Mmlu-Pro

Large language models generate text by sampling each token from a predicted distribution, and a temperature parameter sets how far the draw strays from the most probable tokens. Truncation samplers such as top-p and min-p discard the least probable tokens before the draw, so that sampling at high temperature stays coherent. Their reported accuracy gains come from temperatures of 1.5 to 3, while the defaults of deployed systems cluster between 0.6 and 1.0. Whether they change accuracy at those defaults, and for which models, has not been measured. We test thirteen open-weight models on GSM8K and MMLU-Pro at temperatures 0.7, 1.0, and 1.3 in one controlled pipeline, ten of them under eight decoding configurations. Six of the thirteen models lose 17 to 38 accuracy points on MMLU-Pro between 0.7 and 1.3, and the other seven lose at most 10. The lost accuracy comes from generations that run to the token limit or never state an answer. These results suggest that truncation samplers improve accuracy primarily when higher temperatures substantially degrade model performance. Where accuracy remains stable across temperatures, none of the tested truncation samplers improves on plain temperature sampling.

Adding answer options can lower multiple-choice scores without improving assessment validity. Turkish MMLU Pro examines this distinction using 12,000 Turkish-source questions across 58 sections. Each question retains its stem, five original options and source key, and receives five options copied from other questions in the same section. Sentence-embedding retrieval proposes candidates; a language model selects existing identifiers. Deterministic verification reconstructs all 60,000 additions. A 25-model calibration exposes scoring and generation-budget effects. Five evaluations produce source-key accuracies of 34.8%-81.4%. On 981 shared questions, one API-served model falls from 93.7% with five choices to 83.1% with ten; 102 of 115 lost correct responses select borrowed options. The decrease is 24.4 percentage points on heuristically flagged negative stems and 5.9 points elsewhere. A completed human-checked audit of 200 sampled questions, with undocumented reviewer tool use, yields 47 and 31 multiple-answer judgments across the two record sets, 25 of the latter unresolved. These records support concern about ambiguity, while their dependence and incomplete reviewer-method documentation limit validation. Because order and labels also change, the paired comparison measures augmentation as implemented. The contribution is a traceable construction and an analysis of its validity limits, not evidence that lower ten-choice scores measure knowledge better.

Although MMLU is widely adopted as a benchmark for calibrating general AI capabilities, we psychometrically demonstrate that its aggregate score primarily evaluates a model's factual retrieval capacity rather than its reasoning ability. By calibrating item difficulty for 1,000 open-weights language models over 14,042 MMLU test items using Item Response Theory, we show that evaluating both abilities via a single test is inherently flawed. Difficulty is then regressed on a deterministic, text-extractable framework of structural complexity. Applying a joint Wald test with subject-clustered covariances demonstrates that the MMLU conflates fundamentally separable constructs. The mapping from structural complexity to difficulty is not invariant across the benchmark's STEM and non-STEM partitions. This finding has practical consequences. Aggregate leaderboard ranks track non-STEM accuracy more closely than STEM accuracy, so selecting a Top-50 model on the aggregate for a reasoning-intensive deployment displaces roughly 22% of the STEM-appropriate choices. Furthermore, when controlling for the multiple-choice guessing floor natively inside the response model, we find that higher-ability models continue to degrade more steeply under increased reasoning depth. The MMLU aggregate therefore weights retrieval capacity and reasoning stability unequally, inadvertently favoring models optimized for retrieval. We release our deterministic framework as a reproducible auditing instrument and recommend disaggregated reporting.

Universal machine-learning interatomic potentials (u-MLIPs) aim to generalize across diverse configurations. Benchmarks enable reproducible evaluation but may not expose failures outside their predefined scope. Here, we show that physics-informed search can complement benchmark-based evaluation by uncovering hidden failure modes. We introduce MLIP Detective, an agentic framework for active failure mode discovery. Starting from benchmark evidence, MLIP Detective generates falsifiable, physics-informed failure hypotheses, screens them with inexpensive simulations, and escalates only the most suspicious cases to human experts together with proposed verification protocols. Without issue-specific prompting, MLIP Detective identified and characterized a systematic anomaly in MACE-MPA-0: the model predicted some relaxed adsorbate-surface systems involving O- or F-containing adsorbates to be higher in energy than their corresponding separated fragments. Using cross-model comparisons, MLIP Detective further inferred a likely training-data origin for the anomaly, consistent with recent reports.

