3148 articles
SynC, a deep graph clustering framework, leverages synergistic relationship between representation learning and structure augmentation via a Transform Input Graph Auto-Encoder (TIGAE). The model shares weights across two stages to reduce parameters and improves generalization on low homophily graphs.
Survey on universal time-series representation learning. Proposes taxonomy based on three fundamental elements for state-of-the-art deep learning methods. Covers extraction of hidden patterns without manual feature engineering, with resources and future research directions.
Empirical study of 117 experiments (0.2B–3.8B parameters, 2B–128B tokens) on scaling laws for Code LLMs. Code requires higher data-to-parameter ratio than natural language. Farseer law outperforms Chinchilla. Code-NL mixtures benefit NL under resource constraints but harm it at higher compute budgets.
arXiv study on long-term time series forecasting (LTSF). Authors show that a simple linear layer (affine mapping) dominates performance on standard benchmarks. Analysis reveals models learn similar transition matrices, capture periodic patterns well but fail on non-periodic signals. Code available.
Property-guided program synthesis approach reduces LLM costs by replacing simple numeric scores with formal property verification. When a property is violated, the system provides concrete counterexamples to guide repair. On PDDL planning domains, this method generates 7× fewer programs and drastically reduces evaluation costs while improving solution quality.
Formal study of imperfect world model exploitation in RL. Authors define exploitation as divergence between policy preferences in the model versus true environment. They prove exploitation is essentially unavoidable on large policy sets and establish theoretical bridge with reward hacking.
Red-Bandit is a red-teaming framework that adapts real-time specialized LoRA experts for different attack styles (manipulation, slang) via reinforcement learning. A multi-armed bandit algorithm dynamically selects the optimal expert based on target model response safety. State-of-the-art results on AdvBench with more readable prompts.
Prompt optimization framework inspired by reinforcement learning to improve long-term planning in LLM multi-turn interactions. Method modifies only task instruction via turn-by-turn feedback and experience replay. Significant improvements on text-to-SQL and task-oriented dialogue, generalizes across LLM agents.
Study on LLMs' (including OpenAI o1) ability to solve and generate linguistic puzzles from Linguistic Olympiads. Models outperform humans on most puzzle types except writing systems and understudied languages. Automated puzzle generation could expand interest in linguistics and support rare language dissemination.
LongAct is a benchmark for evaluating autonomous planning in long-horizon household tasks specified via free-form instructions. HoloMind, a VLM-driven agent with DAG-based hierarchical planner, Multimodal Spatial Memory, and Episodic Memory, achieves 59% goal completion and 16% full-task success with GPT-5 and Qwen3-VL models.
ClawForge is a benchmark framework for CLI agents testing persistent state and conflict handling. 17 scenarios, 6 ability categories. Seven frontier models evaluated: best score 45.3%, widest gap 17-90% driven by whether agents inspect existing state before acting.
CarbonScaling is a hardware-aware analytical framework modeling carbon footprint of frontier LLM training. It integrates neural scaling laws, distributed training strategies, accelerator modeling, and operational/embodied carbon accounting. Source code released on GitHub.
AutoLLMResearch introduces an agentic framework to automate configuration of expensive LLM experiments. The system learns from low-fidelity experiments to extrapolate toward promising high-fidelity configurations. LLMConfig-Gym provides a multi-fidelity environment with >1M GPU hours of verified experiment outcomes.
arXiv paper proposing a theoretical framework for detecting causal bias in generative AI models. Authors formalize causal fairness specific to generative models (vs standard ML), derive causal decompositions to quantify bias impacts across different causal pathways, and demonstrate their methodology by analyzing race and gender bias in large language models.
VLATIM, a new benchmark based on The Incredible Machine 2, evaluates Vision-Language Models' logical reasoning in point-and-click puzzle games. Results reveal a significant gap: large proprietary models excel at planning but struggle with precise visual grounding, failing to match human-level problem-solving.
EnactToM is an evolving benchmark with 300 multi-agent embodied tasks in 3D household environments with partial observability. It tests functional Theory of Mind—acting optimally on implicit beliefs—rather than literal belief questions. All seven frontier models score 0.0% on hard task completion, with 93% of failures traced to epistemic coordination breakdowns.
