Verifying corporate financial disclosures against the International Financial Reporting Standards (IFRS) is a core auditing task in which experts cross-reference disclosures with hundreds of fine-grained regulatory paragraphs: a laborious, costly, and error-prone process that motivates growing interest in large language models (LLMs) as compliance assistants. Yet single-model accuracy remains insufficient for high-stakes use, and naive ensembling is vulnerable to systematic per-model biases that can drag aggregate accuracy below that of the strongest individual annotator. This gap is addressed with a multi-stage framework, supported by a new benchmark of 19,998 IFRS regulation-disclosure pairs comprising 11,161 unanimous multi-LLM labels and 8,837 human annotations on items where the LLMs disagreed. The framework couples a three-stage domain-adapted sentence embedding pipeline (BERT masked language modeling on SEC EDGAR filings, unsupervised TSDAE and SimCSE adaptation, and supervised contrastive alignment) with a learned ensemble classifier that fuses frozen embeddings and LLM probability features from GPT-4.1, Claude Sonnet 4, and Gemini 2.5 Pro via Tree-structured Parzen Estimator (TPE) optimized gradient-boosted trees. On the disagreement set, the ensemble achieves 81.7% weighted F1, exceeding the strongest single LLM by 10.4 points and majority voting by 14.6 points, and outperforms eleven unsupervised aggregation baselines; the retrieval pipeline attains Recall@15 of 45.3% on unseen regulations. The dataset, model checkpoints, annotation platform, and classifier implementation will be publicly released.
Large Language Models (LLMs) have achieved remarkable progress on complex reasoning tasks, yet the mechanisms by which they acquire reasoning capabilities and the efficiency of their reasoning remain poorly understood. In this paper, we investigate the path-finding problem over directed graphs, a symbolic abstraction of multi-step reasoning. We first characterize the architectural requirements for this task, proving that one-layer transformers require $\Omega(N^2)$ size on dense graphs to implement the key DFS child-selection primitive. While a two-layer transformer implements full DFS with $O(N\log(N))$ size. Thus, a second layer is necessary and sufficient for near-linear DFS reasoning. We then extend this expressivity analysis to deeper models: for every $k\geq3$, a $k$-layer transformer implements a bounded-reachability DFS policy with look-ahead horizon $1+2^{k-3}$, establishing an exponential increase in reachability horizon with depth. Finally, going beyond expressivity, given appropriately curated training data, we show that two-layer transformers can provably learn to execute DFS via gradient flow and generalize to unseen graphs. Together, our results provide a mechanistic account of how transformers implement graph search, how depth expands bounded look-ahead, and how such reasoning algorithms can emerge through training.
The morphology of astrocytes is closely associated with their ability to support and modulate neural circuit functions; however, its reconstruction from volume electron microscopy (vEM) data remains labor-intensive and computationally resource-intensive, placing substantial barriers on its broader adoption and scalability despite the high spatial resolution of vEM. We present AstroSeg-vEM, a lightweight, astrocyte-centric framework for morphology-preserving astrocyte segmentation, 3D reconstruction, and quantitative morphological analysis in vEM images. AstroSeg-vEM combines transformer-based 2D segmentation with dedicated post-processing methods and operates directly on high-resolution EM sections, making it well suited to anisotropic vEM data and downstream proofreading workflows while preserving the full in-plane resolution of the original EM data. Across four mouse hippocampal datasets representing normal and morphologically defined stress-associated astrocyte ultrastructural states, AstroSeg-vEM demonstrates robust performance and generalization, achieving average Dice and weighted Dice values of 0.92 and 0.95, respectively, on a held-out test dataset. Its lightweight architecture enables efficient, high-fidelity reconstruction on consumer-grade hardware, providing a scalable and accessible platform for large-scale astrocyte ultrastructure analysis and future astrocyte-neuron interaction studies.
Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.
Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.