Mixture-of-Experts (MoE) models decouple total parameters from per-token compute, but deployment still requires storing every expert. Recent theory shows that pruning experts with the smallest router-norm changes during fine-tuning can preserve accuracy, but assumes full fine-tuning. We test whether lightweight adaptation can recover this signal. We briefly fine-tune with a parameter-efficient adapter, rank experts by the induced $\ell_2$ router change, and prune the least-changed experts in one shot. On Mixtral-8$\times$7B-Instruct (44.83% MMLU-Pro), router-only LoRA trains 0.002% of parameters and outperforms all-module LoRA at matched rank with half the experts removed (27.54% vs. 24.42%); signal quality declines as adaptation spreads to attention and expert weights. Accuracy improves monotonically with LoRA rank, reaching 28.76%. IA3, which leaves router weights frozen, matches direct router adaptation, whereas unconstrained additive adapters degrade the signal. Router-guided MMLU-Pro accuracy decays quasi-linearly rather than collapsing, remains nearly 1.8 times that of magnitude-based or random pruning at maximal compression, and reduces memory by 49% and per-token latency by 37%. At 25% compression, retention is competitive with methods using full activation statistics. The criterion also transfers to Qwen1.5-MoE fine-tuned for mathematics, retaining 49.7% mean accuracy over eleven benchmarks with half the experts removed while random pruning falls to single digits. Router sensitivity under lightweight fine-tuning therefore makes provably motivated expert pruning practical at scale.

Large language models often reason at length before answering, increasing cost and latency. Prompts and trained settings can shorten this reasoning, but a shorter trace may only show that the model stopped sooner. Here, we evaluate paired runs of the same question at matched reasoning horizons across 198 GPQA Diamond and 500 MMLU-Pro questions. We test a numeric/concision prompt that announces a token limit for Qwen3-14B and the trained effort settings of gpt-oss-20b and -120b. The Qwen prompt shortens reasoning traces by 12-17%, while accuracy changes at matched token limits are small and mixed. A concise/early-answer instruction raises MMLU-Pro accuracy by 3.8 percentage points at 512 tokens, including +2.7 points when both runs are unfinished. Its gain at 2,048 tokens is uncertain. For gpt-oss, candidate-logit answers from completed low- and medium-effort reasoning are 14.5-26.3 points more accurate than matched-horizon high-effort answers. Most of the 512-token advantage comes from lower effort finishing earlier, while differences among unfinished runs are smaller and mixed. Wrong early answers often concentrate probability on the chosen option, so earlier stopping does not uniformly improve probability quality. In these tests, a tight deadline can favor lower effort or a concise instruction, whereas allowing high effort to finish can recover higher final accuracy. Evaluations should report correct completion before a deadline, the answer obtained when a run is stopped, differences among unfinished runs, and probability assigned to the correct answer separately.

Recent advances in language models have enabled collaborative settings in which multiple models leverage one another's capabilities, iteratively improving, transforming, and extending each other's outputs. Each agent sees what the others assert before it answers, so peer opinion competes with the model's own parametric knowledge, and a wrong majority can overturn an answer the model would otherwise get right. We measure that displacement in 23 open-weight models, 19 conditions, and three datasets, yielding more than a million graded responses. A unanimous wrong majority reverses 22.8% of a model's correct MMLU answers, 54.8% on GPQA and 71.0% on SimpleQA, and 84-89% of the reversed answers match the peers' answers. Existing mitigations aim to increase Resistance, the rate at which a model keeps its correct answer under this pressure, which is only half of what a collaborating agent needs. We pair it with Receptivity, the rate at which a model adopts a correct peer answer after initially answering incorrectly. We score six methods on both axes, four drawn from prior work and two of our own. Each gains Resistance only by losing Receptivity, and their means fall on a single Resistance-Receptivity frontier with $R^2$ between 0.80 and 0.90. Reflection, the strongest published method, gains 7.9 points of MMLU Resistance and gives up 15.3 of Receptivity. Reasoning is the one exception. On GPQA and SimpleQA it trades like the rest, but on the MMLU subjects whose answers a model can derive for itself it raises Resistance by 7.2 points and Receptivity by 9.6 at once, the only intervention we find that improves both.

