LLM Evaluation

Measuring the creativity of large language models (LLMs) is essential for designing methods that can improve creativity and for enhancing our scientific understanding of this ability. To accomplish this, it has become common in recent years to administer tests of human creativity to LLMs. Although these tests provide a convenient and fully automated way to score "creativity," their validity as measures of machine creativity has not been established, and these tests already have limited validity as predictors of human creativity. To address this problem, we conduct the first large-scale, systematic study assessing the effectiveness of human creativity tests for predicting the creative achievement of LLMs across three target constructs: creative writing, divergent thinking, and scientific ideation. We find that the Divergent Association Task (DAT) and the Conditional DAT are the best predictors of creative writing and divergent thinking, respectively, but that test effectiveness varies significantly by construct, and no single test predicts all constructs well. Moreover, contrary to popular belief, no existing test reliably predicts scientific ideation ability. Motivated by this problem, we introduce the Divergent Remote Association Test (DRAT), a vocabulary-space test that assesses both convergent and divergent thinking in a single instrument. The DRAT is the first and only creativity test for LLMs that is a significant predictor of scientific ideation ability, demonstrating robustness across major design choices. Furthermore, the performance gain of the DRAT is not recoverable from any linear combination of the Divergent Association Task and the Remote Associates Test, indicating that assessing divergent and convergent thinking in the same test is essential to reliably predicting scientific ideation ability.

Off-the-shelf large language models (LLMs) are increasingly used to automate text annotation, yet their effectiveness remains underexplored for underrepresented languages and specialized domains where the class definition requires subtle expert understanding. We investigate LLM-based annotation for a novel legal NLP task: identifying the presence and sentiment of credibility assessments in asylum decision texts. We introduce RAB-Cred, a Danish text classification dataset featuring high-quality, expert annotations and valuable metadata such as annotator confidence and asylum case outcome. We benchmark 21 open-weight models and 30 system-user prompt combinations for this task, and systematically evaluate the effect of model and prompt choice for zero-shot and few-shot classification. We zoom in on the errors made by top-performing models and prompts, investigating error consistency across LLMs, inter-class confusion, correlation with human confidence and sample-wise difficulty and severity of LLM mistakes. Our results confirm the potential of LLMs for cost-effective labeling of asylum decisions, but highlight the imperfect and inconsistent nature of LLM annotators, and the need to look beyond the predictions of a single, arbitrarily chosen model. The RAB-Cred dataset and code are available at https://github.com/glhr/RAB-Cred

Personalisation is a standard feature of conversational AI systems used by millions; yet, the efficacy of personalisation methods is often evaluated in academic research using simulated users rather than real people. This raises questions about how users and their simulated counterparts differ in interaction patterns and judgements, as well as whether personalisation is best achieved through context-based prompting or weight-based fine-tuning. Here, in a large-scale within-subject experiment, we re-recruit 530 participants from 52 countries two years after they gave their preferences in the PRISM dataset (Kirk et al., 2024) to evaluate personalised and non-personalised language models in blinded multi-turn conversations. We find preference fine-tuning (P-DPO, Li et al., 2024) significantly outperforms both a generic model and personalised prompting but adapting to individual preference data yields marginal gains over training on pooled preferences from a diverse population. Beyond length biases, fine-tuning amplifies sycophancy and relationship-seeking behaviours that people reward in short-term evaluations but which may introduce deleterious long-term consequences. Replicating this within-subject experiment with simulated users recovers aggregate model hierarchies but simulators perform far below human self-consistency baselines for individual judgements, discuss different topics, exhibit amplified position biases, and produce feedback dynamics that diverge from humans.

Large language models (LLMs) are increasingly deployed in settings where the available context is incomplete or degraded. We argue that an LLM generating answers under incomplete context can be viewed as an implicit imputer, and evaluated against a criterion from the multiple imputation (MI) literature: uncertainty should scale with the amount of missing information. We assess this criterion on SQuAD, using a controlled framework in which context availability is varied across five levels. We evaluate two answer-level uncertainty measures that can be estimated from repeated sampling: sampling-based confidence (empirical mode frequency) and response entropy. Confidence fails to reflect increasing missingness: it remains high even as accuracy collapses. Entropy, by contrast, increases with context removal, consistent with the MI analogy, and explains substantially more variance in accuracy than confidence across all evidence levels (quadratic $R^2$ gap up to 0.057). We further introduce a black-box diagnostic $ρ_R(α)$ that estimates the proportion of baseline uncertainty resolved by context level $α$, requiring only repeated sampling with and without context. These results suggest that entropy is a more responsive black-box uncertainty measure than confidence under incomplete context.

