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.
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.
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.
A benchmark score comes from a single phrasing of each problem. That single phrasing is treated as if it stood for the whole space of ways the same problem could be asked, but it does not. We show that rephrasing a problem while keeping its meaning and answer fixed routinely flips a model's answer in both directions, so some failures become successes and some successes become failures. We call this drift. BenchDrift generates meaning-preserving variations of benchmark problems along four axes, namely linguistic, referential, pragmatic, and structural, and measures how often, and why, correctness flips under each. Across eight models and three benchmarks (GSM8K, MMLU, MATH-Hard), we observe that drift is large in both directions. Two findings stand out. First, phrasing sensitivity does not fade as models get better. Instead, it changes sign. Weak models gain more from rephrasing than they lose, while strong models lose far more than they gain. We find that the best models on a benchmark are therefore the ones whose scores depend most on the wording they happened to be given. Second, the models largely agree on which rephrasings cost the most correct answers even though they differ in how much they drift, so fragility belongs to the rephrasing and not to the model. Furthermore, rephrasing breaks answers a model was confident about, whether the problem is made shorter or longer. Code and Data: https://github.com/IBM/BenchDrift/tree/demo-ui
Shailja Thakur, Sungeun An, Chad DeLuca et al.· 0 citations
DCASE~2026 Task~5 introduces Audio-Dependent Question Answering (ADQA), which tests whether large audio-language models answer from the audio rather than from textual priors. An Audio-Dependency Filtering (ADF) pipeline combines silent-audio probing, per-option perplexity, a large language model (LLM) commonsense check, and human review to remove items solvable from text alone. The 3000 items that pass form the ADQA-Bench evaluation set, spanning music, speech, and environmental audio. The inaugural edition draws 14 teams and 36 submissions across two tracks defined by total parameter count (up to 100B and under 10B). A Chung-Ang University ensemble of MOSS-Audio-8B-Thinking and Qwen3-Omni-30B reaches the top overall accuracy at \pct{58.33}, and a MOSS-only configuration from the same team leads the sub-10B track at \pct{57.30}. Across the 30 submissions with a comparable development score, evaluation accuracy falls by 11.91 percentage points (pp) on average (median 10.91\,pp) on the hidden evaluation split, which is designed to be harder than the development split. The most common building blocks are: the MOSS-Audio-8B-Thinking backbone (13 of 36 submissions), Low-Rank Adaptation (LoRA) fine-tuning on AudioMCQ-StrongAC, and preference or reinforcement-learning objectives -- Group Relative Policy Optimization (GRPO) in five teams, Group reward-Decoupled Normalization Policy Optimization (GDPO) in two. At test time, prompt engineering is near-universal, and majority or choice-permutation voting is common. Every system misses the same set of 233 evaluation items.
Haolin He, Renhe Sun, Zheqi Dai et al.· 0 citations
LLM debate is usually evaluated by final answers, yet transcripts reveal whether later turns develop new arguments or return to earlier claims in new wording. We study this process with \textit{prior-argument similarity}, which compares extracted argument units with earlier units in the same debate. In controlled eight-turn debates over 71 motions, six languages, and four model agents, Chinese is the only tested language with a consistently positive gap relative to English across three multilingual embedding models. The gap persists across agents, turn positions, regression adjustment, metric variants, extraction-length controls, a second-extractor subset, and cross-encoder tail rescoring. Manual calibration shows weak item-level alignment but a high-similarity tail enriched for substantive repetition. A diversity-aware prompt lowers \textit{prior-argument similarity} across languages, yet does not significantly narrow the Chinese--English gap. Multilingual debate evaluation should therefore measure argumentative development over time and report both average and gap terms.
Test-time scaling (TTS) reliably improves reasoning in large language models, but whether it transfers to small open vision-language models remains unclear. We examine this on EXAMS-V, a multilingual visual multiple-choice benchmark, comparing self-consistency, describe-then-reason with PRM-guided beam search, and two post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. What matters is the conditions under which TTS runs, not the search or verification machinery. The largest factor is parseability: an early prompt format left many chains reasoning correctly yet never committing to an answer letter, which a standard answer cue and a guided repair step largely remove. A larger decoding budget removes the rest: raising the per-chain token limit from 1k to 2k recovers 3.7 pp, whereas sampling more chains (8 to 16) adds only 0.15 pp. Once chains have room to finish, elaborate methods contribute little: PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost, and neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies. The largest gain comes instead from the policy model itself (+11.4 pp). Our best configuration reaches 84.1% on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.