As text-to-video (T2V) models are increasingly positioned as world simulators, it is essential to evaluate whether their videos conform to physical commonsense. However, existing benchmarks rely on annotators who rate each video with a single score based on intuition, yielding ratings that are subjective and do not tell which physical principle is violated. We introduce PhyRubric, a rubric-grounded diagnostic framework for evaluating physical commonsense in T2V generation. For each caption, PhyRubric constructs a caption-specific rubric before any video is seen, decomposing the caption into semantic and physical items, and derives scores from item-level verdicts grounded in timestamped evidence. We instantiate this framework as PhyRubric-Bench, which spans 34 categories across five physical domains, and scale it up to PhyRubric-Train with 3,788 videos, on which we fine-tune the open-source PhyRubric-Judge. Empirical results across ten T2V models show that physical commonsense remains far from solved, and that PhyRubric-Judge correlates more closely with human judgments than existing open-source judges while matching the frontier proprietary Gemini 4 Argon. Further analysis provides actionable insights for T2V models.
Much of the variation in financial time series is driven by market-wide shocks that cannot be predicted from an asset's own history. Per-asset self-supervised learning treats this variation as irreducible noise. Under a latent factor model, we show that conditioning prediction on the realized market factor makes the asset's market exposure necessary for optimal prediction, while reducing the noise against which the asset-specific component is learned. Building on this insight, we introduce Market-Conditioned JEPA (MC-JEPA), a joint-embedding predictive architecture for financial time series. During training, we construct a leave-one-out market condition for each asset from the future target embeddings of other assets and incorporate it through factor loadings derived from the asset's own representation. On 11 years of minute-level Chinese A-share data, MC-JEPA improves over per-asset JEPA across nine targets spanning returns, market exposure, risk, and liquidity, with the strongest and most statistically significant gains in exposure, risk, and liquidity. It achieves the best RankIC on all liquidity targets and on return horizons from two hours to five days, outperforming time-series foundation models and self-supervised baselines, and delivers the best portfolio performance in a historical backtest. Ablations show that the return and exposure gains require the condition to be aligned with the realized market movement. After restoring the volatility scale removed by normalization, the predictor's factor loadings rank stocks by market beta nearly as well as a supervised probe, without labels. A synthetic factor-model study with known parameters shows that conditioning improves the recovery of both market exposure and the asset-specific component.
Tokenisers such as Byte-Pair Encoding (BPE) shorten text by repeatedly merging an adjacent pair of symbols into a new one; fewer tokens mean less computation. We ask whether an efficient algorithm can choose $k$ merges that leave a string of length $n$ almost as short as possible; doing nothing is already within a factor $n$. We prove that, unless P = NP, no polynomial-time algorithm comes within a factor $c\,n/\log n$, for some constant $c>0$, already for one binary string and for BPE-style merges that replace every occurrence. The key idea is that $k$ merges form a grammar with $k$ rules: a vertex-cover gap becomes a grammar-size gap, which we amplify with relabelled copies, encode in binary and realise with explicit merges. For merges that may skip occurrences, a simple algorithm achieves $O(n/\log\log n)$, and no Karp reduction can prove hardness beyond order $n/\log n$ unless coNP $\subseteq$ NP/poly. Over a one-letter alphabet, both tokenisation problems left open by Kastreva et al. are strongly NP-complete on datasets with repeated strings.
Evaluating a learned AC optimal power flow (AC-OPF) surrogate on one grid serialization cannot show whether its accuracy depends on that encoding. We test this dependence with thirteen metamorphic relations, each specifying when a transformation is valid and how to compare its outputs. Our audit covers 17 model and pipeline configurations: two released trained checkpoints and 15 random-weight architecture or mechanism controls. Every configuration violates at least six relations, even at reporting tolerances ten times tighter or looser. For GridSFM-Open v1.1, source inspection and intermediate-tensor interventions point to an indexing bug and three noncanonical or numerically unstable preprocessing choices; a repaired checkpoint remains untested. In a separate case500 training experiment, augmentation lowers transformed-view error in three of four models and median discrepancy and constraint-violation P95 in all four. These benefits depend on the training protocol: some relations worsen, and every evaluated prediction remains AC infeasible. A GPU/CPU eigensolver mismatch accounts for most of the large raw-Laplacian-baseline gain. We recommend reporting representation consistency alongside accuracy, feasibility, and cost. These results establish neither deployment readiness nor universal invariance.
