Theoretical analysis of standard transformers with softmax and low precision, proving they can simulate Turing machines via Chain-of-Thought. Authors construct hardmax transformers with ternary activations, then convert to equivalent softmax without unrealistic parameter magnitudes. Results validated on Sudoku reasoning.
AdaGraph is a graph-native clustering algorithm that overcomes the curse of dimensionality by operating on kNN graph topology instead of Euclidean distances. Tested on 10 synthetic benchmarks (d=10 to 5000) and three scientific domains (genomics, NLP, materials science), it outperforms HDBSCAN, WGCNA, and other methods without requiring k specification.
Study of 'cognitive poisoning': malicious tools accumulate trust through benign feedback before becoming harmful. TRUST-Bench (1,970 episodes) and VISTA-Guard propose defense via final-action risk scoring from interaction trajectory. Prompt-centric heuristics fail; trajectory-aware scoring achieves 84.2% in-domain performance.
GRASP is an agentic system for multi-hop question answering that decomposes queries into dependency-aware plans and uses a three-layer hierarchical graph (entities, propositions, passages). On MuSiQue, 2WikiMultihopQA, and HotpotQA, GRASP achieves highest accuracy while consuming 40-50% fewer tokens than IRCoT+HippoRAG2.
InvDesFlow-AL combines diffusion and active learning for inverse design of functional materials. The model achieves RMSE 0.0423 Å in crystal structure prediction (+32.96% vs existing methods) and systematically generates low-formation-energy materials. Validation: discovery of Li₂AuH₆ as BCS superconductor at 140 K.
FastOCR introduces a training-free framework to accelerate OCR on Vision-Language Models by exploiting dynamic visual fixation. Through KV cache pruning, the model reduces visual tokens processed to 5% per decoding step while retaining 98% accuracy on Qwen2.5-VL, achieving 3.0× attention latency reduction.
SomaliWeb v1: Somali corpus of 819,322 documents (~303M tokens) with BPE-16K tokenizer and language-identification benchmark. Reveals major defects in existing distributions (HPLT v2: 17.3% duplicates, 56.1% mojibake). Tokenizer 40.2% more efficient than GPT-4's cl100k_base on FLORES-200.
TABROUGE, a deterministic reward metric based on Longest Common Subsequence, improves LLM table reasoning without training. RE-TAB, a plug-and-play framework using TABROUGE, gains 26.7 pp across six backbones and three benchmarks, reducing test-time scaling samples by 33%.
Mosaic, a probabilistic weather forecasting model, addresses three spectral degradation failures in ML-based prediction: spectral damping, high-frequency aliasing, and residual leakage. With 214M parameters at 1.5° resolution, it matches models trained 6× finer and generates well-calibrated ensembles in 12s for 10-day forecasts on H100.
New metric called Refusal Index (RI) measures LLMs' ability to refuse questions beyond their knowledge. RI correlates refusal probability with error probability using Spearman's rank correlation. Testing across 16 models and 5 datasets shows LLM refusal behavior remains fragile despite high factual accuracy.
BacktestBench is the first large-scale benchmark for automated quantitative backtesting, containing 18,246 annotated QA pairs across 6 million real market records. AutoBacktest, a multi-agent system, translates natural language strategies into reproducible backtests via a Summarizer, SQL Retriever, and Python Coder. Evaluation on 23 mainstream LLMs.
LoopQ is a post-training quantization (PTQ) framework for recursive language models (LoopLMs) that reuse Transformer blocks. It addresses three challenges: distribution shift across roles, state reuse between loops, and recursive error accumulation. Under W4A4 quantization, LoopQ improves downstream accuracy by 68.8% and reduces perplexity by 87.7% versus static PTQ baseline.
Study of membership inference attacks (MIA) on masked diffusion language models (MDLM). Researchers extract 46-dimensional feature vectors from reconstruction loss at different masking ratios and train XGBoost and MLP classifiers. On MIMIR benchmark, XGBoost achieves AUC 0.878 (peak 0.930), outperforming SAMA baseline by 0.062 AUC. ELBO trajectory alone drives most of the signal.
arXiv paper demonstrating that multi-turn conversations in low-resource African languages (Afrikaans, Kiswahili, isiXhosa, isiZulu) bypass safety mechanisms in commercial LLMs. Testing ChatGPT, Claude, DeepSeek, Gemini, and Grok shows jailbreak rates from 52.7% to 83.6% depending on model. Translation quality is the critical success factor.