Backtesting large language models on historical events requires reasoning exclusively from information available before a specified cutoff date. Yet models routinely leak post-cutoff knowledge from pre-training into their reasoning, inflating apparent accuracy and undermining evaluation validity. Prompt-based constraints fail when suppressed content is causally related to the prediction, and knowledge unlearning cannot address this problem because temporal compliance is instance-specific: the same fact may be legitimate evidence for one cutoff date and a violation for another. Rather than erasing knowledge, the model must learn temporal discipline: selecting evidence conditioned on each instance's cutoff date. We propose TEMPO (Temporal Enforcement via Mode-separated Policy Optimization), which trains this discipline via two contributions: (1) a two-mode reward where a leakage mode drives post-cutoff claims to zero as a hard prerequisite before a performance mode optimizes task performance; and (2) a GRPO-based training pipeline that enables the model to discover temporally valid reasoning strategies. We prove that training monotonically decreases leakage, converges to the leak-free optimum, and improves task performance once compliance is achieved. On three prediction tasks and two models, TEMPO reduces leakage from 213% to 0.63.7% across all conditions, with task performance improving 6~13% where strong pre-cutoff signals exist and maintained where the prediction task is inherently difficult from valid information alone.

Leaderboards rank frontier models on independent axes but do not reveal whether capabilities reinforce or trade off across releases -- and at the frontier, this interaction is the more informative signal. We decompose paired SWE-bench and GPQA Diamond scores into a population coupling trend and per-release residual ($h$-field) that diagnoses capability emphasis from two public benchmark scores. Across 34 models from 10 labs (2024--2026), capabilities cooperate ($r = +0.72$, $p < 10^{-6}$), but cooperation varies systematically: per-lab coupling slopes span $5\times$ (Google $1.15$ vs. DeepSeek $0.23$), and labs pivot -- DeepSeek reversed from reasoning-rich to coding-first ($Δh = 15.9$~pp); Anthropic oscillates between coding excursions and recovery. The population regression serves as an isocline phase boundary: the same $\sqrt{(a/b)\cdot B_1}$ classifier that identifies the base-scale coupling transition [Amin, 2026] classifies frontier models and already detects mixed-phase behavior at the next transition (two models below the GPQA--IFEval isocline). The $h$-field is not just diagnostic -- it tells you what to change. Pretraining establishes coupling at $0.871$ while RLHF adds $0.081$ [Amin, 2026]: pretraining-level shifts are permanent (DeepSeek's four-release reversal persists), post-training shifts are reversible (Anthropic's three coding excursions each recover within one release), and inference compute alone shifts $h$ by $+7.8$~pp without retraining. Knowing which component dominates determines whether to retrain or wait. We provide a three-step diagnostic (locate, classify, predict), a per-lab measurement-priority table, and seven falsifiable predictions with timestamped criteria. Five post-cutoff releases fall within the 95% prediction interval. Code, data, and an interactive dashboard: https://zehenlabs.com/cape/.

Large language models (LLMs) can fluently generate student-like responses, making them attractive as simulated students for training and evaluating AI tutors and human educators. Yet such simulators are typically evaluated by output similarity to real students, not by whether they behave like students with coherent misconceptions during interaction. We introduce a controlled framework for evaluating misconception faithfulness, whether a simulator maintains a misconception-driven belief state and updates selectively when feedback addresses the underlying misconception. Central to our framework is a misconception-contrastive feedback protocol that compares targeted feedback against two controls: misaligned feedback (targeting a different but plausible misconception) and generic feedback (only identifying answer is wrong). We propose Selective Flip Score (SFS), which quantifies how much more often a simulator flips its answer under targeted feedback than under contrastive controls. Across seven LLMs (4B-120B), multiple datasets, and prompting strategies, simulators exhibit near-zero SFS, correcting their answers at similarly high rates regardless of feedback relevance. Further analyses reveal a sycophantic failure mode: models behave less like students with misconceptions but more like problem-solvers who treat any corrective signal as a cue to abandon the simulated belief and re-solve from internal knowledge. To address this, we develop a post-training pipeline spanning supervised fine-tuning (SFT), preference optimization, and reinforcement learning (RL) with an SFS-aligned reward; SFT yields notable gains up to +0.56, and SFS-aligned RL provides more consistent improvements than preference optimization. Our results establish misconception faithfulness as a challenging yet trainable property, motivating a shift from static output matching toward interactive, belief-aware student modeling.