We examine how prompt tone affects both accuracy of the LLM answers and inference cost as reflected in output-token consumption. Experiments were performed to understand the trade-offs between accuracy and inference cost on a 570 Question MMLU dataset for LLM models prompted in seven different tones from sycophantic to threatening. Our results show that the output-token-length variation substantially exceeded accuracy variation across all models. Output-token consumption varied by up to 44.3% across tone conditions. We also analyzed the tradeoff between the accuracy of the answers and the average output token length in the reasoning process. For the ChatGPT models 4o and 5-nano, the rude tone is quite dominant. For the Gemini models 2.5 Flash and 2.5 Flash Lite, the rude and neutral tones are dominant on the Pareto-optimal frontier. We find that prompt tone influences not only answer quality but also the amount of billable inference resources consumed by modern LLMs.

This paper reports on a collaboration between the Directorate-General for Translation (DGT) and the European Master's in Translation (EMT) to localise the MMLU dataset into 11 European languages. Beyond creating a more inclusive benchmark for LLM evaluation, the project offers master's students authentic, project-based professional training in translation, revision, project management, and multilingual coordination, while highlighting key methodological, administrative, and workflow challenges.

The dominant practice in language model evaluation is to report a single accuracy number per model and declare the higher one better, without testing whether the gap could plausibly be sampling noise. On benchmarks of a few thousand items, and under temperature sampling where a model can differ from itself run to run by more than the reported gap between models, this practice routinely overstates confidence in headline claims. The statistical machinery to fix this -- confidence intervals, paired significance tests, power analysis, clustered standard errors, multiple-comparison correction -- is well established, but no standard, pip-installable tool packages it in the shape an evaluation actually takes: a per-item results table. We present evalci, a pure-Python library (numpy/scipy/pandas only) that turns a per-item results table into a publication-ready claim -- e.g., "Model A beats Model B, $Δ=3.1$ pts, 95% CI [1.2, 5.0], paired permutation $p=0.002$, $n=1{,}319$" -- in one function call, with adapters for lm-evaluation-harness and HELM output. Every routine is validated against an independent reference (statsmodels, or brute-force exact enumeration) rather than only against itself. As a case study, we re-analyze a public comparison of nine language models' MMLU accuracy and find that 3 of the 8 adjacent leaderboard-rank gaps are not statistically significant after correcting for the 36 pairwise comparisons the ranking implies. evalci is available at https://pypi.org/project/evalci/ (source: https://github.com/Shreyaskc/evalci, DOI: https://doi.org/10.5281/zenodo.21201815)

Sparse Mixture-of-Experts (MoE) models route each token to a few of many experts, inviting the hypothesis that experts form functional modules tied to capabilities or languages. We test this causally on Command A+, a frontier open-weights MoE (218B total / 25B active; 128 experts, 8 active, +1 shared). We build a routing-mass atlas, pre-register six family-to-axis hypotheses before any intervention, and ablate each family at inference time against a size-matched random-expert null, measuring whether it selectively breaks its own axis (worst off-target effect at most one third of on-target). Crucially, we test the same families under four metrics and a held-out, independent-corpus run with bootstrap confidence intervals. Our finding is cautionary: robust functional modularity is rare and measurement-dependent. Of six pre-registered families, only one, the Arabic-language family, is a clean selective module that survives an independent corpus and a conservative statistical bar (1/6; a more permissive pre-registered point rule admits 3/6, but that count is threshold-sensitive). Every other family has a real causal effect yet fails selectivity, and its apparent modularity flips with the measurement: with the corpus, the metric, and the statistical bar. A positive control on Qwen3-30B-A3B recovers its published disjoint structure, confirming the method detects modularity when present. The verdict reproduces on the un-quantized BF16 model, ruling out a 4-bit quantization artifact. We conclude that ablation-based modularity verdicts are not safe unless the corpus, metric, and statistical bar are controlled. We release the atlas and ablation data.

