–800 times faster per item and an order of magnitude cheaper.
RubatoCodec adds budgeted adaptive speech tokens and streaming decode
Code and weights are public; adaptive allocation beat fixed-rate coding on reconstruction and intelligibility at low event rates.
CulturalRiddles benchmarks LLMs on riddles from 62 cultures
Across 28 model configurations, even frontier systems reach at most 22% accuracy in some communities.
Qwen-Audio-3.1-ASR adds diarization and broader audio understanding
The family keeps 30 languages and 16 Chinese dialects, with file, streaming, and interactive variants.
Diffusion language models vary compute cost by task and language
Weighting math answers by decoding speed beat majority voting by up to 4.2 points at no extra cost.
Paired-residual removal improves multilingual LLM text embeddings
The inference-time method was tested on 18 languages and English bitext mining against more than 200 languages.
First-token cueing recovers closed-book facts LLMs fail to start
Prefilling only the gold answer's first token fixes a fifth to a half of multi-token English QA errors across 16 open models.
Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?
LLMs refuse more often in Indic languages when told they are under a safety test, including some benign requests.
Language Is a Security Surface for Agent Guards
Most smaller agent guards change safety verdicts when requests are translated, especially into rarer languages.
Equal CER, Unequal Damage: Severity-Weighted Character Error Rate for OCR
semCER weights OCR character edits by judged severity, matching readers better than standard CER.
Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning
LASEF finds English reasoning skeletons help smaller models slightly, but language choice is context-dependent.
The Missing Minimal Pair: Stereotype Evaluation in LLMs
Researchers propose dual minimal pairs and a mutual-information metric for more reliable LLM stereotype evaluation.
Exploiting Acoustic and Content-Oriented Speaker Verification Attacks Against Multilingual Voice Anonymization
Acoustic ASV attackers generally beat content-oriented ones against multilingual voice anonymization.
HINTT Submission to the 2nd MLC-SLM Challenge: Comparing Cascaded and Unified Approaches to Diarization and ASR
HINTT's cascaded diarization and Qwen3-ASR pipeline beat a unified speech LLM on MLC-SLM Task 1.