General-purpose safety benchmarks for large language models do not adequately evaluate disability-related harms. We introduce DisaBench: a taxonomy of twelve disability harm categories co-created with people with disabilities and red teaming experts, a taxonomy-driven evaluation methodology that pairs benign and adversarial prompts across seven life domains, and a dataset of 175 prompts with human-annotated labels on 525 prompt-response pairs. Annotation by four evaluators with lived disability experience reveals three findings: harm rates vary sharply by disability type and will compound in non-text modalities, terminology-driven harm is culturally and temporally bound rather than universally assessable, and standard safety evaluation catches overt failures while missing the subtle harms that only domain expertise can recognize. Disability harm is simultaneously personal, intersectional, and community-defined: it cannot be isolated from the full context of who a person is, and general-purpose benchmarks systematically miss it. We will release the dataset, taxonomy, and methodology via Hugging Face and an open-source red teaming framework for direct integration into existing safety pipelines with no additional infrastructure.

Large Language Models (LLMs) achieve strong performance on logical reasoning benchmarks, yet their reliability remains uncertain. Existing evaluations rely on static benchmarks, which fail to assess robustness under logically equivalent transformations and often overestimate reasoning capability. We propose LGMT (Logic-Grounded Metamorphic Testing), an oracle-free framework that leverages first-order logic (FOL) to evaluate LLM reasoning. By deriving metamorphic relations from formal logical equivalences, LGMT constructs semantically invariant test cases and detects reasoning defects through cross-case consistency checking. Experiments on six state-of-the-art LLMs show that LGMT exposes substantial hidden defects missed by traditional reference-based evaluations. We further find that models are particularly sensitive to symbol-level and conclusion-level variations, and that advanced prompting such as Few-shot CoT only partially mitigates these issues. These results suggest that LLM evaluation should move beyond isolated correctness toward robustness under logical invariance. LGMT provides a principled and scalable approach for diagnosing reasoning failures.

Evaluation of foundation models often rely on aggregate scores from benchmarks that lack comprehensive coverage and metadata for a fine-grained evaluation. We introduce a framework for automated benchmark generation. Our framework generates evaluation problems grounded in reference material, such as textbooks, producing benchmarks with broad coverage, rich metadata, and robustness to contamination. The pipeline employs a multi-agent architecture for problem generation and a solution-graph-driven strategy that significantly improves the reliability of ground truth solutions. Using the framework, we generate three benchmarks in Machine Learning, Corporate Finance, and Personal Finance. Expert review finds a significantly lower ground-truth error rate than previous benchmarks such as MMLU and GSM8K. Evaluation of 12 commercial and open-source models shows that our benchmarks achieve near-uniform competency coverage and surface performance differences across models that existing benchmarks fail to capture. We will open-source the framework and our curated benchmarks soon.

Estimating question difficulty is a critical component in evaluating and improving large language models (LLMs) for question answering (QA). Existing approaches often rely on readability formulas, retrieval-based signals, or popularity statistics, which may not fully capture the reasoning challenges posed to modern LLMs. In this paper, we introduce Q-DAPS (Question Difficulty based on Answer Plausibility Scores) method, a novel approach that estimates question difficulty by computing the entropy of plausibility scores over candidate answers. We systematically evaluate Q-DAPS across four prominent QA datasets-TriviaQA, NQ, MuSiQue, and QASC-demonstrating that it consistently outperforms baselines. Moreover, Q-DAPS shows strong robustness across hyperparameter variations and question types. Extensive ablation studies further show that Q-DAPS remains robust across different plausibility estimation paradigms, model sizes, and realistic settings. Human evaluations further confirm strong alignment between Q-DAPS's difficulty estimates and human judgments of question difficulty. Overall, Q-DAPS provides an interpretable, scalable, and bias-resilient approach to question difficulty estimation in modern QA systems.