Scientific design often requires balancing competing objectives that are expensive to evaluate through experiments or simulations. Multi-objective Bayesian optimization (MOBO) addresses this challenge by using probabilistic models to guide the selection of promising designs. In many applications, tunable inputs map to observable intermediate responses, which are then combined through known functions to compute the final objectives. Prior work on Bayesian optimization of composite functions shows that learning from these intermediate responses, which we call Composite modeling, can substantially improve sampling efficiency in single-objective optimization, compared to methods that condition only on black-box observations of the objective function. The benefits of Composite modeling in MOBO, however, remain understudied. In this work, we study the benefits of Composite modeling in MOBO and investigate the factors driving variation in its gains. We first establish single-objective conditions under which intermediate observations provide no additional predictive information beyond objective values, where composite and black box methods use the same warped Gaussian process prior. Controlled numerical experiments with versions of four MOBO methods across eight synthetic and application-inspired scientific-design benchmarks examine how these conditions relate to optimization gains in the multi-objective setting, where the black box approach models the objective functions with a standard multi-output Gaussian process. Both investigations support the claim that composite modeling tends to help more when the function mapping intermediate responses to final objectives is non-bijective and non-affine.
Objective: Concept-based interpretation has emerged as a promising approach for explaining blackbox models using human-understandable, high-level concepts. However, most existing methods require annotated concept labels to train the model to identify concepts in the input, and obtaining such annotations in clinical text is often expensive. This study aims to enable concept-based interpretation for language models without concept supervision on clinical prediction tasks. Methods: We propose a novel framework that enables concept-based interpretation without concept supervision through a tailored cross-attention bottleneck. Clinical text and the text descriptions of medical concepts are first tokenized and embedded by a pretrained language model. A cross-attention module then estimates the presence of each concept for each text token, producing concept relevance scores that capture how concepts are involved throughout the note. These scores are pooled and used in a linear prediction layer to model clinical outcomes in an interpretable manner. Results: The performance of the proposed framework was compared against blackbox models across two tasks: chest trauma severity stratification using a private dataset and an in-hospital mortality prediction task using the publicly available MIMIC-III dataset. The proposed framework achieved AUROC values within 0.02 of the corresponding blackbox models. The framework also enabled cohort- and patient-level interpretation, capturing the overall concept importance across the cohort and patient-specific concept contributions. Conclusion: Our proposed framework realizes concept-based interpretation without requiring ground-truth concept presence labels. It improves the transparency of clinical text predictions while maintaining strong performance in clinical prediction tasks.
Edge NPU performance depends not only on hardware capabilities, but also on model transformation and the overall deployment pipeline. We propose a benchmarking methodology that separates these effects by comparing three configurations: Baseline, Vendor Guided NPU Only (VG-NPU), and Vendor Guided End-to-End (VG-E2E), where VG-NPU preserves common model I/O and host-side processing while VG-E2E follows the vendor-native deployment pipeline. We apply this methodology to four edge NPUs and three vision workloads. VG-NPU improves both accuracy and Runtime-call latency in 10 of 12 NPU--workload combinations, whereas VG-E2E reduces System-level latency in all combinations but improves accuracy in only four. Moreover, NPU-level performance rankings do not always match End-to-End rankings, and VG-E2E provides an additional 10.4--70.7\% reduction in System-level latency over VG-NPU. These results demonstrate the importance of evaluating vendor-guided optimization separately at the controlled model-transformation and full deployment levels when comparing heterogeneous commercial Edge NPU platforms in practice.