We use training-data attribution as an interpretable tool for capability discovery, mapping which regions of the pretraining corpus support social-reasoning versus STEM-reasoning in OLMo3-7B. Training-data attribution measures how strongly each training document influences a model's predictions on a benchmark, but document-level scores are too noisy to identify which corpus regions support which capabilities, and prior work has emphasized factual knowledge rather than reasoning. We compute gradient-based attribution (TrackStar via Bergson) over a working set drawn from the de-duplicated Dolma3 mix, aggregate influence across WebOrganizer's 24-format x 24-topic taxonomy (576 bins), and contrast benchmark pairs in a 2x2 design that varies domain (social vs. STEM) and capability type (reasoning vs. knowledge): SocialIQA and MMLU Social Sciences against ARC-Challenge and MMLU STEM. Social and STEM reasoning draw on qualitatively distinct corpus regions, and the contrast is sharper at the reasoning level than at the knowledge level. Targeted machine unlearning provides partial causal validation: forgetting high-attribution topic bins (e.g., Literature for SocialIQA) degrades the aligned benchmark more than within-bin random baselines, and we open-source all code, sampling manifests, the bin-level influence matrix, and unlearning checkpoints.

Sparse autoencoders (SAEs) are typically trained to reconstruct the \textbf{entire} residual stream through a sparse dictionary, implicitly assuming that all activation content is amenable to sparse, monosemantic decomposition. We question this assumption and hypothesize that activations contain a low-rank, dense component that is computationally important to the model yet inherently unsuitable for sparse representation, which serves as a major source of the persistent dense latents widely observed in trained SAEs. To test this, we add a small rank-$r$ linear bottleneck in parallel with standard SAEs (BatchTopK and Matryoshka), allowing dense structure to be absorbed before sparse reconstruction. On Gemma-2-2B layer 12, a rank-24 bottleneck reduces dense latent count by up to 84% while improving sparse probing and targeted probe perturbation on both architectures at matched sparsity. The absorbed component is (i) \textbf{structurally identifiable} as the top principal components and outlier dimensions; (ii) \textbf{causally necessary}, with removing it raising next-token cross-entropy by 7.5$\times$, far exceeding the 2.8$\times$ from removing the geometrically near-identical top-24 PCA directions; and (iii) \textbf{redundantly encoded by sparse dictionaries}, with ablating 787 maximally aligned sparse features raising cross-entropy by only 2.9$\times$ and ablating 2,048 topic-aligned features leaving MMLU topic classification virtually unchanged, whereas removing the scaffold drops it from 98.7% to chance. Together, our findings identify a compact, semantically informative and causally important component of residual stream activations (which we term a \textbf{computational scaffold}) that standard sparse dictionaries represent inefficiently, suggesting that the scope of sparsity-based interpretability methods warrants careful re-examination.

Across two public LLM leaderboards, many displayed pairwise rankings do not meet a conventional paired-test resolution target under the actual paired evaluation design: 11 of 40 Open LLM Leaderboard v1 pairwise comparisons and 4 of 9 MMLU-Pro top-10 adjacent-rank pairs are unresolved at (alpha, 1-beta) = (0.05, 0.8). The MMLU-Pro count rises to 6/9 under real subject-level clustering and stays at 5-6 out of 9 in 99.9% of category-bootstrap resamples. We frame paired LLM evaluation as a hypothesis-testing problem, invert level-alpha, power-(1-beta) tests, and report a per-pair resolution ratio q = N/N* as the primary diagnostic. A sharp small-effect expansion with an explicit second-order constant shows that the widely-used unpaired Cohen-h-plus-(1-rho) shortcut deviates from the correct N* by approximately a factor of two in the close-comparison regime, a deficit that three of five off-the-shelf calculators(Cohen 1988, G*Power, R pwr) silently inherit when the user post-multiplies their per-arm output by (1-rho). The unresolved-pair pattern remains under multiplicity correction and anytime-valid sequential testing.

The guidance of scaling laws has increased the resource demands of modern large language models (LLMs), yet it remains questionable whether these models utilize resources effectively under a fixed budget. Previous research has proved superposition as a key contributor to loss. By leveraging the Neural Feature Ansatz, we extend superposition from parameter space to gradient space and define it as neural interaction. We find that under a fixed budget, good generalization is usually accompanied by efficient neural interactions, and the model can be placed in an efficient interaction interval by adjusting its depth-width ratio ($R_{D/W}$). In addition, as the budget scales up, the efficient interaction interval of the model remains relatively stable. By comparing existing small scale dense LLMs, we observe that models operating near this interval tend to perform better on the MMLU-Pro benchmark. Our findings reveal that the $R_{D/W}$ influences resource utilization efficiency and thereby affects generalization, providing insights into model shape initialization and the understanding of model generalization mechanisms. Code for Neural Interaction Law is available at: https://anonymous.4open.science/r/Neural_Interaction_Law-D788

