New inference-time method DCO (Dynamic Contextual Orthogonalization) to reduce hallucinations in LLMs. Based on hypothesis that hallucinations manifest as orthogonal noise relative to semantic manifold of residual stream. Evaluated on Llama-3 (8B/70B) with improvements on XSum, NQ-Swap, IFEval and TriviaQA.
CORE is a framework training MLLMs to detect multimodal fake news by identifying semantic or physical conflicts across modalities. Using an annotated corpus (CAC) with conflict factors, CORE generalizes to unseen manipulations in few-shot or zero-shot settings, outperforming existing methods.
Novel topological approach to interpretability in deep neural networks. Authors introduce 'coherence', a geometric property where each neuron responds to contiguous regions of state space. They propose Coh, a differentiable objective function based on Fréchet variance, validated on MNIST and BERT embeddings.
FiRe-OPD introduces fine-grained on-policy distillation combining trajectory filtering and soft token reweighting. Validated on AIME 2024 (+6.25 strong-to-weak) and Miner (+18.81 multi-teacher), the method outperforms recent token-level OPD approaches in stability and performance.
HyperPatch introduces a hypergraph-based sequential knowledge editing framework for LLMs. It addresses N-ary Structural Drift, where complex relation updates fracture relational integrity. On MQuAKE-CF and MQuAKE-T benchmarks, it achieves 96.24% and 21.06% relative gains in Hop-wise Accuracy over strongest baselines.
Novel framework combining importance-aware news compression and process-level retrieval supervision for time series forecasting. A reward model estimates each article's forecasting utility for sequential fusion, while a PRM ranks supplementary-news candidates based on error profile. Experiments on finance, energy, traffic, and bitcoin benchmarks show improved accuracy and fewer refinement iterations.
Conditional hypothesis generation framework for LLM-based text analysis incorporating researcher-specified covariates. Addresses stratum imbalance and sign reversal through feature-covariate interactions and inverse-frequency reweighting. Validated on synthetic and real-world datasets in computational social science.
Researchers propose Bank of Values (BoV), replacing context-dependent value vectors with context-free vectors stored as sparse parameters in the last third of layers. On 135M and 780M models, BoV improves validation loss and performance across 21 benchmarks with reduced compute and memory.
Study of demographic bias in skin lesion classification using ResNet-based models. Three strategies evaluated: single-task, reinforcing multi-task, adversarial learning. Sex bias stems from data imbalance; age bias consistently favors younger groups. Mitigation strategies show partial effectiveness depending on data distribution.
Framework combining conformal prediction and collaborative filtering-style annotator representation to analyze LLM behavior against human annotators in content moderation. Introduces Ghost Prediction metric to quantify model-human divergences. Evaluation across 4 LLMs and 4 datasets shows larger models more confident on texts with no human alignment, revealing structural demographic bias.
Study showing visual graphs help LLMs organize multi-hop reasoning better than text representations. Graph-structured mind maps guide models without direct answer hints, unlike flattened text graphs. Advantage persists after supervised fine-tuning and KL-based distillation.
Study on linguistic productivity in LLMs: models reproduce entrenchment (frequent structures) but fail to implement preemption (observed absence of structures). Large models handle contextual coercion but overgeneralize patterns never seen in training data.
Ψ-Bench is a benchmark assessing LLMs' ability to persuade realistic users through conversation. 10 frontier models tested on 3 real-world scenarios. Access to user profiles yields 18.24% performance gain. Code available.
ML method for real-time binary road surface classification (dry/damp grip vs snow/ice slip) using production vehicle signals during cruising. Feature-based and end-to-end frameworks using wheel speeds, torques, longitudinal acceleration, steering angle, yaw rate. Validated on public-road data.
HDPO (Hint-Guided Diversified Policy Optimization) enhances LLM reasoning through reinforcement learning with verifiable rewards. The method prompts models to first generate multiple candidate solution outlines (hints), then select the most reliable one. Two stages: Cold Start for structured reasoning, then hint-guided RL to diversify and improve solution reliability.
Theoretical study on geometric limits of feature representation in transformers. Authors establish a framework based on linear representation and superposition hypotheses, showing representational capacity depends on vectors-to-dimensions ratio (k/d) rather than raw count. Analysis of dozens of open-source models reveals two classes based on orthogonality constraint ε.
MASC (Margin Self-Correction) is a language-model unlearning method that efficiently reduces the logit gap between the original next token and its alternatives, without requiring downstream evaluation. Tested on TOFU, MUSE News, and MUSE Books, it achieves competitive forget-retain trade-offs at a fraction of existing baselines' computational cost.
Dataset of 410,499 tropical species covering plants, aquatic, and pet animals. Integrates taxonomic identifiers (GBIF, NCBI, iNaturalist, etc.), adds cross-domain ontology, Chinese vernacular names (99.50% coverage) with explicit provenance, and CITES linkage. Deposited on Zenodo.
RelGT-AC extends RelGT for autocomplete tasks on relational databases. The model combines column masking, a unified task head (classification/regression), and a TF-IDF encoder for free text. On 7 RelBench v2 tasks, it outperforms GraphSAGE on 3 regression tasks and gains +10 AUROC on text-heavy eligibility tasks.
arXiv paper on LLM representations: lexical overlap influences embeddings more than semantic content, persists across layers and architectures, and degrades performance in summarization and model editing tasks.