Benchmark scores for medical large language models (LLMs) are difficult to interpret when scoring criteria cannot be traced to clinical evidence. Here we developed Med Evidence Bench (MEB; https://www.medevidencebench.org.cn/), comprising 1,536 physician-constructed items across three levels: guideline knowledge, standardized clinical scenarios and real-world laboratory scenarios. Each criterion stating a medical claim is linked to a versioned source-text span. Across ten LLMs, the Evidence Trace Score (ETS) was 0.975-0.997 for guideline knowledge, 0.292-0.658 for standardized scenarios and 0.182-0.447 for real-world scenarios, and rankings varied across levels. For real-world scenarios, scores for identifying clinical information (mean, 0.770) exceeded those for matching evidence (0.165) or formulating recommendations (0.160). On 1,296 criterion-level judgments, automated grading agreed with the majority vote of three physicians (macro-F1, 0.930; 95% confidence interval, 0.915-0.945), although agreement was lower for evidence application. These findings from Chinese-language materials indicate that guideline recall does not ensure patient-specific evidence use and do not establish clinical effectiveness.
Computational cognitive modeling has traditionally relied on manual model construction, which is a time-intensive process that requires substantial domain expertise. Although recent LLM-based approaches have begun to automate model generation, they often prioritize predictive accuracy at the expense of structural plausibility and parameter identifiability. We introduce DisCo, an LLM-guided framework that discovers interpretable symbolic cognitive models through evolutionary search with two structural components: a Validator that filters structurally invalid models, and a Cleaner that promotes parameter identifiability by removing redundant parameters. Across three behavioral tasks, DisCo consistently discovers models that achieve performance competitive with task-specific baselines. Ablation analyses show that the Validator is important for recovering psychologically meaningful structure. Without this component, models tend to violate behavioral constraints and show less consistent associations with external cognitive measures. These results suggest that LLM-driven model discovery, when coupled with appropriate structural constraints, can yield cognitively plausible models that go beyond predictive fit.
RLVR trains language models to reason by rewarding correct final answers. A widely used algorithm, group-relative policy optimization (GRPO), gives every token of a solution the same credit, so dense process rewards have been proposed to credit intermediate steps, for example with a Monte-Carlo reward that scores each step by sampling continuations from it. Comparing such a reward with GRPO per training update, or vetting it with a gradient test on a single batch, leaves out what it costs: in our setting, the continuations alone take 91% of all generated tokens. We introduce a matched-compute protocol that gives every training intervention the same budget, measured in generated tokens and training GPU-hours, and tests it in preregistered studies with controls for side effects. Across 1,668 runs on six models, matching compute reverses the ranking. On Qwen2.5-1.5B, the Monte-Carlo reward leads standard GRPO per update, yet ends 14.3 percentage points lower at equal tokens and 7.8 points lower at equal training time, and none of four dense-reward variants outperforms GRPO even with as much training time. A per-token gradient test orders these rewards as training does but overestimates each by one to two points, as much as the gains it is meant to detect. In contrast, training on the prompts whose predicted success is nearest 1/2 improves accuracy at matched compute, by 1.7 points on Qwen3-0.6B and 0.8 on Qwen2.5-1.5B. These gains replicate on fresh seeds and unseen problems, and a simple rule, that the candidate prompts are predicted to fail more often than not, predicted where they appear in five of six preregistered tests. Both gains come early in training, and standard GRPO with twice the initial learning rate comes within half a point of them. We argue that training interventions should be validated at matched compute, which overturns some apparent gains and confirms others.
We present Darmok, the first automated system to reach superhuman performance on British cryptic crosswords. Unlike American-style crosswords, cryptic clues disguise an unmarked definition and compositional wordplay beneath a misleading surface reading, sparse grids leave nearly half of each word's letters unchecked, and tournament rules rank a grid with even a single wrong or blank square behind all error-free solutions regardless of speed. Competing alongside an elite human field at the 55th Times Crossword Championship in London, Darmok solved all seven championship puzzles error-free and finished faster than every human competitor in every round. It solved the Final puzzle in 6m 52s, a full 3m 13s ahead of the human champion. Today's frontier AI models can on average solve a whole puzzle in a single try with comparable accuracy and speed. However, cloud API calls still stall, time out, trigger safety refusals, or get stuck in reasoning loops. To simultaneously achieve tournament-grade accuracy, speed, and reliability, Darmok cross-checks clues across three independent models in rounds; uses verified crossing letters to unlock harder clues; and uses a whole-puzzle model and dictionary lookups to fill any remaining gaps without overwriting confirmed words. This design maintained near-zero-error performance across a five-month daily evaluation period. Under normal benchmark conditions, removing backup models, fallback steps, or elaborate prompt instructions barely changes accuracy, making the architecture appear over-engineered. Yet during both the worst days of daily operation and a severe slowdown of our primary model in the live Championship Final, the backup paths carried the solve, showing that as foundation models master the average case, the central challenge in building complex AI systems shifts to resilience in the worst case.