Recent advancements in large language models (LLMs) have shown their remarkable capacities in many NLP tasks. However, their substantial size often presents challenges for deployment. This necessitates efficient techniques for model compression, with quantization emerging as a prominent solution. Despite its benefits, the exact impact of quantization (from 2- to 6-bit) on the performance and accuracy of LLMs remains an active area of research. This paper investigates the performance of eight LLMs at various quantization levels, focusing on tasks such as MMLU-Pro for knowledge processing and reasoning, CRUXEval for code comprehension, and MuSR for reading comprehension. Our results show a consistent trend where higher precision (e.g., 8-bit Q8_0) yields improved performance, albeit with diminishing returns. Aggressive quantization (e.g., 2-bit Q2_K) usually retains acceptable accuracy, though some models show a substantial loss in performance. Our findings indicate that while lower bit precision generally reduces performance, the impact varies across models and tasks. Larger models show greater resilience to aggressive quantization, but can still undergo significant drops at lower precision levels. Mid-sized models in the 7-9 billion parameter range strike an optimal balance between efficiency and resource usage. Such results provide insights into the trade-offs between model size, quantization, and performance.

LLM-as-a-Judge evaluation has become a standard tool for assessing base model performance. However, characterizing performance via the naive estimator, i.e., raw judge outputs, is systematically biased. Recent work has proposed estimators to correct this bias, but their reliability depends critically on judge quality and, for model comparisons, on calibration stability. Sharing calibration across compared models is practically attractive but can introduce severe bias, including cases where the comparison estimate points in the wrong direction with high apparent confidence. We study these failure modes through analytical results, simulations over judge quality ($J$) and cross-model calibration instability ($ΔJ$), and a real-data MMLU-Pro case study with sign reversal. We propose $J$ and $ΔJ$ as diagnostics for when corrected estimates, especially shared-calibration comparisons, are likely unreliable, and provide reporting guidance for LaaJ evaluation.

Can linearly decodable failure signals in LLM hidden states be leveraged to correct those failures? We investigate this classification-correction gap via Overthinking (OT)--a stable behavioral regime (Jaccard >= 0.81, 94% inter-annotator agreement) in medical QA where models answer correctly under resampling yet fail in extended chain-of-thought. OT is linearly decodable at 71.6% balanced accuracy (p < 10^{-16}). Yet five families of fixed linear steering (29 configurations, n=1,273) all yield Delta ~= 0, with identical null results cross-architecture (Qwen2.5-7B) and cross-domain (MMLU-STEM). Three convergent lines of evidence suggest representational entanglement: the OT direction has 85-88% overlap with task-critical computation (specificity ratio <= 0.152); non-targeted shared-direction steering damages accuracy (-12.1pp); and LEACE concept erasure damages accuracy (-3.6pp, p=0.01), while 10 random erasures produce Delta=+0.3pp. The per-instance probe-steering correlation is r=-0.002 (p=0.97). Positively, the same probe enables selective abstention (held-out AUROC=0.610, exceeding all five uncertainty baselines, p=0.009): decodable failure structure supports post-generation reliability estimation even when the fixed linear steering family cannot exploit it for correction.

Social-educational robots designed for socially interactive pedagogical support, such as the Robot Study Companion (RSC), rely on responsive, privacy-preserving interaction despite severely limited compute. However, there is a gap in systematic benchmarking of language models for edge computing in pedagogical applications. This paper benchmarks 25 open-source language models for local deployment on edge hardware. We evaluate each model across three dimensions: inference efficiency (tokens per second, energy consumption), general knowledge (a six-category MMLU subset), and teaching effectiveness (LLM-rated pedagogical quality), validated against five independent human raters using the Raspberry Pi(RPi)4 as the primary platform, with additional comparisons on the RPi5 and a laptop GPU. Results reveal pronounced trade-offs: throughput and energy efficiency vary by over an order of magnitude across models, MMLU accuracy ranges from near-random to 57.2%, and teaching effectiveness does not correlate monotonically with either metric. Among the evaluated models, Granite4 Tiny Hybrid (7B) achieves a strong overall balance, reaching 2.5 tokens per second, 0.90 tokens per joule, and 54.6% MMLU accuracy; high MMLU accuracy does not appear necessary for strong teaching scores. Human validation on four representative models preserved the automated rank ordering (Pearson r = 0.967, n = 4). Based on these findings, we propose a three-tier local inference architecture for the RSC that balances responsiveness and accuracy on resource-constrained hardware.