ScaleLogic, a synthetic logical reasoning framework, demonstrates that RL can teach long-horizon reasoning to LLMs. Training compute follows a power law with proof depth (T ∝ D^γ, R² > 0.99), with exponent γ increasing from 1.04 to 2.60 as logical expressiveness grows. Models trained on more expressive logics transfer better (+10.66 points on downstream benchmarks).
HAAS is a framework for adaptive task allocation between humans and AI systems in software engineering and manufacturing. It combines rule-based governance constraints with contextual-bandit learning. Results show governance is not binary but a tunable design variable: moderate governance improves operational performance and reduces fatigue in manufacturing while remaining competitive as the learner gains experience.
arXiv paper demonstrates that Mixture-of-Experts (MoE) models can outperform dense architectures under strictly equal resource constraints (parameters, training compute, data). Researchers identify an optimal activation rate region consistent across model sizes. Validated on ~200 2B-scale and 50 7B-scale models (50 trillion tokens processed).
MolClaw is an autonomous agent with a three-tier hierarchical architecture (70 skills) for drug molecule evaluation, screening, and optimization. It integrates 30+ specialized resources and achieves state-of-the-art performance on MolBench, a benchmark spanning 8 to 50+ sequential tool calls. Gains concentrate on structured workflow orchestration rather than ad hoc scripting.
CheeseBench evaluates 6 open-weight LLMs (3B-72B) on 9 behavioral neuroscience paradigms (Morris water maze, T-maze, etc.). Qwen2.5-VL-7B achieves 52.6% success on ASCII vs 32.1% random and 78.9% rodent baselines. Scaling >7B yields diminishing returns; longer context and chain-of-thought degrade performance.
Study comparing OpenAI o3 and Google Gemini 2.5 Pro as models of human driving behavior in a simplified merging scenario. LLMs reproduce intermittent operational control and tactical dependencies, but fail to capture responses to dynamic velocity cues. Prompt ablations reveal model-specific inductive biases that do not transfer across LLMs.
Comparative study on surgical AI: multi-billion parameter Vision Language Models fail at neurosurgical tool detection despite extensive training. Scaling experiments show diminishing improvements. Obstacles persist across architectures, suggesting data and compute alone are insufficient.
Study showing that unlearning in LLMs merely suppresses information at surface level—models recover original behavior through minimal fine-tuning. Authors introduce representation-level analysis framework (PCA, CKA, Fisher information) to assess genuine data erasure and identify four forgetting regimes based on reversibility and catastrophicity.
UniversalRAG extends retrieval-augmented generation (RAG) to heterogeneous multi-modal corpora (text, images, videos) with variable granularities. The framework proposes modality-aware routing to avoid intra-modal bias and dynamically retrieve from the appropriate corpus. Validated on 10 multi-modal benchmarks.
Neurosymbolic architecture with ontologies (Role, Domain, Interaction) for enterprise LLM agents. Controlled experiment (1,800 runs, Claude Sonnet 4, Qwen 2.5 72B, Gemma 4 26B): ontology-constrained agents outperform ungrounded agents on metric accuracy and role consistency (p < .001). 2x greater lift in localized domains (Vietnam) where LLM training coverage is weak.
New Interactive Benchmarks evaluation paradigm assesses model reasoning through budgeted multi-turn interaction. Two settings: Interactive Proofs (logic, UI2Html, mathematics with objective feedback) and Interactive Games (strategic reasoning). Reveals substantial gaps in current interactive capabilities.
arXiv study analyzing 10,000+ Google Maps reviews of urgent care facilities (DMV, Florida) using GPT and prompt engineering. Interpersonal factors and operational efficiency emerge as primary satisfaction drivers, while technical quality, finances, and facilities show no significant independent effects. Population density alone correlates with ratings among socioeconomic factors.
RAG-Gym, a framework supervising the search process rather than final answers, improves autonomous search agents. Re²Search++ uses process supervision and reasoning reflection to generate higher-quality queries, achieving significant gains on multi-hop benchmarks with better out-of-domain generalization.
