Typed-decision encoders answer a question about a text in one forward pass, but their behaviour outside English is rarely measured on a fixed yardstick. We release PolyRoute: 322 native-language support messages in 12 languages and 9 writing systems, 156 code-mixed or romanised messages (Hinglish, Spanglish, Arabizi), 967 optional native training rows, and a protocol for a three-way department-routing decision. Labels come from Qwen3.8-27B with a blind second pass (91% of candidates kept); an independent model, gemma4-31b, agrees on 97.2% of test items, and the largest character 5-gram similarity between any test text and any training text is 0.33. A leaderboard of 13 systems gives 61.2% for the English-root checkpoint, 77.3% for the shipped multilingual one and 96.6% for the best fine-tune; TF-IDF with logistic regression reaches 91.9%. On four withheld languages the three multilingual systems score 77%, 90% and 93%. On code-mixed text the English root (85%) is nominally ahead of the shipped multilingual checkpoint (79%; p=0.134). The split resolves differences only above about 4 points. All texts and labels are LLM-generated.
Typed Decision Models such as Jev, Clef and pplx-decider return a probability for each answer an application provides. Applications use them in many languages, but are compared mostly in English. We introduce XDecideBench, a multilingual benchmark of nine suites in 122 languages. On our twelve-language suite, three open models fine-tuned for typed decisions beat their Qwen base models in every language. On new test data with local base models, their accuracy gain is larger in languages with less web text than in those with more. They fix a similar share of base-model errors in both, so the larger gain may reflect weaker base models. The best model on the twelve-language suite is best on only two other suites. The best English model on each suite, chosen in hindsight, is significantly beaten in about a quarter of suite and language combinations. On Belebele, Decision-1 is best in English but falls far behind Jev in Burmese. A confidence threshold fitted on English to allow 5% errors gives a higher error rate on XNLI, even in German and French. XDecideBench shows that results from one suite or from English are a poor guide to how a model and its threshold behave in other languages.
Marek Šuppa, Jaroslav Kopčan
Comenius University in Bratislava · Cisco Systems · Kempelen Institute of Intelligent Technologies
An LLM router picks which model should answer each query. The appeal is that models fail on different questions. Whatever single model is best overall still gets some wrong, and another model in the pool gets many of those right. Getting that choice right every time is the ceiling, and a router is an attempt to approach it. However, recent work reports that routers do not get close. Across 21 routing methods on five benchmarks, sharply different designs land within a fraction of a point of each other, and all of them stay far below that ceiling. Learned routers often fail to beat simply always calling the strongest model. We ask what those missed questions have in common. We set fourteen models to answer all 294 questions, with 7 task types across 3 languages: Korean, English and Hindi. We ran the whole matrix twice, changing nothing, but 5.37% of the 4,116 model-question pairs came out scored differently anyway. Run-to-run movement like that is normal, and we argue that a small win does not show that routing did anything, ours or anyone else's. Counting an answer correct only when the model got it right in both runs, 29 questions on this matrix can be improved with routing. Every correct-answer count here is on that rule. Task type accounts for most of them: assigning each task type one model in advance, chosen once and never updated, improves 21 of the 29. Splitting each task type by language improves 2 more and leaves 6 of 294 unoptimized. That handful is what a learned router would have been built for, and it is smaller than the run-to-run movement above, which is a share of pairs rather than of questions. The static table we adopted answers 262 of 294 questions at $3.33 per run, against the best single model's 245 at $7.69. All of this is fitted and scored on the same 294 questions with no holdout.
Large Language Model (LLM) routers commonly rely on neural query embeddings, with larger encoders expected to better capture query intent and difficulty. Yet scaling Qwen2.5 encoders from 0.5B to 72B parameters brings little improvement in routing accuracy (Figure 1b), suggesting that small encoders may already capture the query properties needed for routing. We therefore investigate which properties matter and whether they can be extracted directly from text without a neural encoder. We introduce REGEXROUTE, a pipeline that uses sparse autoencoders (SAEs) to discover interpretable regular-expression (regex) features. Using unlabeled text, an LLM turns descriptions of grouped SAE latents into regex extractors and refines them to match latent activation patterns. These extractors supply numerical features to a lightweight routing head, eliminating neural encoding at inference (Figure 1a). Across four benchmarks, one fixed set of 128 features achieves 76.43% average routing accuracy, comparable to 76.41% for the strongest neural text encoder baseline, with much smaller latency and strong robustness. These findings establish explicit, interpretable text features as a practical basis for designing and understanding LLM routers.