Qift introduces a zero-free level set for W2A4/KV4 quantization ({±0.5, ±1.5}) leveraging Hadamard rotation. Training-free and codebook-free, it improves perplexity on LLaMA-2-7B and LLaMA-3.1-8B versus standard levels {-2,-1,0,+1}, narrowing the gap to W3A4 in mixed precision while keeping half transformer layers at two-bit.
ChatHealthAI aligns structured EHR representations from a pretrained EHR foundation model with a frozen LLM's semantic space via a task-aware resampler. The multimodal framework integrates longitudinal patient representations with refined clinical event descriptions, improving interpretable clinical reasoning while maintaining competitive predictive performance on the EHRSHOT benchmark.
arXiv study showing LLMs struggle to infer user sociodemographics from single conversation histories. Disparities in advice (legal, medical, financial) are minimal but present. Conversation topics prove more predictive than sociodemographic data and affect LLM outputs unpredictably.
Comparative study of encoder-only Transformer vs LSTM for streamflow prediction in ungauged basins. On NOAA NWM data, LSTM outperforms Transformer. Incorporating downstream information improves all models (+60% median NNSE). Recurrent memory proves better suited for this hydrologic inference task.
QUIVER enriches classical ML models with quantum views based on quantum Fisher information matrix extracted from variational quantum circuits. Tested on QM9 (molecular properties) and JetClass (LHC), the paradigm improves performance without requiring fault-tolerant quantum hardware.
Two automated metrics assess LLM lexical misalignment: Lexical Alignment Score detects term overuse ('suggest', 'additionally', 'strategy'), Triangulated Preference Shift quantifies RLHF impact. Tested on 6 model families (Falcon, Gemma, Llama, Mistral, OLMo, Yi) via PubMed abstracts, no manual annotation required.
Framework for LLM agents operating under underspecified instructions. Proposes Information Gain Reward metric quantifying clarification utility via Bayesian belief updates. Validation on τ-Bench: +3.7% success rate vs no-clarification baseline, +0.3 interaction steps.
Padyam2Gadyam: dataset of 600 classical Telugu poems (13th-17th centuries) with human-verified modern Telugu and English prose translations. Evaluation of 5 contemporary LLMs shows insufficient performance and limitations of existing MT evaluation metrics for this task.
Researchers propose an agent economy where AI agents self-coordinate through auctions and payment exchanges without centralized control, inspired by Hayek's economic theory. This approach generates emergent multi-step reasoning strategies and outperforms baselines on five tasks including mathematical reasoning, financial research, and distributed-system optimization.
G²C-MT proposes graph-guided context selection for document-level machine translation. The system models discourse dependencies between paragraphs via a lightweight graph and uses depth-biased random walks to extract context paths. Tested on DeepSeek-V3, Gemini-2.5-Flash-lite, and Qwen-2.5/3, the approach outperforms baselines across multiple domains.
SNMPBB, first adaptation of nonmonotone Barzilai-Borwein methods to Symmetric NMF. Demonstrates 6× speedup over SymANLS on synthetic data. Extensions for graph clustering (Graph-SNMPBB) and large-scale problems (LAI-SNMPBB) with proven global convergence to first-order stationary points.
Adaptive multifidelity algorithm for machine learning in quantum chemistry. Dynamically determines dataset composition by querying samples at each fidelity level. Reduces data generation costs by up to 30× versus single-fidelity methods and 5× versus standard MFML on coupled cluster energies and excitation energies.
New arXiv study on editing factual opinions in LLMs. FOE benchmark with 261 public figures, 19 issue categories, 2,178 opinion records. Current editing methods fail to maintain consistency between edited opinion and model-generated evidence. Proposes Self-Generated Evidence-Aligned method for opinion-evidence alignment.
PhotoCraft introduces a hierarchical memory system for photo-search agents combining working, episodic, and semantic memory to maintain logical consistency across multi-step reasoning. Training-free approach tested on DISBench achieves up to 18.5% improvement in context-aware retrieval across MLLM backbones.
RL-based method for adaptive sampling control at test-time on LLMs. A lightweight controller trained with RL dynamically decides to stop or continue sampling, balancing answer correctness, latency, and computation cost. MDP formulation with Lagrangian interpretation. Outperforms ASC and ESC on trade-offs.
Theoretical paper on generalization bounds under distribution shift. Proposes framework quantifying extra risk from regime composition mismatch in Markov-switching environments, with exact decomposition and minimax lower bounds. Tested on 25 years of global equity indices.
Internal Coherence Maximization (ICM) automatically generates examples to align AI models with diverse human values without extensive human supervision. Tested across four benchmarks, ICM matches gold label performance. Example coherence improves generalization even at constant accuracy, especially for underrepresented personas.
DMT-CBT models longitudinal therapeutic state evolution in CBT across multiple sessions. The framework maintains structured states, incorporates multimodal data and tool-augmented interventions. DMTCorpus, a synthetic multimodal dataset, demonstrates improved counseling fidelity and therapeutic alliance over post-hoc extraction approaches.
SenseJudge is a customizable judgment framework driven by human preferences for evaluating LLM responses. Paired with SenseBench, a benchmark derived from real-world multi-turn interactions, it outperforms existing methods in adapting to user preferences and achieves model ranking aligned with human judgment.
Deterministic orchestration framework for validating fragmented ESG data (Scope 1-3) with temporal anomaly detection, imbalance-aware ensemble learning, and audit provenance tracing. Synthetic benchmark calibrated against GHG Protocol, PCAF, ISSB standards. Evaluation on classification, calibration, and provenance chain completeness metrics.