Large language models (LLMs) have demonstrated remarkable capabilities, but they still frequently produce hallucinations. These hallucinations are difficult to detect in reasoning-intensive tasks, where the content appears coherent but contains errors like logical flaws and unreliable intermediate results. While step-level analysis is commonly used to detect internal hallucinations, it suffers from limited granularity and poor scalability due to its reliance on step segmentation. To address these limitations, we propose TokenHD, a holistic pipeline for training token-level hallucination detectors. Specifically, TokenHD consists of a scalable data engine for synthesizing large-scale hallucination annotations along with a training recipe featuring an importance-weighted strategy for robust model training. To systematically assess the detection performance, we also provide a rigorous evaluation protocol. Through training within TokenHD, our detector operates directly on free-form text to identify hallucinations, eliminating the need for predefined step segmentation or additional text reformatting. Our experiments show that even a small detector (0.6B) achieves substantial performance gains after training, surpassing much larger reasoning models (e.g., QwQ-32B), and detection performance scales consistently with model size from 0.6B to 8B. Finally, we show that our detector can generalize well across diverse practical scenarios and explore strategies to further enhance its cross-domain generalization capability.

Evaluating large language models (LLMs) in the biomedical domain requires benchmarks that can distinguish reasoning from pattern matching and remain discriminative as model capabilities improve. Existing biomedical question answering (QA) benchmarks are limited in this respect. Multiple-choice formats can allow models to succeed through answer elimination rather than inference, while widely circulated exam-style datasets are increasingly vulnerable to performance saturation and training data contamination. Multi-hop reasoning, defined as the ability to integrate information across multiple sources to derive an answer, is central to clinically meaningful tasks such as diagnostic support, literature-based discovery, and hypothesis generation, yet remains underrepresented in current biomedical QA benchmarks. MedHopQA is a disease-centered multi-hop reasoning benchmark consisting of 1,000 expert-curated question-answer pairs introduced as a shared task at BioCreative IX. Each question requires synthesis of information across two distinct Wikipedia articles, and answers are provided in an open-ended free-text format. Gold annotations are augmented with ontology-grounded synonym sets from MONDO, NCBI Gene, and NCBI Taxonomy to support both lexical and concept-level evaluation. MedHopQA was constructed through a structured process combining human annotation, triage, iterative verification, and LLM-as-a-judge validation. To reduce leaderboard gaming and contamination risk, the 1,000 scored questions are embedded within a publicly downloadable set of 10,000 questions, with answers withheld, on a CodaBench leaderboard. MedHopQA provides both a benchmark and a reusable framework for constructing future biomedical QA datasets that prioritize compositional reasoning, saturation resistance, and contamination resistance as core design constraints.

Conversational scams, such as romance and investment scams, are emerging as a major form of online fraud. Unlike one-shot scam lures such as fake lottery or unpaid toll messages, they unfold through multi-turn conversations in which scammers gradually manipulate victims using evolving psychological techniques. However, existing research mainly focuses on static scam detection or synthetic scams, leaving open whether language models can understand how real-world scams progress over time. We introduce PreScam, a benchmark for modeling scam progression from early conversations. Built from user-submitted scam reports, PreScam filters and structures 177,989 raw reports into 11,573 conversational scam instances spanning 20 scam categories. Each instance is hierarchically structured according to the scam lifecycle defined by the proposed scam kill chain, and further annotated at the turn level with scammer psychological actions and victim responses. We benchmark models on two tasks: real-time termination prediction, which estimates whether a conversation is approaching the termination stage, and scammer action prediction, which forecasts the scammer's subsequent actions. Results show a clear gap between surface-level fluency and progression modeling: supervised encoders substantially outperform zero-shot LLMs on real-time termination prediction, while next-action prediction remains only moderately successful even for strong LLMs. Taken together, these results show that current models can capture some scam-related cues, yet still struggle to track how risk escalates and how manipulation unfolds across turns.

Large Language Models (LLMs) and Vision Language Models (VLMs) have recently shown promising capabilities in various scientific domain. In particular, these advances have opened new opportunities in drug discovery, where the ability to understand and modify molecular structures is critical for optimizing drug properties such as efficacy and toxicity. However, existing models and benchmarks often overlook toxicity-related challenges, focusing primarily on general property optimization without adequately addressing safety concerns. In addition, even existing toxicity repair benchmarks suffer from limited data diversity, low structural validity of generated molecules, and heavy reliance on proxy models for toxicity assessment. To address these limitations, we propose MolDeTox, a novel benchmark for molecular detoxification, designed to enable fine-grained and reliable evaluation of toxicity-aware molecular optimization across stepwise tasks. We evaluate a wide range of general-purpose LLMs and VLMs under diverse settings, and demonstrate that understanding and generating molecules at the fragment-level improves structural validity and enhances the quality of generated molecules. Moreover, through detailed task-level performance analysis, MolDeTox provides an interpretable benchmark that enables a deeper understanding of the detoxification process. Our dataset is available at : https://huggingface.co/datasets/MolDeTox/MolDeTox

