Research Reality Cards

SREE achieves relative improvements of 70% in post-level precision and 59% in F1 compared to state-of-the-art retrieval approaches while retaining at least 30% fewer posts.

SREE achieves the highest exact item-level agreement across all three eRisk collections.

The paper does not specify the authors or provide detailed reproducibility metrics.

URL-Sim achieves higher quantitative accuracy and visual quality in LiDAR scene reconstruction compared to existing methods while maintaining a bounded total primitive count.

Achieves higher accuracy and visual quality on Waymo and KITTI-360 datasets under comparable primitive budgets.

Existing methods are limited to rigid vehicles and may produce redundant primitives, which could affect generalizability.

LandslideFM surpasses general-purpose VLMs in specialized landslide reasoning and scene analysis.

LandslideFM achieves superior quantitative and qualitative performance across public datasets and the newly introduced LandslideVL benchmark.

The reliance on limited labeled data may affect the model's generalizability in diverse real-world scenarios.

SELD consistently outperforms existing baselines in identifying evidence loss in compressed contexts, achieving an 83% detection rate of evidence-removing edits with unchanged answers.

Identified evidence-removing edits in 83% of matched pairs with unchanged answers, compared to 70% for the strongest baseline.

The paper does not specify the authors or provide detailed reproducibility concerns.

EarthRecord provides a principled platform for measuring and advancing multimodal agents' capabilities from raw satellite imagery to reliable land-change analysis.

Evaluations of twelve multimodal agents revealed that leading agents complete most analyses end to end but struggle with mapping from imagery.

The benchmark relies on raw multi-sensor archives, which may not be readily accessible for all researchers.

The proposed method of filtering clinician edits significantly improves model alignment with clinician expectations while maintaining input context fidelity.

Naive fine-tuning on raw edited output increases ungrounded content by over 45%, while the filtering approach enhances clinician-alignment metrics.

Real clinician edits are not publicly available, which may limit the generalizability of the findings.

The study demonstrates that testing on pages from training manuscripts significantly inflates accuracy metrics, with a notable drop in page-level accuracy from 97% to 54%.

Page-level accuracy drops from 97% to 54%, and the expected calibration error (ECE) rises significantly.

The accuracy difference with soft labels could not be confirmed or ruled out due to the limited number of manuscripts.

The StepRefine method significantly reduces hallucinations in MDLMs by employing a step-aware switch between discrete denoising and latent refinement.

StepRefine demonstrates significant hallucination mitigation across different MDLMs on multimodal benchmarks.

The paper does not specify the reproducibility of the results across all MDLMs or the generalizability of the method.

ConceptTime closes the circularity in defining meaningful temporal patterns and segment boundaries, achieving competitive accuracy with state-of-the-art methods while maintaining interpretability.

ConceptTime retains near-98% accuracy on Epilepsy with only 1% of labels across 17 public benchmarks.

The method's reliance on a frozen probabilistic forecaster may limit its adaptability to different domains without retraining.

The zone-based formulation of graph generative models effectively preserves geological properties and supports history matching while remaining within the learned geological manifold.

The optimisation search was conducted in a 160-dimensional joint latent space.

The paper does not specify the reproducibility of the results across different geological settings or datasets.

StatPrompt-LLM achieved a 40.2% reduction in MSE and a 21.4% reduction in MAE under the 10% reduced-prefix configuration compared to Time-LLM.

40.2% reduction in MSE under the 10% reduced-prefix configuration.

Additional model parameters and runtime overhead may affect reproducibility and efficiency.

TimeOmni-o1 outperforms existing models in understanding and forecasting tasks across multiple domains by effectively integrating TSFMs and LLMs.

Achieved superior performance on the MMTR dataset, outperforming existing models in both understanding and forecasting tasks.

The main limitation is the potential difficulty in replicating the multimodal time series reasoning dataset (MMTR) due to its size and complexity.

TopicScope enables the training of compact topic monitors that can be deployed during generation, achieving robust performance with minimal size.

Monitors trained on features from prefill hidden states generalize to decoding, maintaining performance with a maximum loss of 5 macro-F1 points under perturbations.

The classifiers reaching baseline performance without the sparse autoencoder were significantly larger, which may affect deployment efficiency.

ADIndex achieves up to 6.22x prefill and 3.40x decoding kernel speedups while maintaining near-lossless task performance at a context length of 256K.

End-to-end serving speedup of 1.54x at a context length of 128K.

The abstract does not specify potential limitations or reproducibility concerns.

The introduction of Personalized Federated Exploration-Free Value Iteration (PF-EFVI) achieves logarithmic regret without explicit exploration bonuses by leveraging reward diversity.

PF-EFVI attains logarithmic regret under sufficient reward diversity.

