Translating song lyrics while preserving singability requires satisfying musical constraints, including syllable counts, rhyme schemes, and syllable patterns, that standard machine translation ignores. We present a framework that exposes IPA-based phonetic analysis as a suite of composable Agent Skills, each defined by a natural-language description plus deterministic verification scripts that the orchestrating LLM invokes on demand. The agent extracts constraints from source lyrics, then translates through a multi-phase process, calling skills to verify syllable counts and rhyme and iteratively refine syllable-count mismatches. We formalize the three musical constraints, describe the IPA skill suite and agent architecture, and propose three evaluation metrics: Syllable Error Rate (SER), Syllable Count Relative Error (SCRE), and Adjusted Rand Index (ARI). Experiments across three frontier orchestrators, with component ablations, confirm that tool-verified translation dramatically improves constraint satisfaction and preserves rhyme structure substantially better than a supervised baseline, while requiring no task-specific training and no parallel corpora, and transferring to Spanish and Japanese targets with only a small language-specific syllabification adapter.

We introduce TokLens, an open-source toolkit for evaluating tokenizer quality across languages using six intrinsic metrics: fertility, characters per token, compression ratio, normalized sequence length, single-token retention rate, and cross-lingual parity. We evaluate 24 tokenizers from major LLM families across 15 typologically diverse languages and correlate these metrics with downstream performance. Our analysis reveals stark disparities: GPT-2 produces 56x more tokens per word in Japanese than in English, while newer tokenizers like Qwen2.5 and Gemma-2 reduce this gap to under 4x. No intrinsic metric predicts English benchmark performance after controlling for model size. However, on multilingual benchmarks (MMLU-ProX), linear mixed-effects models show that tokenizer metrics significantly predict per-language performance (STRR: 𝛽 = +5.7, z = 18.5, p < 0.001). A controlled experiment on the Qwen2.5 family further shows that languages with higher single-token retention rate exhibit steeper scaling slopes (𝜌 = 0.91, p < 0.001). These results indicate that tokenizer quality is significantly associated with multilingual LLM performance, though the evidence remains correlational and partially confounded with pretraining data composition.

Diffusion language models generate text by iteratively denoising all tokens in parallel, but when and where their hidden states encode whether the output will be functionally correct remains unknown.We present the first probing study of DLM internals, training linear classifiers on hidden states to predict functional correctness.Across two models (LLaDA-8B, Dream-7B) and four tasks, we find that DLMs uniquely accumulate correctness signal across denoising steps (AUC gains of 0.08–0.11 on reasoning tasks), absent in single-pass AR decoding. However, step-0 signal reflects prompt difficulty rather than diffusion-specific computation. Signal emergence is task-dependent: structural tasks show flat profiles while reasoning tasks show gradual buildup. The two models exhibit distinct layer dynamics, with LLaDA concentrating signal in upper layers while Dream redistributes toward lower layers. We further show that probe confidence can identify likely failures, enabling selective generation that avoids 36–98% of wasted compute.

Diffusion language models (DLMs) decode several tokens per step. However, the grammar-constrained decoders evaluated here, IG-CD and LAVE, verify candidate tokens sequentially within each denoising step and treat every violation as a rejection, a lookahead verification, or a fallback to autoregressive decoding. We observe that a violation is local to the first token the parser cannot consume, so the rest of the parallel step can be kept. We propose Dgrammar, a training-free decoder that repairs that frontier token in place under the same forward-pass logits, paired with a grammar-gated backoff schedule for the unmask batch size, asynchronous mask construction, and a bounded grammar-guided tail. On JSONSchemaBench medium with LLaDA-8B, Dgrammar improves schema validity by 8.6 percentage points over the strongest DLM baseline while reducing mean latency by 5.8x and eliminating timeouts. Validity and latency gains also hold on the other JSB splits and on Dream-7B, and experiments with a SMILES grammar demonstrate cross-grammar transfer.