Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit"think in English"is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.
When a tool-using agent is given the same task in a different language, does it still take the same steps? Multilingual evaluation rarely asks: it compares final answers and discards the actions. Yet those actions are the product: they fix cost and latency, decide how the system fails, and are the only auditable part of its behaviour. We make the action policy the measured object across 8 models, 6 parallel benchmarks and 41 languages (2.38M rollouts). The naive measurement fails: five confounds sit between raw trace similarity and any defensible claim, each able to flip a conclusion. Short traces score higher, empty traces score perfectly, unrelated traces agree by chance over half the time, the gap is capped by each model's reproducibility, and a model asked the same question twice in one language answers differently, leaving no baseline. We remove all five, and every correction makes the effect larger. Divergence proves structural, not sampling noise: it survives greedy decoding in every cell and stays flat as temperature rises, even as models grow less self-consistent. Normalised by their own reproducibility, four very different frontier models converge under greedy decoding, each keeping 71-73% of its action policy across languages, with model identity explaining only 5.7% of the variance. Below roughly 10B parameters it breaks down, and the ordering among smaller models is largely an artifact of a chance floor we measure by permutation rather than assume. Agents route non-English tasks through English; this pivot is causally load-bearing, confirmed by a pre-registered prediction across four models, and models will not abandon it when told to. Finally, a single trace-extraction regex, not the model, manufactured a multilingual failure: two worked examples raise one model's measured accuracy twenty-sixfold while its accuracy on readable outputs barely moves.
This work measures the capability that role assumes and finds it lacking under the protocol the role is usually deployed with, one-shot greedy authoring with no test-time reasoning.
Wenhui Chen, Jianlin Chen, Ziyao Lin et al.· 0 citations
Multilingual evaluations report accuracy at a single output-token cap, but languages need different numbers of tokens to express the same content, so the cap is a hidden experimental variable. We test whether the native-vs-translate gap on MGSM (German, Thai, Swahili) is a token-budget artifact for Qwen3-8B and Llama-3.1-8B-Instruct under four prompting strategies. The measured gap swings by up to 57 points across budgets, length normalization moves it by up to 38.9 points where the cap binds, and at tight caps normalization can reverse which strategy scores higher. We prospectively froze the sweep's three Qwen peaks and its near-zero value at 1024 and evaluated them on 540,000 independently hard-capped decodes: a second frozen family of six Holm-corrected tests rejects every null. The frozen test at $B^*=1024$ still fails to reject because native accuracy has already saturated there; above saturation, the residual difference is a strategy-performance gap, not an identified reasoning deficit. The same truncation channel prices a cost-ordered adaptation ladder: a cross-fitted Thai vocabulary extension closes 0.0 points of the gap at the frozen budget and 4.9 points where 19% of traces still truncate. A third frozen family varies only the announced budget at a fixed enforced cap; announcing 128 rather than 2048 tokens moves Thai native accuracy by 5.1 points, so accuracy is not a function of the enforced cap alone. A correct-emission timing identity computed from one long-cap run matches the three pre-specified MGSM peaks to 0.65 points and, in an exploratory Qwen-only analysis of three further benchmarks, tracks held-out items to 0.92 points, locating the peak exactly in five of seven cells. Treat the output cap as an independent variable and report accuracy across the budget regime, not at a single budget.
Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers'trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.
A. Algaba, Francesca Carlon, Lynn Delcon et al.· 0 citations
When a language model must choose one answer from a large space of equally valid options, a format clause --"Reply with JSON only"-- changes which answer it chooses. We re-run the One-Word Census (arXiv:2607.12796): 31 wide-answer-space category prompts asked of 44 models, now with the reply requested in JSON -- no schema enforcement, no constrained decoding, only the request. Convergence deepens sharply: on the unconstrained"Pick a word"prompt the modal answer rises from 41% to 64% of the pool and distinct answers fall from 52 to 36; mean answer-choice surprisal drops from 1.80 to 1.58 bits. The tax is progressive: six of 44 models move individually (BH-FDR q=.10), all toward the mode, led by the most distinctive models, while the conformist floor is immobile. It is a sharpener, not a re-indexer -- the plain-chat modal answer survives in 28 of 31 categories. Defaults are register-indexed: a within-run re-sample (n=20) finds JSON shifts 53% of a model's stable chat defaults, mostly back to the crowd, and installs defaults absent from chat (Claude Fable 5 answers"cerulean"for colour 0% of the time in chat, 100% in JSON). Full-battery controls reveal a register gradient: compression is significant and specific to the answer-delivery formats models are trained to speak (JSON -0.22 bits, p=.0002; XML -0.19, p=.002), absent for YAML and CSV, and reversed for an arbitrary bracket wrapper (+0.13, p=.009) -- weighing the mechanism toward tool-use post-training. Enforcing the schema at the decoder (response_format) compresses no further than the request (-0.03 bits): the collapse lives in the model's response to the register, not the decoder. Structured output is how software consumes language models, and that surface is served by a measurably more homogeneous model than the chat surface on which models are evaluated, compared, and chosen.
Teams measuring whether large language models (LLMs) recommend a brand face a reproducibility problem: ask the same question twice and the answer moves. Practice resamples each prompt a few times (commonly five) and averages, treating within-prompt resampling as the source of the noise. But a measured brand score moves for at least four separable reasons: within-prompt resampling, prompt paraphrase, model identity, and query language. We specify a crossed random-effects (generalizability-theory) decomposition that partitions the total variance of a response-level brand outcome into these four sources, and embed the components in a decision-study allocation that returns how many repeats, paraphrases, models, and languages to buy for a target reliability. We apply it to a fully crossed corpus of 12,933 LLM responses on 20 Central and Eastern European brands, 8 languages, and 3 models (GPT-5.2 and Gemini 3 Flash in parametric mode, Perplexity in grounded retrieval), with a stability subset of 1,435 cells resampled about five times. The outcome is per-response multilingual sentiment polarity. Query language is the largest systematic facet (26.5% of the variance of one response) against 1.5% for brand identity (ICC 0.0146), so a single AI answer carries almost no brand-discriminating signal. Once a cell term isolates pure resampling, resampling is 34.8% of variance and the brand-in-context interaction 29.6%; brand-by-language is 8.6% (a bilingual penalty) while brand-by-model and brand-by-prompt are near zero. Per unit of query budget, adding languages and models reduces relative-error variance far more than adding repeats: a repeat past the fifth reduces it by only 0.0003. Brand-ranking reliability stays low, near 0.01 for a single answer and about 0.36 at the full crossed design, so reliability is bought by spreading across languages and models, not by repeating one prompt.
D. Żatuchin· 1 citation· ⚡1
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduJun 26, 2026
To help robots do chores in places like homes and factories, a new approach from MIT uses one language model to clarify users’ instructions, then another to ignore irrelevant info.
New MIT research could lead to better materials for a fossil-fuel-free process for making the chemical that's essential to fertilizer and other products.