The paper does not specify the authors or provide detailed experimental results, which may limit reproducibility.

The study successfully implemented autonomous synchrotron X-ray experiments that adapt to evolving human-defined objectives, showcasing effective collaboration between human and agentic systems.

61-67% of messages involved interpretation or judgment, indicating a significant level of cognitive collaboration.

Sustained operation on physical instruments remains largely unexplored, which may limit reproducibility.

AptM-GCD achieves state-of-the-art performance in generalized category discovery by effectively utilizing localized attribute evidence and multimodal prompts.

AptM-GCD demonstrates consistent state-of-the-art performance across six datasets.

The reliance on specific attribute embeddings may limit generalizability to other datasets or domains.

PR-OVD achieves an average 6.5% proportional improvement on novel categories over state-of-the-art methods in open-vocabulary object detection.

6.5% proportional improvement on novel categories.

The paper does not specify the authors or provide detailed reproducibility metrics.

The proposed protocol reveals that aggregate mAP overestimates occluded-scene performance and introduces metrics for joint-angle error and temporal stability.

Delta_occ is negative in every combination tested, indicating that aggregate mAP masks occlusion-specific differences.

The paper does not specify the authors, which may limit reproducibility and verification of results.

Larger variants of language models exhibit internal geometric alignment with human psychological constructs, while base variants do not show this alignment.

The correlation between geometry strength and behavioral amplification was found to be r = 0.68, p = 0.008.

The study does not establish that the measured directions mediate the cross-persona answers, which raises questions about the reproducibility of the findings.

BEVCast achieves a 66% reduction in BEV encoder time while maintaining comparable accuracy to full-resolution grids.

At a 10% refinement budget, BEVCast encodes 28% of the BEV queries.

The method's performance may vary based on the specific driving scenarios and datasets used for training and evaluation.

HaNDF provides a flexible and powerful prior for hand-object interaction, enabling pose refinement and physically plausible manipulation synthesis.

Extensive evaluations demonstrate the effectiveness of HaNDF in optimizing hand poses conditioned on object interactions.

The paper does not specify authors or provide detailed reproducibility metrics.

InstaGen improves PSNR over the SVG2 method by 3.4 dB and 2 dB on two different models while maintaining matched attention density.

InstaGen achieves a PSNR improvement of 3.4 dB over SVG2 at matched attention density.

The paper does not specify the authors or provide detailed implementation guidelines, which may hinder reproducibility.

Real-world identifications of AI-generated writing are contextually dependent and often leverage sociolinguistic norms rather than merely identifying 'AI-like' features.

The study employs qualitative coding and computational analysis to reveal diverse evidentiary types used in AI identification.

The findings may be context-specific and influenced by the sociolinguistic norms of the groups studied, which could limit generalizability.

The study finds that large language models exhibit overcompression of fine-grained emotional distinctions, clustering primarily into positive and negative groups despite retaining fine-grained predictive information.

The KMeans clustering revealed that the models primarily separate emotions into two groups: positive and negative.

The study does not provide specific author names or detailed methodology, which may limit reproducibility.

The study demonstrates that the MCQ format leads to quicker commitment to answer tokens and significantly reduces refusal rates compared to open-ended formats.

MCQ activations are approximately 3x further from the refusal direction than open-ended formats.

The findings may not generalize across all model architectures or datasets beyond WMDP.

Subliminal learning occurs in vision-language models, allowing preferences to transfer from a biased teacher to a student through unrelated training data.

Students trained on biased-teacher outputs showed a preference difference of up to ~10 percentage points in the animal preference task.

The study's findings may vary with the target preference, model, and the method of number generation, which could affect reproducibility.

FinEvolveKG improves knowledge graph construction by 19.4-30.9% over a frozen prompt baseline and outperforms an unconstrained prompt workflow in various evaluations.

The complete two-stage block-constrained workflow (v4) shows a 19.4-30.9% improvement over the frozen prompt baseline (v1).

The specific author names and detailed methodology are not provided in the abstract, which may hinder reproducibility.

The proposed RL-ATTACK framework achieves lower post-attack accuracy and greater query efficiency compared to untrained attackers and a strong optimization-based baseline.

Achieves post-attack accuracy of 2% within 20 queries on natural language inference tasks.

No single reward design or training setting consistently outperforms others, which may complicate reproducibility.

The PUMA framework enables effective building damage mapping with only a small number of labeled examples by utilizing weak supervision techniques.

PUMA demonstrated effectiveness in mapping damage during the 2025 Los Angeles wildfires with limited supervision.

The reliance on weak supervision may introduce variability in representation learning, affecting reproducibility.