Prevailing machine-learned interatomic potential (MLIP) uncertainty-quantification methods rely on ensembles of independently trained backbones. These methods scale unfavorably with foundation-scale MLIPs, and their member-disagreement signals correlate weakly with per-molecule prediction error. Here we probe the frozen per-atom representations of a pretrained MLIP with a compact discriminative classifier, recasting MLIP uncertainty quantification as selective classification rather than error regression. The resulting method, PROBE (Post-hoc Reliability frOm Backbone Embeddings), produces a per-prediction reliability probability that monotonically tracks actual error without modification to the underlying model. Across large held-out evaluation sets and two structurally distinct MLIP architectures, PROBE outperforms ensemble disagreement as a binary reliability signal, which strengthens with the expressiveness of the backbone representation, implying a favorable scaling trajectory toward foundation-scale MLIPs. Multi-head self-attention additionally yields per-atom importance maps, providing chemically interpretable diagnostics at no additional computational cost. PROBE is post-hoc and architecture-agnostic, and is directly deployable on any MLIP that exposes per-atom representations.

Compositional machine-learning (ML) systems assemble runtime behavior from libraries of independently re-trained capability modules. Replacing one module raises a regression-testing question that static dependence analysis cannot answer: which existing compositions stay valid, and at what test cost? We frame capability updates as regression test selection (RTS) and contribute four results. First, a paired cross-version swap protocol isolates the marginal effect of a single module update. Second, on two contact-rich manipulation tasks we characterize a dominant-skill effect: one capability module reaches 88.0% atomic success while siblings stay at or below 32.0%, and its inclusion shifts composition success by up to 52 percentage points; a controlled weight-space interpolation tracks composition success against atomic quality point-by-point (pooled Pearson r=0.94), and the effect replicates on a second task, where the governing module must lie on the critical path of the phase sequence. Third, off-policy behavioral-distance metrics fail to identify the dominant module. Fourth, a margin-gated Hybrid Selector matches full revalidation at zero per-decision test cost (75.0% gold-label agreement, with no detectable difference) and reaches 81.25% match at half of full-revalidation cost, beating a cost-matched random budget (Monte-Carlo p=0.039). A resolution analysis shows that coarse evaluation overstates the apparent advantage of full revalidation. The atomic-quality probe gives a principled test-selection criterion for capability-update regression testing in compositional ML systems.

A predecessor pilot (Cacioli, 2026) found that Llama-3-8B implements prompted sandbagging as positional collapse rather than answer avoidance. However, fixed option ordering in MMLU-Pro left open whether this reflected a model-level position-dominant policy or dataset-level distractor structure. This pre-registered follow-up (3 models, 2,000 MMLU-Pro items, 4 conditions, 24,000 primary trials) added cyclic option-order randomisation as the critical control. The pre-registered item-level same-letter diagnostic did not confirm deterministic position-tracking (same-letter rate 37.3%, below the 50% threshold). However, pre-specified supporting analyses revealed that the response-position distribution under sandbagging was highly stable under complete content rotation (Pearson r = 0.9994; Jensen-Shannon divergence = 0.027, compared to 0.386 between honest and sandbagging conditions). Accuracy spiked to 72.1% when the correct answer coincidentally occupied the preferred position E, and fell to 4.3% at position A. The data provide strong evidence for a soft distributional attractor: under sandbagging instruction, the model enters a low-entropy response-position basin centred on E/F/G that is highly stable and largely content-invariant at the aggregate level. Qwen-2.5-7B served as a negative control (non-compliant, no distributional shift). These results provide evidence, at the 7-9 billion parameter scale, that response-position entropy is a promising black-box behavioural signature of this sandbagging mode.

As Large Language Models (LLMs) are deployed globally, ensuring their safety and alignment across multiple languages becomes paramount. However, safety behaviors often vary unpredictably between languages, posing significant challenges for consistent and ethical AI. In this work, we systematically investigate the dynamics of multilingual alignment, exploring whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities. We introduce RefusEU, a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models. Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.

