NavOne reformulates vision-language navigation (VLN) as one-step global path planning on top-down maps. The framework directly predicts dense path probabilities via single end-to-end forward pass, achieving 8x speedup over map-based methods and 80x over egocentric approaches. Introduces R2R-TopDown dataset.
Method to expand LLMs to new languages without costly alignment phase. Converts dense model to Mixture-of-Experts architecture with language-specific experts, then transfers alignment capabilities via post-training delta fusion. Improves performance on new languages while preserving original abilities.
UniER is a unified benchmark for personalized exercise recommendation, comparing two paradigms: ILER (item-level) and PLER (path-level). The framework introduces Weighted Cognitive Gain (WCG) metric and evaluates 18 methods across 9 datasets. Results show systematic dominance of PLER and reveal ILER's pedagogical failures under extreme sparsity and noise.
Benchmark of LLMs on multi-label legal precedent treatment classification. Expert-annotated dataset of 239 real-world citations. Gemini 2.5 Flash achieves 79.1% on high-level classification, GPT-5-mini 67.7% on fine-grained schema. Novel Average Severity Error metric to measure practical impact of misclassifications.
D²Evo is an RL framework to enhance LLM reasoning. It addresses scarcity of medium-difficulty samples by mining anchors matched to model capability and training a Questioner to generate diverse questions at appropriate difficulty. Results: outperforms existing methods on math benchmarks with <2K real samples.
A neuro-symbolic post-training framework trains a 4B-parameter model to generate cubing heuristics for SAT via SFT+DPO. The model achieves pass@5=53 on 100 SAT competition benchmarks, matching the best symbolic heuristic and surpassing Claude-Sonnet-4 (50). Data comes from an MCTS pipeline exploring splitting decisions over SAT competition formulas.
Theoretical analysis of convergence for MARL algorithms in urban traffic control. Authors formalize stability of multi-agent systems using independent Q-learning on each traffic signal, extending single-agent asynchronous value iteration convergence proofs to the multi-agent case via stochastic approximation methods.
TTG (Token Games) is an evaluation framework where language models challenge each other by creating programming puzzles. The system uses pairwise duels and Elo ratings to compare 10 frontier models. Results match existing benchmarks (Humanity's Last Exam) for under $200 USD without human puzzle curation.
LPG (Latent Policy Guardrail) is a safety framework for LLMs that adapts security policies at inference time without retraining. It compresses reasoning into 10 latent tokens, achieves 84.5% accuracy and 77.9% F1 on benchmarks, while running 11× faster than Qwen3-4B-Thinking.
ESI-Bench is a benchmark for embodied spatial intelligence spanning 10 task categories on OmniGibson. Agents must combine perception, locomotion, and manipulation to actively accumulate evidence. Experiments show active exploration outperforms passive approaches, but failures stem primarily from poor action choices rather than weak perception.
DISCA, a training-free inference-time method, culturally aligns LLMs using World-Values-Survey-grounded persona agents. Tested across 20 countries and 7 backbones (2B–70B), it reduces cultural misalignment by 10–24% on MultiTP without weight modification.
Babel is a black-box jailbreak method exploiting a vulnerability in LLM safety alignment: safety relies on sparse attention heads, leaving representational space weakly monitored. Through optimized obfuscation and iterative refinement, Babel achieves 82.67% success on GPT-4o and 78.33% on Claude-3-5-haiku within ~40 queries.
Economic model combining scaling laws and microeconomic theory to characterize profit optimization in LLM training. Analyzes how model size, token budget, and computational costs interact. In compute-bound regime, optimal spending tracks hardware efficiency (FLOPs/$) near-linearly. In data-bound regime, it scales as D²/E.
Economic model combining scaling laws and microeconomic theory to characterize rational behavior of LLM training firms. Analyzes profit maximization under compute-bound and data-bound regimes: in compute-bound, optimal model size tracks hardware efficiency (FLOPs/$) at near-linear rate; in data-bound, optimal training expenditure scales as D²/E.