Post-training improves language-model reasoning, but how it changes the use of internal information and verbalized chain of thought (CoT) remains unclear. We study successive SFT, DPO, and RLVR checkpoints of Olmo-3-7B-Instruct on multi-digit addition, using carry values as known intermediate state. Probes recover carry information from internal activations with similar accuracy in SFT and RLVR, although manipulating this information does not reliably control model outputs. In contrast, the usefulness of verbalized reasoning differs: CoTs produced by RLVR yield higher accuracy across receiving checkpoints than those produced by SFT or DPO. We also show that carry-related statements receive higher importance scores after RLVR than matched random controls. DPO shows similar sensitivity to carry-related statements but without comparable accuracy gains. These results suggest that post-training changes the usefulness of written reasoning and sensitivity to its content in meaningful ways, although greater sensitivity does not necessarily coincide with better problem performance.
This technical note introduces tabular foundation models (TFMs) to accounting research and evaluates their performance in one common benchmark task: predicting firms' next-year earnings changes. TFMs produce numeric predictions and learn in context, without post-release training or tuning, making them natural and easy to use for forecasting. The TFMs we study are pretrained on synthetic tabular data, so they avoid look-ahead bias from memorizing firms' outcomes. We compare Google Research's TabFM with a fine-tuned open-weight large language model developed in recent accounting research and with several tree-based models. TabFM is more accurate than all comparison models. It is also faster and less computationally intensive than the large language model. Several tree-based specifications produce predictions more quickly than TabFM, so choosing between them involves a trade-off between accuracy and speed. We conclude that TFMs have promise as a new tool for forecasting in accounting and finance. However, our evaluation is based on a single benchmark task, and no class of models is likely to dominate in all settings.
Sparse autoencoders (SAEs) are commonly evaluated by variance explained and ablation damage. We show how massive activations undermine both evaluations, and propose a correction. Massive activations cause up to 99.9% of activation variance to lie along the first principal component (PC1). Because the PC1 component has no unique sparse decomposition, each SAE run rebuilds it from a different set of features. These features dominate importance rankings: an average of 47 of the 50 most important features are unmatched in a separate run, compared with about 27 of a random 50. A dictionary can also report 99.96% of variance explained while leaving 33% of the variance outside PC1 unexplained. We bypass PC1 by training the SAE on activations with PC1 removed, then adding PC1 to the reconstruction. In affected models, the top-50 feature sharing rate rises from 4-39.5% to 73-99.5%. At the same TopK sparsity, the resulting SAE degrades the model up to 44% less and reduces content reconstruction error by up to 33.9%.
BF16 is widely used to pretrain large language models because it has the same dynamic range as FP32 and converges similarly to FP32 training. However, in reinforcement learning with verifiable rewards, where learning rates are typically low, we find that standard mixed-precision BF16 training converges more slowly than FP16 training at the same learning rate. We attribute this gap to weight sticking: when FP32 shadow weights are quantized to BF16 using round-to-nearest, small updates often leave the BF16 weights used for rollout generation and training unchanged. Consequently, several optimizer steps may be needed before updates to the shadow weights are reflected in the BF16 weights. Raising the BF16 learning rate reduces weight sticking and matches FP16's convergence speed, but it also increases the KL divergence from the reference model. We instead apply stochastic rounding when quantizing the FP32 shadow weights to BF16. This approach matches FP16's convergence speed across different models and datasets while reducing the KL divergence relative to BF16 training with a larger learning rate. Unlike FP16, it keeps BF16's wider dynamic range and needs no loss scaling.