BioProAgent combines LLMs with finite state machines to plan biological experiments in wet-labs. The system enforces a Design-Verify-Rectify workflow and reduces token consumption by ~6× through symbolic abstraction. On BioProBench, it achieves 95.6% physical compliance versus 21.0% for ReAct.
New arXiv paper introducing OAG (Online Algorithms with Unreliable Guidance), a model for ML-augmented online decision-making separating predictive and algorithmic components. Presents DTB (drop-or-trust-blindly) compiler converting standard online algorithms into learning-augmented versions. Demonstrates optimal guarantees on bipartite matching, caching, and uniform metrical task systems.
Novel LGPT (Learnable Graph Pooling Token) method to integrate graphs into LLMs. Uses learnable tokens to represent graphs without information loss. 4.13% improvement on GraphQA benchmark without LLM fine-tuning.
A framework uses language models to identify 'alien' research directions—coherent with existing literature but unlikely under current researcher distribution. On 16,068 AI/NLP papers, the method explores 3.5-7× broader conceptual space than baselines while maintaining scientific coherence.
Long reasoning models (LRMs) generate redundant chains of thought uncorrelated with correctness. The paper discovers LRMs implicitly know when to stop thinking. SAGE (Self-Aware Guided Efficient Reasoning) exploits this via a novel sampling paradigm, improving accuracy and efficiency on mathematical benchmarks.
LLMs exhibit 'conversational inertia' in multi-turn agent scenarios: they over-imitate their previous responses instead of exploring. Authors identify this bias through attention analysis and propose Context Preference Learning to favor low-inertia responses. Validated across 8 agent environments.
Supervised sparse auto-encoders improve model interpretability by aligning learned features with human semantics. Tested on Stable Diffusion 3.5, they enable compositional generalization and image editing through feature-level intervention.
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%.
Study on Bangla fake news detection using CNN. Combines semantic, statistical, and character-level features on BanFakeNews-2.0 dataset. Results show that hybrid feature combinations significantly improve recall and F1-scores compared to individual features.
arXiv paper argues agentic AI cannot be responsible without explicit traceable provenance. Authors formalize provenance through causal attribution function and responsibility tensor, demonstrate it is computable and intervenable online, and identify responsibility gaps in current multi-agent systems.
Investigation of extrinsic gender bias in Bangla pretrained language models. Four manually annotated task-specific benchmark datasets constructed (sentiment analysis, toxicity detection, hate speech, sarcasm detection) with minimal-pair gender perturbations. RandSymKL debiasing strategy proposed, combining symmetric KL divergence and cross-entropy loss. Implementation and datasets publicly released.
Study reveals a safety vulnerability in personalized dialogue agents: long-term memory biases intent inference and legitimizes harmful queries. PS-Bench benchmark shows personalization increases attack success rates by 15.8%–243.7% versus stateless baselines. A lightweight detection-reflection method is proposed to mitigate this safety degradation.
RL-trained Lean theorem provers suffer mode-collapse at inference: doubling sampling from k=32 to k=64 on miniF2F-test with DeepSeek-Prover-V1.5-RL solves zero additional theorems (42/244). Fixed structural diversity of 15 tactic skeletons recovers +45% relative improvement at k=16 (+12.3±4.2 theorems). Phenomenon is RL-specific and orthogonal to scaling.
Comprehensive survey of reinforcement learning post-training methods for LLMs. Unifies RLHF (DPO), RLVR (PPO, GRPO) and SFT within a single policy gradient framework. Detailed technical analysis of offline and iterative approaches with standardized notation for direct comparison.
MirrorBench is a benchmarking framework to evaluate user-proxy agents in conversational systems. It combines 6 metrics (MATTR, Yule's K, HD-D, GTEval, Pairwise Indistinguishability, Rubric-and-Reason) to measure realism of LLM-generated user utterances across 4 public datasets. Open-source code released.
arXiv paper on homogenization in LLMs: models reproduce and amplify human biases through mode collapse. Authors propose a framework to characterize homogenization in terms of normativity (queer theory) and introduce 'xeno-reproduction' as a mitigation strategy promoting diversity. Experiment on Claude 3.5 Haiku demonstrates gender bias.