The Two-Ledger Model effectively reconciles settlement-grain billing data with invocation-grain usage records, achieving over 99% cost coverage in simulations.

The reconciliation residual recovers unmetered spend within 0.24 percentage points while flagging provider billing errors.

The model's effectiveness may vary based on the accuracy of trace propagation and the completeness of invocation-grain records.

PLatour significantly outperforms open-source baselines on generation tasks while using a substantially smaller sender model.

PLatour surpasses Mostik's reported results on shared mathematics and coding benchmarks.

Controlled experiments indicate that performance gains may arise from uninformative sender signals or receiver-only training.

BlockLock significantly reduces latency by 53.5-55.0% and improves throughput by 115.2-122.2% compared to existing batch-invariant solutions while preserving output consistency.

BlockLock-Rows reduces latency by 53.5-55.0% and improves throughput by 115.2-122.2%.

The method's performance is evaluated under specific deployment conditions, which may limit generalizability.

The AutoIP library allows for the autoformalization of classical linear regularization theory using AI agents, emphasizing the need for human verification of formal statements.

The library organizes tasks for reviewers to compare formal statements with natural-language counterparts, reducing the need to read proofs.

The main limitation is the risk associated with the accuracy of formal statements rather than the proofs themselves.

Correcting scoring errors in benchmarks can significantly improve acceptance rates of model responses, with an increase of over 80 percentage points in some cases.

On the 96 LegalBench SARA numeric cases, correcting the scorer raised acceptance by more than 80 percentage points for each of three models.

The study relies on specific benchmarks that may not generalize to all policy-based decision systems.

SBL+S-FLEX provides more focal localization and tighter uncertainties than BMN+sLORETA, suggesting greater statistical power for identifying bilateral auditory sources.

SBL+S-FLEX yields lower Wasserstein localization error and smaller posterior uncertainty compared to BMN+sLORETA.

Residual miscalibration remains, and transfer across mismatched SNR, sparsity, and head models is condition dependent.

KDASO significantly outperforms current general-purpose and domain-specific baselines in liability adjudication tasks.

Achieved an 8.96% improvement when transferring skills to concept-similar scenarios.

The dynamic evolution of skill libraries may pose challenges in maintaining consistency across different scenarios.

SteelAgent demonstrates significant variability in predictive performance based on the choice of held-out units, highlighting the importance of context in model evaluation.

Pooled out-of-fold $R^2$ dropped from 0.859 to 0.113 when steel families were held out.

Calibration can deteriorate on held-out groups, affecting the reliability of predictions.

The proposed method significantly improves the quality-coverage frontier, reducing the failure distance metric on ImageNet-256 with minimal compute.

Achieved a reduction in FD_{r,6} from 19.22 to 5.91 at the 67% Recall point on ImageNet-256.

The paper does not specify the authors or provide detailed reproducibility guidelines.

The proposed Gaussolve framework effectively dissolves inconsistent mesh content and refills it with Gaussian primitives, resulting in higher visual fidelity and faster rendering.

Achieved higher visual fidelity with fewer Gaussian primitives and faster rendering in experiments on structured and semi-structured scenes.

The method's performance may vary based on scene complexity and specific viewing conditions, which could affect reproducibility.

FlexFloor increases response generation throughput by 52.74% while generating 65.40% more response tokens per unit of occupied computing time.

Increases response generation throughput by 52.74%.

The paper does not specify the authors or provide detailed experimental setups, which may hinder reproducibility.

T$^2$-Splat improves PSNR by 0.71dB over MeshSplatting on Mip-NeRF 360 while allowing for flexible control of mesh complexity.

Supports face budgets as low as 10% of the original face count.

The specific authors and detailed methodology are not provided, which may hinder reproducibility.

JPTA outperforms existing methods like TransCLIP and StatA in adapting vision-language models, particularly in scenarios with sparse and prior-shifted batches.

JPTA consistently shows improved performance across standard benchmarks, especially in low effective-class regimes.

The paper does not specify detailed reproducibility metrics or datasets used for validation.

Pretrained representations improve classification accuracy over random-weight counterparts but do not consistently outperform strong non-pretrained baselines.

Pretraining improves classification accuracy by up to 23 percentage points on five of six datasets.

Frozen representations do not reliably outperform the strongest non-pretrained baseline in any setting, raising concerns about their practical utility.

The study demonstrates that interactive backdoor attacks can effectively manipulate dialogue systems to achieve attacker-specified objectives while maintaining normal utility.

All three attack methods achieved their objectives in experiments, preserving normal utility.

The tested training-data filters failed to retain nearly all poisoned dialogues, raising concerns about the effectiveness of defenses.

PRISM, a large-language-model pipeline, identifies and drafts outreach messages for variant reinterpretation needs in online health communities, surfacing 123 candidate classification differences.