[Abridged] Production LLM deployments receive feedback from a non-random fraction of users: thumbs sit mostly in the tails of the satisfaction distribution, and a naive average over them can land 40-50 percentage points away from true system quality. We treat this as a topic- and sentiment- stratified selection-bias problem and propose a three-agent hierarchical Bayesian pipeline that does not require ground-truth labels on individual interactions. A Topic Clustering Agent partitions the stream via UMAP + HDBSCAN over text embeddings; a Bias Modeling Agent fits a two-stage hierarchical Beta-Binomial under NUTS, inferring per-topic selection rates $s_c$ and quality $q_c$ with partial pooling; a Synthesis Agent reweights $q_c$ by true topic prevalence $\hatπ_c = n_c/N$ to report a bias-corrected aggregate posterior $\bar Q = \sum_c \hatπ_c q_c$ with credible interval, plus drift signals for online recalibration. Validation uses UltraFeedback (N=10,232 retained interactions, $C=18$ clusters, $Q^\star=0.6249$) with simulated topic- and sentiment-dependent selection biases. We compare five Bayesian variants against Naive and IPW baselines. A mild prior on the feedback channel (typical positive-feedback rate and negative-to-positive ratio, both readable from any production dashboard without labels) keeps Hierarchical-Informed within 4-13 pp of $Q^\star$ as the bias ratio sweeps from 1:1 to 30:1, with 95% credible intervals covering $Q^\star$ in 50/50 random-seed replicates at $κ_{\max}=10$. Without channel-side priors, every weak-prior variant misses $Q^\star$ by 22-33 pp: the per-cluster sufficient statistics admit a one-parameter family of equally good fits, and the prior on the bias channel (not on latent quality) is what breaks the degeneracy.

Large language models (LLMs) are increasingly used to simulate human behavior, but their ability to simulate $individual$ privacy decisions is not well understood. In this paper, we address the problem of evaluating whether a core set of user persona attributes can drive LLMs to simulate individual-level privacy behavior. We introduce PrivacySIM, an evaluation suite that benchmarks LLM simulation of user privacy behavior against the ground-truth responses of 1,000 users. These users are drawn from five published user studies on privacy spanning LLM healthcare consultations, conversational agents, and chatbots. Drawing on these user studies, we hypothesize three persona facets as plausible predictors of privacy decision-making: demographics, previous experiences, and stated privacy attitudes. We condition nine frontier LLMs on subsets of these three facets and measure how often each model's response to a data-sharing scenario matches the user's actual response. Our findings show that (1) privacy persona conditioning consistently improves simulation quality over no-persona conditioning, but even the strongest model (40.4% accuracy) remains far from faithfully simulating individual privacy decisions. (2) A user's stated privacy attitudes alone may not be the best predictor because they often diverge from the user's actual privacy behavior. (3) Users with high AI/chatbot experience but low stated privacy attitudes are the most challenging to simulate. PrivacySIM is a first step toward understanding and improving the capabilities of LLMs to simulate user privacy decisions. We release PrivacySIM to enable further evaluation of LLM privacy simulation.

Large Language Models (LLMs) achieve strong performance on standard knowledge evaluation benchmarks, yet recent work shows that their knowledge capabilities remain brittle under question variants that test the same knowledge in different forms. Robustness augmentation of existing knowledge evaluation benchmarks is therefore necessary, but current LLM-assisted generate-then-verify pipelines are costly and difficult to scale due to low-yield variant generation and unreliable variant verification. We propose SAGE (Scalable Automated Generation of Robustness BEnchmarks), a framework for scalable robustness augmentation of knowledge evaluation benchmarks using fine-tuned smaller models. SAGE consists of VariantQual, a rubric-based verifier trained on human-labeled seed data, and VariantGen, a variant generator initialized with supervised fine-tuning and further optimized with reinforcement learning using VariantQual as the reward model. Experiments on HellaSwag show that SAGE constructs a large-scale robustness-augmented benchmark with quality comparable to the human-annotated HellaSwag-Pro at substantially lower cost, while the fine-tuned models further generalize to MMLU without benchmark-specific fine-tuning.