FormuLLA fine-tunes LLMs (Llama2, GPT, Claude) on 1400+ FDM formulations to recommend pharmaceutical excipients and predict filament mechanical properties. Llama2 outperforms; smaller models suffer catastrophic forgetting even with this dataset size.
Vision-OPD introduces regional-to-global self-distillation to improve fine-grained visual understanding in MLLMs. The framework transfers the model's privileged perception on evidence-centered crops to its full-image policy via token-level KL divergence minimization on on-policy rollouts. Competitive results on fine-grained visual understanding benchmarks without external models or ground-truth labels.
AgroCoT is a VQA benchmark with 4,759 Chain-of-Thought samples designed to evaluate reasoning capabilities of Vision-Language Models in agriculture. Evaluation of 30 VLMs (proprietary and open-source) reveals significant gaps in zero-shot reasoning, highlighting the importance of CoT for precision farming applications.
QuickLAP fuses physical and language feedback to learn robot reward functions in real time using a Bayesian framework. LLMs extract reward feature attention masks and preference shifts from free-form utterances, integrated with physical corrections via closed-form update rule. Achieves 70% error reduction vs physical-only and heuristic multimodal baselines in semi-autonomous driving simulator.
ALIGN is a vision-language framework to infer precise accident coordinates from Bangla news reports and map-based cues. Using an agentic architecture combining OCR, LLM, and vision-language models, the system reduces localization error from 10.9 km to 0.593 km on validation data and 0.465 km on official Dhaka Metropolitan Police records.
WELD is the first emotion dataset in naturalistic workplace context spanning 30.1 months (Nov 2021–May 2024) with 49 employees from a Chinese software company. 733,780 seven-class facial-expression probability vectors validate three established phenomena and reveal six asymmetric emotional regimes. Exposes FER model bias: over-prediction of 'angry' on neutral Asian faces (0.194 vs 0.05).
ProtoSiTex is a semi-interpretable framework for fine-grained multi-label text classification. It combines unsupervised prototype discovery with supervised classification using hierarchical loss. Experiments on a new hotel reviews benchmark and two public benchmarks demonstrate SOTA performance with faithful, human-aligned explanations.
CoLLM-NAS combines two complementary LLMs for neural architecture search: a Navigator LLM guides search direction, a Generator LLM synthesizes candidates. On ImageNet and NAS-Bench-201, the method reduces search costs by 4–10× while outperforming existing NAS methods.
Position paper arguing that AI model evaluations should be grounded in an explicit theory of capability rather than treating scores as direct measurements. Authors empirically demonstrate that reported performance depends strongly on evaluator modeling assumptions and propose an 'Evaluation Card' to document underlying decisions.
ERA, an AI system combining LLM and Tree Search, automatically generates expert-level scientific software. It discovered 40 novel bioinformatics methods outperforming top human-developed approaches, generated 14 epidemiological models surpassing the CDC ensemble for COVID-19 hospitalization forecasting, and produced expert-level solutions for geospatial analysis and neural activity prediction.
EndoCogniAgent is a closed-loop agentic framework for iterative endoscopic diagnosis. It couples fine-grained visual evidence acquisition and multi-step reasoning via self-consistency validation (knowledge and temporal consistency). On EndoAgentBench (6,132 QA pairs from 11 datasets), the system achieves 85.23% accuracy on perception and 71.13% clinical acceptance on reasoning tasks.
Novel approach to emotion modeling using fine-tuned generative language models to output continuous intensity scores (0-100) instead of discrete classification. Demonstrates improved generalization and transfer to sentiment and arousal, particularly valuable for finance applications.
GenoMAS is an LLM-based multi-agent framework for gene expression analysis. Six specialized agents orchestrated via typed message-passing protocols combine structured workflows with autonomous adaptability. On GenoTEX benchmark: 89.13% correlation for preprocessing, F1 of 60.48% for gene identification (+10.61% and +16.85% vs prior art).