DiVT (Disentangled Visual Tokenization) clusters patch embeddings into coherent semantic units for MLLMs, creating discrete meaningful visual tokens instead of continuous streams. Adapts token budget to image complexity, reducing memory and latency while improving LLM compatibility.
CarbonScaling is a hardware-aware analytical framework modeling carbon emissions during frontier LLM training. It integrates neural scaling laws, distributed training strategies, accelerator modeling, and operational/embodied carbon accounting. Source code available on GitHub.
LLM cascade systems, designed to balance efficiency and performance by routing complex queries to powerful models, are vulnerable to targeted adversarial attacks. A novel attack exploits lightweight models and internal decision mechanisms to simultaneously degrade accuracy and cost-efficiency.
Spiker-LL is an FPGA accelerator for spiking neural networks (SNNs) enabling adaptive on-device learning. Built on Spiker+ architecture, it implements the STSF local learning rule with minimal overhead. On MNIST/F-MNIST/DIGITS: 93% accuracy, sub-millisecond latency, <0.1 mJ per inference, DSP-free.
STRIDE-AI is a threat modeling framework for assessing generative AI system security. It bridges NIST AI RMF standards and OWASP LLM Top 10, defines a six-phase assessment lifecycle, and operationalizes through a web tool. Validation on a deployed LLM chatbot reduced attack success rate from 80% to 15%.
Multi-agent LLM teams outperform human teams in creativity (Cohen's d=1.50) across 4,541 AI ideas versus 341 human ideas on six tasks. Advantage driven by novelty while maintaining usefulness. LLM teams benefit from efficient exploration (high semantic spread, shorter paths), while humans benefit from smooth conversational flow (high local coherence, frequent pivots).
HINT-SD proposes targeted self-distillation for training long-horizon LLM agents. The method uses full-trajectory hindsight to identify failure-relevant actions and applies feedback-conditioned distillation only on targeted action spans. On BFCL v3 and AppWorld, it improves over dense per-turn feedback baselines by up to 18.80% while achieving 2.26× lower time per training step.
StreamPro introduces StreamPro-Bench, a benchmark evaluating proactive video streaming understanding across three dimensions: perception, temporal reasoning, and proactive agency. The framework proposes CB-Stream Loss to address supervision imbalance and applies GRPO with multi-grained rewards. Results: 41.5 on StreamPro-Bench vs 10.4 previously, 78.9 on StreamingBench-RTVU.
ManiSoft is a benchmark for vision-language manipulation with soft robotic arms. It includes a simulator coupling realistic soft-body dynamics with contact-rich interactions, 4 deformable control tasks, and 6,300 scenes with expert trajectories. Testing 3 policy models shows promising results in clean scenes but substantial performance drop under randomization.
TABOM, a post-training method for Diffusion Language Models, aligns optimization with the multi-step easy-to-hard decoding trajectory observed at inference. Via Boltzmann modeling of unmasking preferences, it derives a tractable pairwise ranking objective that reduces training-inference discrepancy and improves performance on new domains.
WASIL is a dataset of 8,529 turns of Arabic voice interactions with LLMs, including audio, ASR hypotheses, responses, and user feedback (14.2% dislikes). Covers Modern Standard Arabic and 4 major dialects. Enables isolating ASR errors from intrinsic unanswerability via annotation and multi-judge LLM evaluation.
Study reveals an 'unlearnability' phenomenon in Reinforcement Learning with Verifiable Reward (RLVR) for LLMs. Some hard examples remain unlearnable even with correct rollouts. Cross-example gradient analysis shows fundamental representation flaws: low gradient similarity and ungeneralizable reasoning patterns. Data augmentation fails to improve gradient similarity.
arXiv paper proposing a formal framework for combining LLM and human evaluations. Uses a doubly robust estimator (missing data approach) to determine optimal sample sizes of human ratings needed for benchmark validation, shifting LLMs from substitutive to auxiliary role.