This article analyzes the use of Large Language Models (LLMs) as support for the conceptual modeling of relational databases through the automatic generation of Entity-Relationship (ER) diagrams from natural language requirements. The approach combines different language models with prompt engineering techniques to evaluate their ability to identify entities, relationships, and attributes in a conceptually consistent manner. The experimental evaluation involved three LLMs, each subjected to three prompting techniques (Zero-Shot, Chain of Thought, and Chain of Thought + Verifier), applied to the same requirements scenario with progressively increasing complexity. The generated diagrams were qualitatively analyzed through direct comparison with the textual requirements, considering the structural and semantic adherence of the modeled elements. The results indicate that, although LLMs show reasonable performance in less complex scenarios, their reliability decreases as the complexity of the requirements increases, with a rise in inconsistencies, ambiguities, and failures in representing constraints. These findings reinforce that, in their current state, LLMs are not sufficiently mature for reliable use in complex scenarios, and the cost of validation may offset the apparent productivity gains.

The performance of Large Language Models (LLMs) on downstream tasks is fundamentally constrained by the capabilities acquired during pre-training. However, traditional benchmarks like MMLU often fail to reflect a base model's plasticity in complex open-ended scenarios, leading to inefficient model selection. We address this by introducing a new task of predicting post-training potential - forecasting a base model's performance before post-training. We propose RuDE (Rubric-based Discriminative Evaluation), a unified framework that bypasses the generation gap of base models by leveraging response discrimination. Guided by our systematic 4C Taxonomy, RuDE constructs controlled contrastive pairs across diverse domains by fine-grained rubric violations. Extensive experiments demonstrate a correlation greater than 90% with post-training performance. Crucially, validation via Reinforcement Learning (RL) confirms that RuDE effectively identifies high-potential smaller models that outperform larger counterparts, offering a compute-efficient mechanism for foundation model development.

Large language models (LLMs) are increasingly used to simulate human populations via persona prompting, often under the assumptions that richer persona descriptions improve behavioral fidelity, similarly sized attribute combinations are equally simulatable, and persona definitions generalize across tasks. In this work, we formalize these assumptions and systematically evaluate them across multiple architectures, scales, and simulation settings. We identify a fundamental limitation we term persona manifold collapse, where increasingly expressive persona specifications lead to systematic contraction of representational and behavioral diversity. Across models, increasing persona complexity consistently reduces inter-persona separation in latent space and weakens behavioral differentiation in downstream simulation tasks. These effects persist across multiple analyses as richer personas fail to preserve human subgroup disagreement, performance varies across attribute combinations of similar size, and adding descriptive detail often degrades rather than improves simulation fidelity. Surprisingly, simple Age-Gender personas consistently outperform richly specified Ideal Customer Profiles (ICPs) across industries, achieving substantially higher downstream prediction accuracy. We find that collapse is not uniform across attributes. Certain combinations remain behaviorally stable and preserve stronger alignment with human responses, forming localized regions we term alignment bridges. Together, our results provide empirical and conceptual foundations for understanding the limits of persona-conditioned simulation, highlighting the need for representation-aware persona construction rather than increasing persona expressivity alone.

We measure procedural-skill SFT contribution across three Qwen3.5 dense scales (0.8B, 2B, 4B) on a 200-task / 40-skill holdout, with Claude Haiku 4.5 as a frontier reference. The corpus is 353 rows of (task + procedural-skill block, Opus chain-of-thought, judge-pass) demonstrations. Main finding. Under matched-path LLM-only scoring, the SFT-attributable procedural-$Δ$ lift is roughly uniform across sizes: $+0.070 / +0.040 / +0.075$ at 0.8B / 2B / 4B. Variation in post-SFT $Δ$ ($-0.005$, $+0.100$, $+0.065$) is dominated by a W-shaped pre-SFT base trajectory ($-0.075$, $+0.060$, $-0.010$, Haiku-4-5 at $+0.030$): the 5-step procedure hurts 0.8B and 4B, helps 2B, and helps frontier Haiku modestly. SFT works hardest in absolute terms where the base struggles with the procedure -- a regime-asymmetric pattern with a falsifiable prediction at 8B/14B. Methodology. (i) A bench format-compliance artifact: 83.5% of the holdout uses a deterministic ANSWER-line extractor that under-counts free-form-prose conclusions; our LLM-only re-judge reveals it was systematically biased against the curated condition. (ii) A negative-iteration sequence at 0.8B: three well-formed recipe variants cluster post-SFT curated pass-rate within a 2 pp band, constraining the absolute-pass-rate ceiling to base capacity rather than recipe. Cross-family judge validation. GPT-5.4 via OpenRouter on all 7 configurations (2800 paired episodes) agrees on the direction of every per-student finding: Cohen's $κ\geq 0.754$, agreement $\geq 93.25\%$, max headline $Δ$ shift $\leq 0.035$ pp. Two earlier framings -- "format-only learning at 0.8B" and "SFT contribution shrinks at 4B" -- were path-mismatch artifacts; this paper supersedes both. Single-seed evaluation; threats itemised in the paper.