CooT is a multi-agent framework using in-context learning for real-time adaptation to unfamiliar partners. Evaluated on Overcooked and Google Research Football, it outperforms population-based methods, gradient-based fine-tuning, and Meta-RL baselines without parameter updates.
Study on latency of computer-use agents on OSWorld: LLM calls for planning and reflection dominate total time. 16 agents tested require 2.7–4.3× more steps than optimal human trajectories. Each successive step takes 3× longer than initial steps.
AMARIS enhances rubric-based RL by integrating persistent evaluation memory. The system accumulates evaluation diagnostics over time, retrieves them via static and semantic search, and continuously adapts reward rubrics. Experiments show performance gains with ~5% time overhead.
Survey on integrating foundation models (LLMs, Gemini, Grok) with federated learning to enable artificial personalized intelligence (API). Proposes personalized federated intelligence (PFI) paradigm combining privacy, generalization, and edge personalization, trustworthy adaptation, and refinement via retrieval-augmented generation.
Automated LLM-based pipeline to generate and tag knowledge components (KCs) for open-ended programming problems. KCGen-KT framework leverages LLM-generated KCs for knowledge tracing. Evaluation on two real-world student code submission datasets shows it outperforms existing KT methods and human-written KCs on future response prediction.
Prompt2Fingerprint introduces a framework for LLM fingerprinting via parameter generation. Instead of fine-tuning each model separately, a specialized generator maps textual descriptions to low-rank parameter increments in a single forward pass, eliminating retraining costs.
Paper presents an AI agent with semantic and episodic memory systems inspired by cognitive science. Authors design and release "The Room" environment (OpenAI Gym-compatible) showing that combining both memory types outperforms single systems. Multi-agent collaboration improves performance.
DashAttention introduces a differentiable hierarchical sparse attention method using adaptive α-entmax transformation to select variable numbers of KV blocks. Unlike NSA and InfLLMv2, it maintains full differentiability and achieves 75% sparsity with accuracy comparable to full attention. GPU-aware Triton implementation provides significant speedup.
Vision-OPD introduces regional-to-global self-distillation to improve fine-grained visual understanding in MLLMs. The framework transfers the model's privileged perception on evidence-centered crops to its full-image policy via KL divergence minimization between token distributions. Competitive results on fine-grained visual understanding benchmarks without external models or ground-truth labels.
Study of 38 models on 8,900 scholarly references: factual recall quality follows a sigmoid combining model size and topic frequency in training data. These two variables explain 60% of variance across dense models, rising to 74-94% within individual model families.
Semantic Generative Tuning (SGT) aligns visual understanding and generation in unified multimodal models by using image segmentation as a generative proxy. High-level semantic tasks improve feature linear separability and visual-textual attention allocation, outperforming decoupled training approaches.
Reversa is a reverse documentation engineering framework converting legacy systems into operational specifications for AI agents. A multi-agent pipeline extracts implicit business rules, synthesizes architecture, and generates traceable specifications with confidence marking. Case study: COBOL-to-Go ATM migration producing 517 claims, 10 identified gaps, and 53 Gherkin scenarios.
COOPO is a hybrid offline-online reinforcement learning algorithm that cycles between KL-regularized offline training and online fine-tuning. Periodic returns to offline training eliminate catastrophic forgetting and distribution drift. On D4RL benchmarks, COOPO reduces online interactions while improving final returns compared to state-of-the-art hybrids.
Study of trade-offs between estimation accuracy, differential privacy, and communication cost in federated learning. Proposes FedHybrid and FedNewton, improvements over FedAvg and FedSGD with finite-sample MSE upper bounds and minimax lower bounds. Evaluation on logistic regression and neural networks (MNIST, CIFAR-10).
Distillation of tabular foundation models (TabICLv2) into boosted trees (XGBoost/CatBoost) for ultra-fast CPU inference. Solves soft target collapse via stratified out-of-fold labeling. Across 153 datasets: 0.882 macro-mean AUC (96.5% of teacher) at 1.9 ms on CPU, 38–860x speedup. Open-sourced as TabTune library.