Agent-predicted reinterpretation-relevant discussion is rare (0.288%) but concentrated in specialist genetics communities (6.30%).

Factual accuracy is necessary but not sufficient for clinically acceptable outreach, indicating potential limitations in the model's applicability.

Automated feedback to tutors increased the elicitation of learner ideas and reasoning by 8-15%, but did not change the overall balance of talk between tutors and learners.

Feedback to tutors increased learner reasoning by 8-15%.

Learner-facing feedback had little effect due to group-level reporting and limited learner interaction.

TNT achieves best-in-class performance on a wide range of data streams with varying concept drift types.

TNT was evaluated on 24 real and semi-real classification data streams, demonstrating superior performance.

The paper does not specify limitations or reproducibility concerns.

TraCE achieves competitive task performance in action recognition and video question answering while improving interpretability of motion sources in video understanding.

Maintains an average accuracy drop of only 1.0 points for action recognition and a nominal gain of 0.4 points for VQA across six benchmarks.

The reliance on specific benchmarks may limit generalizability and reproducibility across diverse video datasets.

Octopus achieves 49.6% accuracy on BrowseComp-Plus, outperforming single-agent GRPO by 17.5 points and dense-retrieval agents up to 250 times larger while increasing training throughput by 3.3 times.

Increased training throughput by 3.3 times and reduced peak KV memory by 2.8 times.

The abstract does not provide specific details on the reproducibility of the results or the dataset used.

The proposed Conformal Reservation for Orchestration (CRO) algorithm effectively reserves compute and bandwidth allocations that meet pipeline demands with probabilistic guarantees, reducing resource waste significantly.

CRO halves the resource waste compared to the Chebyshev-bound baseline while satisfying target probabilistic coverage across all risk levels.

The paper does not specify the authors or provide detailed experimental setups, which may hinder reproducibility.

Challenge-aware in-between retrieval improves accuracy on 8 of 9 biomedical datasets, demonstrating the importance of timing and specificity in knowledge recruitment.

GPT-4.1-mini MedQA accuracy improved from 46% to 54% with Challenger-triggered recruitment.

Naively combining retrieval and specialist recruitment degrades performance, indicating a need for careful coordination.

The introduction of an evidence gate significantly reduces the corruption of correct proposals while still allowing for necessary repairs in multi-agent language-model systems.

The evidence gate reduced corruption of correct proposals while retaining most repairs, with specific performance differences observed between models Qwen and Mistral.

The results are model-dependent, indicating that interface design influences reviewer behavior and may not provide a universal safeguard across different models.

The authors derive a closed-form condition that determines when a tiered cascade for LLM inference is beneficial compared to sending all queries to a backstop.

The feasibility condition is bracketed by a signal-free floor and a perfect-signal ceiling, indicating the conditions under which tiered inference is effective.

The design requirements proposed may not be satisfied simultaneously by existing trained-in primitives, which could limit practical implementation.

Verified tool-health state significantly reduces mean observable policy regret by 96.9% and increases single-dispatch completion from 69.4% to 95.8%.

Mean observable policy regret reduced from 4,252 ms to 131 ms.

The study's empirical extensions regarding runtime state and heterogeneous selection remain open and may affect reproducibility.

INOS achieves over 4,600× speedup in nonlinear simulations compared to classical finite-element solvers while maintaining high predictive accuracy.

Achieves higher predictive accuracy than neural solver baselines trained on dynamic simulations, evaluated zero-shot.

The paper does not specify the authors or provide detailed reproducibility guidelines.

The proposed rhetorical structure grammar successfully parses 55.21% of California state bill discussions in the dataset.

Achieved an F1 score of 52.8% using a logistic regression-based classifier.

The performance of the pipeline may be limited by the complexity of legislative discourse and the specificity of the grammar.

CCDWiki improves the quality of conceptualization and intervention in CBT dialogue systems without requiring backbone retraining.

Substantial improvements in conceptualization and intervention quality demonstrated through experiments.

The paper does not specify the authors or provide detailed experimental conditions, which may affect reproducibility.

Deep neural speaker embeddings exhibit cross-language generalisation that aligns with human behaviour, demonstrating accurate identity discrimination and sensitivity to language switching.

Embeddings showed cross-language generalisation that paralleled human behaviour, supported by 18 state-of-the-art models.

The study may have limitations in generalising findings beyond the specific bilingual speech conditions tested.

PEEK improves image generation quality by reducing exposure bias, achieving FID improvements up to 49.3% with minimal extra computation cost.

FID improvements up to 49.3% at minimal extra computation cost.

The one-time calibration process may vary in time based on model complexity, which could affect reproducibility.