Multimodal Large Language Models (MLLMs) have facilitated Multimodal Summarization with Multimodal Output (MSMO), wherein systems generate concise textual summaries accompanied by salient visuals from multimodal sources. However, current MSMO evaluation remains fragmented: text quality, image-text alignment, and visual diversity are typically assessed in isolation using unimodal metrics, making it difficult to capture whether the modalities jointly support a faithful and useful summary. To address this gap, we introduce MM-Eval, a unified evaluation framework that integrates assessments of textual quality, cross-modal alignment, and visual diversity. MM-Eval comprises three components: (1) text quality, measured using OpenFActScore for factual consistency and G-Eval for coherence, fluency, and relevance; (2) image-text relevance, evaluated via an MLLM-as-a-judge approach; and (3) image-set diversity, quantified using Truncated CLIP Entropy. We calibrate MM-Eval through a learned aggregation model trained on the mLLM-EVAL news benchmark, aligning component contributions with human preferences. Our analysis reveals a text-dominant hierarchy in this setting, where factual consistency acts as a critical determinant of perceived overall quality, while visual relevance and diversity provide complementary signals. MM-Eval improves over heuristic aggregation baselines and provides an interpretable, reference-weak framework for comparative evaluation of multimodal summaries.

A single base LLM now comes with dozens of post-training variants, quantized, LoRA-adapted, or distilled, and each has to be checked before release. Existing evaluations provide only a partial picture: benchmark scores and likelihood screens say that a variant has degraded, similarity scores such as CKA and SVCCA say how its features moved, and nothing connects the two. We connect them with one structural fact and one design choice: the prediction head is linear, so feature geometry reaches the loss, and we compare the two feature sets through an orthogonal map, which leaves the geometry being measured unchanged. From these we derive PRISM, a closed-form upper bound on the cross-entropy risk gap between a target model and a proxy variant, and prove that it splits exactly into three measurable axes: scale, shape, and head. Each axis names a failure mode and where to intervene: low-bit quantization distorts shape and, at the lowest bit-widths, also inflates activation scale; quantizing the output embedding inflates the head term, which dominates the bound at high bit-widths. Because the shape term is differentiable, the same geometry becomes a regularizer that curbs catastrophic forgetting more than experience replay. Across two model families and five benchmarks, PRISM ranks quantized and fine-tuned variants from a single forward pass at mean Spearman above 0.8, and eight reference sequences already recover the ranking that the risk gap itself needs over a hundred to reach.

Autoregressive language models are widely used for text evaluation, however, their left-to-right factorization introduces positional bias, i.e., early tokens are scored with only leftward context, conflating architectural asymmetry with true text quality. We propose masked reconstruction as an alternative paradigm, where every token is scored using full bidirectional context. We introduce DiffScore, an evaluation framework built on Masked Large Diffusion Language Models. By measuring text recoverability across continuous masking rates, DiffScore eliminates positional bias and naturally establishes an evaluation hierarchy from local fluency to global coherence. We further provide diagnostic tools unavailable to autoregressive frameworks: multi-timestep quality profiles that decompose scores across masking rates, and bidirectional PMI decomposition that disentangles fluency from faithfulness. Experiments across ten benchmarks show that DiffScore consistently outperforms autoregressive baselines in both zero-shot and fine-tuned settings. The code is released at: https://github.com/wenlai-lavine/DiffScore.

During disasters, extracting causal relations from social media can strengthen situational awareness by identifying factors linked to casualties, physical damage, infrastructure disruption, and cascading impacts. However, disaster-related posts are often informal, fragmented, and context-dependent, and they may describe personal experiences rather than explicit causal relations. In this work, we examine whether Large Language Models (LLMs) can effectively extract causal relations from disaster-related social media posts. To this end, we (1) propose an expert-grounded evaluation framework that compares LLM-generated causal graphs with reference graphs derived from disaster-specific reports and (2) assess whether the extracted relations are supported by post-event evidence or instead reflect model priors. Our findings highlight both the potential and risks of using LLMs for causal relation extraction in disaster decision-support systems.