ZEDA, a self-distillation framework, converts post-trained static MoE models into dynamic variants. On Qwen3-30B-A3B and GLM-4.7-Flash, it reduces 50% of expert FLOPs with marginal accuracy loss and achieves 1.20× end-to-end inference speedup.
Comparative study of tabular foundation models (TFMs) vs classical models on credit default prediction. On Home Credit and Lending Club datasets, context construction strategy (balanced vs uniform sampling) explains more AUC-ROC variance than model choice: +3-4 AUC points. With 5K-10K balanced examples, TFMs match classical GBDTs while improving default-class recall.
Neural network weights constitute a first-class data modality. This position paper proposes treating checkpoints as generative data: synthesizing weights on demand matches or exceeds fine-tuning while reducing adaptation costs by orders of magnitude. High-performing models occupy structured regions of weight space (symmetry, modularity, shared subspaces).
New SPBM method for constrained optimization in deep learning. Combines penalty methods, barrier methods, and exponential dual averaging to handle non-convexity and non-smoothness. Demonstrates effectiveness on fairness, physics-informed networks, and symbolic knowledge integration with linear overhead up to 10k constraints.
SAME is an autoencoder for stereo music and general audio achieving 4096× temporal compression while maintaining reconstruction quality. The architecture combines a transformer backbone, semantic regularization, phase-aware reconstruction losses and improved discriminators. Two variants (SAME-L and SAME-S) are released in open-weights.
CATA introduces a continual machine unlearning method for vision-language models (VLMs). It represents each unlearning request as a task vector and aggregates historical vectors by suppressing conflicting components, ensuring forgetting effectiveness, model fidelity, and persistence against knowledge re-emergence.
Researchers demonstrate typographic attacks against household manipulation robots using CLIP. By placing adversarial stickers, they achieve 67.8% attack success rate on HomeRobot benchmark in Habitat simulation, causing physical grasping and transport errors of wrong objects.
Automatic generation of fuzzy cognitive maps (FCMs) from text using LLM agents that chunk text into overlapping segments. Convex mixing of chunk FCMs produces a cyclic FCM knowledge graph. Operator-level Bayesian inference generates "de-chunked" FCMs. Demonstration on Thucydides Trap model: 7 out of 8 FCMs predicted armed conflict. Gemini 3.1 served as chunking agent.
RAT (Randomized Advantage Transformation) estimates Tikhonov-regularized natural policy gradients via direct backpropagation without explicit Fisher matrix construction. The method applies the Woodbury formula and randomized block Kaczmarz iterations on on-policy mini-batches. Results match or exceed established natural-gradient methods on continuous and visual control benchmarks.
OverEager-Gen is a benchmark measuring out-of-scope actions by autonomous coding agents on benign tasks. On Claude Code, removing the consent declaration raises the overeager rate from 0% to 17.1%. The study validates 500 scenarios across 4 products (Claude Code, OpenHands, Codex CLI, Gemini CLI) and 6 base models.
Study evaluating three off-the-shelf LLMs to estimate difficulty of educational items without response data. Across 6 primary-school mathematics domains, Spearman correlations show moderate-to-strong alignment with empirical difficulties. Pairwise comparisons outperform absolute judgements; adding token probabilities and few-shot examples improves results.
Key-Gram is a conditional-memory framework separating linguistic knowledge from visual reasoning for embodied control. It decomposes instructions into task-specific key-grams, retrieves linguistic priors via O(1) hashed lookup, and injects them into hidden layers. Achieves 29.5% gains on RoboTwin2.0, 35.8% on LIBERO-Plus, 15.4% on real-world tasks.
STT-Arena is a benchmark of 227 interactive tasks measuring LLMs' ability to detect and adapt to spatio-temporal changes. Claude-4.6-Opus achieves under 40% accuracy. Authors identify three recurring failure modes and propose STT-Agent-4B combining iterative trajectory refinement with online RL.
A supervised method called Manifold Probe discovers representation manifolds in superposition within neural networks. Tested on Llama 2-7b, it identifies linear manifolds for time and space, and demonstrates causal control by steering model completions about release years of movies and songs.