Discrete diffusion models generate code through iterative refinement. CDC (Constrained Diffusion for Code) integrates constraints directly into the denoising process without additional training, combining mathematical optimization and program analysis to improve functional correctness, security, and syntax.
First systematic study of source attribution for generated 3D assets. Researchers build a benchmark covering 22 3D generators and propose a hierarchical multi-view multi-modal Transformer detecting fingerprints (cross-view inconsistencies, geometric artifacts, frequency-domain signatures). Results: 97.22% accuracy under full supervision, 77.17% with only 1% training data.
OpenJarvis presents a decomposed personal AI stack architecture with five primitives (Intelligence, Engine, Agents, Tools & Memory, Learning) to run tasks locally without routing sensitive data to cloud. Using LLM-guided spec search, on-device specs match or exceed cloud accuracy on 4/8 benchmarks, reduce API cost by ~800x and end-to-end latency by 4x.
Asynchronous RAG framework predicting when and what to retrieve using three components (retrieval predictor, context monitor, query generator). Achieves 43.5% end-to-end latency reduction and 62.4% time-to-first-token improvement by exploiting semantic precursors in generation dynamics while maintaining answer quality.
EvoMemBench is a unified benchmark evaluating LLM agent memory across two axes: scope (in-episode vs. cross-episode) and content (knowledge-oriented vs. execution-oriented). Comparison of 15 memory methods: long-context baselines remain highly competitive, retrieval-based methods dominate knowledge-intensive tasks, procedural methods excel for execution-oriented tasks.
Researchers propose the Refusal Index (RI), a metric measuring LLMs' ability to refuse questions beyond their knowledge. RI correlates refusal probability with error probability using Spearman's rank correlation. Testing across 16 models and 5 datasets shows LLMs refuse unreliably despite high factual accuracy.
TailedTS is a benchmark of 24.69 billion data points from Wikipedia 2024 page views, designed to test time series forecasting models under heavy-tailed and non-Gaussian conditions. The dataset shows 5% of pages account for 70% of traffic, and introduces a periodicity quantification framework revealing high-traffic pages have weaker periodic structure.
SkillMOO optimizes skill bundles for coding agents via multi-objective search (NSGA-II) on pass rate and inference cost. Across 16 SkillsBench SE tasks, the framework achieves top pass rate rank on 11/12 non-zero-pass tasks, reducing costs up to 31.7% and gaining up to 21 percentage points in pass rate.
arXiv paper introducing a trace-optional evaluation protocol decomposing token efficiency of reasoning LLMs. Analyzes 14 open-weight models on CogniLoad, GSM8K, ProofWriter, ZebraLogic by separating completion rate, conditional correctness, and generated length. Identifies three failure modes: logic-limited, context-limited, or verbosity-limited.
Attractor-Vascular Coupling Theory (AVCT): mathematical framework showing cardiac attractor geometry encodes blood pressure information. Calibrated LightGBM model on smartphone PPG achieves MAE 2.05 mmHg (SBP) and 1.67 mmHg (DBP) in strict leave-one-subject-out cross-validation (46 subjects, 29,684 windows), meeting AAMI/IEEE SP10 criteria. PPG-only ablation matches ECG+PPG within 0.05 mmHg.
CompactAttention optimizes chunked prefill attention for long-context LLMs using Block-Union KV Selection. The method converts block-sparse masks into GQA-aware per-group KV block tables, eliminating explicit KV compaction. On LLaMA-3.1-8B, it achieves 2.72× speedup at 128K tokens while maintaining accuracy near dense attention (RULER benchmark).
Theoretical study proving multi-layer cross-attention is optimal for multi-modal in-context learning. Authors show single-layer linear self-attention fails to recover Bayes-optimal predictor, but linearized cross-attention mechanism achieves Bayes optimality with gradient flow.