We investigate the use of Large Language Models (LLMs) for zero-shot prediction of Ryff Psychological Well-Being (PWB) scores from spontaneous speech. Using a few minutes of voice recordings from 111 participants in the PsyVoiD database, we evaluated 12 instruction-tuned LLMs, including Llama-3 (8B, 70B), Ministral, Mistral, Gemma-2-9B, Gemma-3 (1B, 4B, 27B), Phi-4, DeepSeek (Qwen and Llama), and QwQ-Preview. A domain-informed prompt was developed in collaboration with experts in clinical psychology and linguistics. Results show that LLMs can extract semantically meaningful cues from spontaneous speech, achieving Spearman correlations of up to 0.8 on 80% of the data. Additionally, to enhance explainability, we conducted statistical analyses to characterise prediction variability and systematic biases, alongside keyword-based word cloud analyses to highlight the linguistic features driving the models' predictions.

While existing benchmarks demonstrate the near-perfect performance of large language models (LLMs) on various tasks, this apparent saturation often obscures the need for rigorous evaluation of their reliability. In real-world deployment, however, achieving extremely high reliability (e.g., "five-nines" (99.999%) vs. "three-nines" (99.9%)) is fundamentally critical, as this gap results in an order-of-magnitude increase in failures, which is catastrophic in reliability-critical applications. Still, estimating such a rare failure probability with tight confidence bounds requires prohibitively large LLM inference sizes, making standard Monte Carlo evaluation infeasible under limited compute budgets. In this paper, we observe that LLM failures exhibit strong systematic patterns: across broad parameterized input spaces, a small subset of inputs disproportionately accounts for the majority of failures. Leveraging this observation, we propose to learn a sampling distribution concentrated on failure-prone inputs via the cross-entropy method (CEM). We evaluate our framework on three LLMs, Qwen2.5-Math-7B-Instruct, gpt-oss-20b-low, and Gemini 2.5 Flash Lite, across parameterized GSM8K templates and achieve up to 156.22x reduction in required inferences compared to naive uniform sampling. Our estimates reveal that models with indistinguishable accuracy on standard benchmarks can differ substantially in estimated failure rates, underscoring that reliability is a distinct and measurable axis of model quality. Our simple yet practical framework enables the evaluation of extreme reliability in LLMs, a distinct and underexplored dimension of evaluation beyond existing benchmarks, for their growing use in reliability-sensitive applications.

The irreversible nature of blockchain transactions makes the identification of smart contract vulnerabilities an essential requirement for secure system development. While Large Language Models (LLMs) are increasingly integrated into developer workflows, their reliability as autonomous security auditors remains unproven. We assess whether current generative models are a viable replacement for, or only a complement to, traditional static-analysis tools. Our findings indicate that LLM efficacy is undermined by both inherent lexical bias and a lack of rigorous validation of external data inputs. This reliance on non-semantic heuristics, such as identifier naming, leads to a high frequency of false positives. Furthermore, prompting techniques reveal a trade-off between precision and recall. These results were derived using our custom automated framework, which achieves 92% accuracy in classifying model outputs.

Diversity is essential for language-model applications ranging from creative generation to scientific discovery, yet modern LLMs often collapse into a narrow subset of plausible outputs. While prior work has developed benchmarks for measuring this lack of diversity, less is known about how the step-by-step probability distributions at inference time cause the problem. We introduce a validity--diversity framework that attributes diversity collapse to how an LLM allocates probability mass across valid and invalid continuations during decoding. This framework decomposes the bottleneck into two complementary forms of miscalibration. First, order calibration: valid tokens are not reliably ranked above invalid tokens, so rank-based cutoff rules must trade off between recovering valid continuations and admitting invalid ones. Second, shape calibration: probability mass is overly concentrated only on few valid continuations while having a heavy-tail of mixed valid and invalid tokens, so maintaining high validity limits diversity. We formalize both mechanisms and show that local failures compound across decoding steps, producing strong sequence-level losses in diversity. Empirically, we develop controlled diagnostics for probing these bottlenecks, including tasks with exactly known valid sets and oracle cutoff baselines. Across 14 language models spanning multiple families and scales, we find that diversity collapse is not merely a limitation of particular sampling heuristics, but a consequence of order and shape miscalibration in the LLM distribution.