Aggregate metacognitive quality scores mask within-model variation across MMLU benchmark domains. We administered 1,500 MMLU items (250 per domain, under an a priori six-domain grouping) to 33 frontier LLMs from eight model families and computed Type-2 AUROC per model-domain cell using verbalized confidence (0-100). Total observations: 47,151. Every model with above-chance aggregate monitoring showed non-trivial domain-level variation. Applied/Professional knowledge was reliably the easiest benchmark domain to monitor (mean AUROC = .742, ranked top-2 in 21 of 33 models); Formal Reasoning and Natural Science were reliably the hardest (one of the two ranked bottom-2 in 27 of 33 models). The three middle domains were statistically indistinguishable (Kendall's W = .164). A subject-level coherence analysis (within-domain similarity ratio = 0.95) confirms the six-domain grouping is a pragmatic benchmark taxonomy, not a validated latent construct. Within-family profile-shape clustering is significant for Anthropic, Google-Gemini, and Qwen (permutation p < .0001) but not DeepSeek, Google-Gemma, or OpenAI. Gemma 4 31B showed a +.202 AUROC improvement over Gemma 3 27B. Three models classified Invalid on binary KEEP/WITHDRAW probes produced normal profiles under verbalized confidence, confirming probe-format specificity. Bootstrap 95% CIs on 198 cells have median width .199. Split-half aggregate stability r = .893; profile-level split-half is weaker (grand median r = .184). These results show stable benchmark-domain variation obscured by aggregate metrics, and support benchmark-stage domain screening as a step before deployment in specific application areas.

While self-consistency is known to improve performance on symbolic reasoning, its effect on the recall of encyclopedic knowledge is unclear due to a lack of targeted evaluation grounds. To address this, we establish such a knowledge recall split for the popular MMLU benchmark by applying a data-driven heuristic from prior work. We validate this split by showing that the performance patterns on the symbolic reasoning and knowledge recall subsets mirror those of GSM8K and MedMCQA, respectively. Using this solid ground, we find that self-consistency consistently improves performance across both symbolic reasoning and knowledge recall, even though its underlying CoT prompting is primarily effective for symbolic reasoning. As a result, we achieve an 89% accuracy on MMLU, the best performance to date with the use of GPT-4o.

Machine unlearning for large language models (LLMs) aims to remove targeted knowledge while preserving general capability. In this paper, we recast LLM unlearning as an asymmetric two-task problem: retention is the primary objective and forgetting is an auxiliary. From this perspective, we propose a retention-prioritized gradient synthesis framework that decouples task-specific gradient extraction from conflict-aware combination. Instantiating the framework, we adapt established PCGrad to resolve gradient conflicts, and introduce SAGO, a novel retention-prioritized gradient synthesis method. Theoretically, both variants ensure non-negative cosine similarity with the retain gradient, while SAGO achieves strictly tighter alignment through constructive sign-constrained synthesis. Empirically, on WMDP Bio/Cyber and RWKU benchmarks, SAGO consistently pushes the Pareto frontier: e.g., on WMDP Bio (SimNPO+GD), recovery of target model MMLU performance progresses from 44.6% (naive) to 94.0% (+PCGrad) and further to 96.0% (+SAGO), while maintaining comparable forgetting strength. Our results show that re-shaping gradient geometry, rather than re-balancing losses, is the key to mitigating unlearning-retention trade-offs.

Meaningful multilingual evaluation must test models in the target language and educational context. Urdu, spoken by more than 230 million people, lacks a broad MMLU-style benchmark built from native educational sources. We introduce UrduMMLU, a benchmark of 26,389 Urdu MCQs across 26 subjects and five domains, collected from native Urdu MCQ banks and public examination PDFs. Unlike translation-based benchmarks, UrduMMLU combines academic subjects with content specific to Urdu and regional education. We label the exam-derived portion through dual human annotation with strict consensus filtering. We evaluate 30 LLMs under English and Urdu prompts, yielding 60 zero-shot evaluations, and further evaluate four open-source LLMs under multiple few-shot settings across both prompt languages. Gemini-3.5-Flash performs best, reaching 90.23% and 90.45% accuracy, while no other model exceeds 85%. The strongest open-source model trails by 7.78 and 9.12 points, and many models lose 25 to 40 points on Urdu-centered Humanities subjects compared with STEM. Few-shot prompting yields only modest gains. Results on UrduMMLU show that current LLMs have uneven Urdu knowledge, particularly for content grounded in the regional context.