Critical study of LLM-based trading agents (FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, FLAG-Trader). Reported Sharpe ratios do not constitute deployment evidence: temporal contamination, unmodeled frictions, and insufficient predictive calibration invalidate results. Proposes P1-P6 reporting protocol and modular architecture with LLMs as auditable information interfaces.
Token-level confidence trajectories in LLMs encode geometric signals linked to reasoning trace correctness. Without access to text or hidden states, low-dimensional representations separate correct from incorrect traces on GSM8K, MATH, and MMLU. NeuralConf, a lightweight estimator, improves confidence-weighted answer aggregation over majority voting.
arXiv paper showing LLMs exhibit a knowing-doing gap in tool use: recognition of tool necessity vs. actual invocation diverge. Testing 4 models on arithmetic and factual QA reveals 26.5-54% mismatches. Hidden state probing shows cognition and action signals become nearly orthogonal in late layers, with most failures at the cognition-to-action transition, not in recognition itself.
Perovskite-R1 is a specialized LLM based on QwQ-32B, fine-tuned on 1,232 scientific publications and 33,269 candidate materials to discover precursor additives optimizing perovskite solar cells. The model generates solutions for defect passivation and improves stability/performance, experimentally validated.
DCFold is a single-step generative model achieving AlphaFold3-level accuracy with 15x inference acceleration. Using a Dual Consistency training framework and Temporal Geodesic Matching scheduler, it maintains predictive fidelity for structure prediction and binder design.
EvoMemBench is a unified benchmark evaluating LLM agent memory along two axes: scope (in-episode vs cross-episode) and content (knowledge vs execution-oriented). Comparison of 15 memory methods: long-context baselines remain highly competitive, retrieval-based methods dominate knowledge-intensive tasks, procedural methods excel at execution-oriented tasks.
PROTEA is an interface for offline debugging and refinement of multi-agent LLM workflows. It evaluates intermediate outputs with configurable rubrics, localizes bottlenecks via workflow graph visualization, and generates targeted prompt revisions. On two production-adjacent workflows, PROTEA improves document-inspection accuracy from 64.3% to 83.9% and recommendation Hit@5 from 0.30 to 0.38.
Tabular foundation models (TFMs) for healthcare are distilled into lightweight models using context-leakage-aware distillation. Across 19 medical datasets and 6 TFMs, student models retain ≥90% of teacher AUC while running 26× faster on CPU, preserving calibration and fairness critical for health applications.
AgentArk distills multi-agent intelligence into a single LLM through three hierarchical strategies: reasoning-enhanced fine-tuning, trajectory-based augmentation, and process-aware distillation. The resulting model preserves single-agent computational efficiency while retaining multi-agent reasoning and self-correction capabilities.
CounterCount is a diagnostic framework to evaluate counting bias in vision-language models. Tests show VLMs perform well on factual images but degrade significantly on counterfactual images where visual attributes contradict learned priors. An inference-time attention modulation strategy improves accuracy by up to 8% across multiple VLMs.
DocReward is a document reward model evaluating structure and style of professional documents, independent of textual quality. Trained on DocPair (117K document pairs, 32 domains), it outperforms GPT-4 by 14.6 percentage points and effectively guides agents via RL toward higher structural and stylistic professionalism.
DriveMoE proposes a Mixture-of-Experts architecture for end-to-end autonomous driving. The model combines Vision MoE (dynamic camera selection based on driving context) and Action MoE (specialized expert activation for different behaviors). Built on Drive-π₀ baseline, DriveMoE achieves SOTA on Bench2Drive by avoiding mode averaging.
LegalCheck automates drafting of objection response letters in Dutch municipalities using RAG and context-augmented generation. Deployed in Amsterdam, the system generates near-final legal drafts in minutes instead of hours, capturing 80–100% of essential content while maintaining expert review and legal consistency.
SocialMemBench benchmarks AI memory systems in multi-party social groups (430 personas, 7,355 conversation turns, 1,031 QA pairs). Gemini 2.5 Flash reaches 0.721 on small networks vs 0.98 expected. Four open-source frameworks (Mem0, LangMem, Graphiti, Cognee) score 0.12-0.18, well below references (0.345-0.369), revealing a measurable gap.