Empirical study comparing 11 LLMs (GPT-4o, Gemma-3, Phi-4, etc.) across 10 everyday occupational tasks. GPT-4o delivers superior performance but at higher cost; smaller models achieve strong results with better efficiency. Proposes task-aware sufficiency assessments over performance-maximizing benchmarks.
SonarSweep fuses sonar and vision for underwater 3D reconstruction via plane sweeping. The end-to-end deep learning framework overcomes single-modality limitations by adapting the plane sweep algorithm for cross-modal fusion. Results in simulation and real-world environments, public release of code and first synchronized stereo-camera and sonar dataset.
SteganoPrompt, an open-source web tool, detects verbatim copies of assignment prompts submitted to LLMs. It encodes an invisible instruction in the prompt via the Unicode Tags block (U+E0000–U+E007F), creating a detectable signature in the model's response. Tested across 7 LLM families, the approach bypasses limitations of post-hoc detectors and requires no cooperation from model providers.
HIBCG introduces a theoretically grounded approach to learn sparse coordination graphs in multi-agent RL. Using graph information bottleneck, it determines edge existence and message capacity with formal guarantees on learned topology and differential capacity allocation across agent groups.
Theoretical paper on transformer training in mean-field regime (infinite depth and width). Authors model training as controlling a neural PDE (vs ODE for ResNets), establish well-posedness of forward pass, derive explicit formulas for Wasserstein gradients, and prove gradient flow convergence to global minima under NTK injectivity conditions.
Study of internal preferences in LLMs using linear probes on residual-stream activations. Researchers identify a shared preference vector in Gemma-3-27B and Qwen-3.5-122B that predicts and causally controls model choices. This vector remains stable even when the model adopts radically different personas (helpful assistant vs evil persona).
New credit assignment method for reinforcement learning with LLMs. IBPO (Implicit Behavior Policy Optimization) uses counterfactual trajectories to convert sparse terminal rewards into step-sensitive learning signals, reducing gradient variance and improving stability on mathematical and code reasoning benchmarks.
Comparative study of 6 EEG foundation models across 8 datasets beyond clean accuracy. Robustness analysis (noise, channel dropout), interpretability via Attention-Aware Layer-Wise Relevance Propagation, and expressiveness through block-wise probing. Findings: no single model dominates all failure modes; models focus on task-appropriate brain regions but decode corrupted content poorly.
Adaptive Layerwise Perturbation (ALP) addresses off-policy issues in LLM RL by injecting learnable perturbations into hidden states across all layers. This reduces heavy-tailed importance ratios, stabilizes training, and improves performance on math and multi-turn reasoning tasks while boosting exploration.
Theoretical paper on generative model evaluation. Authors show standard criteria (marginal matching) don't guarantee covariance structure preservation. They introduce D_Sigma = ||Sigma_P - Sigma_Q||_F to measure dependence fidelity, with formal proofs and validation on Fashion-MNIST VAE, RNA-seq (TCGA-BRCA, n=1111), and Alzheimer's data (n=113).
First machine unlearning framework for masked diffusion language models (LLaDA, Dream). MDU minimizes KL divergence from prompt-conditional to prompt-masked unconditional distribution at each masked position, with temperature scaling for privacy-utility trade-off. Code released.
New GPRL (General Preference Reinforcement Learning) method replaces scalar reward models with General Preference Model (GPM) using k skew-symmetric subspaces. Tested on Llama-3-8B-Instruct: 56.51% win rate AlpacaEval 2.0, outperforms SimPO and SPPO on Arena-Hard, MT-Bench, WildBench by preventing single-axis reward hacking.
RoboMME is a standardized benchmark for evaluating memory in vision-language-action (VLA) models for long-horizon robotic manipulation. 16 tasks test temporal, spatial, object, and procedural memory. 14 memory-augmented VLA variants built on π0.5 show effectiveness is highly task-dependent.