Researchers identify Entropy-Gradient Inversion, a negative correlation between token entropy and logit gradients, as a geometric fingerprint of Large Reasoning Models' reasoning capability. They propose Correlation-Regularized Group Policy Optimization (CorR-PO), an RL method embedding this inversion signature into reward regularization, outperforming baselines across multiple reasoning benchmarks.
arXiv study demonstrates that color features alone (RGB/HSV histograms, statistical moments) achieve 89% accuracy in binary cancer/benign classification in histopathology, excluding morphological information. Authors propose these simple features as lightweight pre-screening tool before complex deep learning models.
DBES is a diagnostic framework for evaluating expert specialization in Mixture-of-Experts models. Five theoretically grounded metrics measure domain isolation and routing specialization. Testing on Qwen, DeepSeek, and GLM reveals distinct specialization paradigms. Targeted post-training on specialized expert paths improves performance by 66–94% using only 15% of original training resources.
SafeLens introduces a two-tier video moderation architecture (fast-and-slow) to reduce inference costs. The framework filters SafeWatch dataset to 2.4% via influence-guided filtering and augments with Chain-of-Thought traces. It outperforms SafeWatch-8B, OmniGuard-7B, GPT-5.4, and Gemini-3.1-pro on real and AI-generated video benchmarks.
GAMMA is a quantizer-agnostic mixed-precision framework that automatically allocates bit precision per module without quantization-aware training. Using teacher-forced hidden-state reconstruction and integer programming, it achieves +12.99 Avg. over fixed baselines on Llama/Qwen 8B-32B, matching 3-bit quality at 2.5-bit average.
Prompt2Fingerprint introduces a framework for LLM fingerprinting via parameter generation. Instead of fine-tuning each model separately, a specialized generator maps textual descriptions to low-rank parameter increments in a single forward pass, eliminating the computational overhead of existing methods.
Six modern tabular foundation models form a highly redundant ensemble (mean Q-statistic 0.961). On 153 OpenML classification tasks, the best ensemble (two-level cascade stacking) gains +0.18% accuracy at 253× compute cost. Friedman-Nemenyi analysis places three ensembles and the best single model in the same equivalence group. Greedy selection is recommended as practical default.
Reinforcement learning framework for predicting customer trajectories in retail spaces. RL-based approach outperforms TSP/PNN heuristics (average 28% deviation from shortest paths) by modeling bounded rationality. Validated on real convenience store data: RL predictions better align with observed behavior, more accurate impulse purchase rates and shelf traffic estimates, enabling practical layout optimization.
Paper on mitigating NBTI aging in arithmetic multipliers used in AI. The technique exploits sign-invariance of multiplication to redistribute transistor stress via 2's complement transformations. Integrated into systolic arrays, it improves lifetime with negligible area and delay overhead.
Researchers train KinGPT (25M parameters) on chess data and demonstrate that high benchmark scores of chess-trained LLMs stem primarily from pattern-matching rather than genuine rule understanding. LLM-Modulo, a verifier-in-the-loop framework, improves RedPajama 3B from 1.2% to 21.2% best-move accuracy. Training code, datasets, and model checkpoints open-sourced.
Qumus is the first embodied AI quantum materials experimentalist: an autonomous robotic mini-laboratory capable of hypothesis generation, protocol planning, and experimental execution on 2D quantum materials. It achieved first-time AI creation of graphene and fabrication of atomically thin field-effect transistors via van der Waals stacking, with closed-loop error correction.
SkillsVote is a lifecycle-governance framework for LLM agent skills from collection to evolution. It profiles a million-scale open-source corpus for quality and verifiability, then decomposes trajectories into skill-linked subtasks with outcome attribution. Results: +7.9pp on Terminal-Bench 2.0 (GPT-5.2) and +2.6pp on SWE-Bench Pro.
UbiSLAM presents a real-time mapping and localization system for dynamic indoor environments using a network of fixed RGB-D cameras. This approach reduces computational load on robots and improves navigation accuracy and human-robot interactions, while requiring automatic calibration and optimized communication protocols to handle blind spots.