3148 articles
FOL2NS is a neurosymbolic framework converting first-order logic formulas into natural language sentences. It handles deeply nested structures with varying quantifier depths rarely captured by existing corpora. Results show fluent generation but increasing challenges with structural complexity.
iPOE is a prompt optimization method guided by automatically generated or human explanations. It creates annotation guidelines that direct optimization through removal, addition, shuffling, and merging operations. On 4 datasets, iPOE improves performance by up to 31% over prompts without guidelines and 35% over random guidelines.
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.
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.
PPAI is an interoperability system for personalized LLM agents on edge devices. It enables P2P collaboration by routing queries to specialized remote agents. The system proposes a prototype-based query-agent scoring mechanism and a multi-agent Bayesian game for load balancing. Results: +7.96% accuracy improvement, 16.34% latency reduction.
RISE, an inference-time semantic reranking framework, improves predictions on hard examples in Rhetorical Role Labeling by leveraging contrastively learned label representations. Without retraining, it achieves +9.15 macro-F1 gain on 8 domain-specific datasets across 7 language models.
PACC (Prosodic Adjustment with Conversational Context) converts read speech to natural conversational speech using deep neural networks and HiFi-GAN. The approach improves intonation, stress, and rhythm for real-time applications (virtual assistants, customer service). New benchmarks established on MOS and speech conversion tasks.
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.
Flow-Direct introduces a training-free guidance framework for flow models using a persistent non-parametric guidance field. Analytically derived from the log-density ratio between base and reward-weighted target distributions, this field accumulates all evaluated samples to improve feedback efficiency and enable reusability without additional reward evaluations.
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.
Study on universal adversarial attacks in NLP. Authors propose a method combining POS filtering and perplexity-based loss to generate natural-sounding triggers. On SST (sentiment analysis), triggers achieve 0.04-0.12 accuracy. Adversarial training improves model robustness from 0.12 to 0.48.
Study of prompt compression on LLaDA, an 8B-parameter DLLM, using LLMLingua-2. Evaluation on GSM8K, DUC2004, ShareGPT at 2× compression ratio shows semantic preservation does not ensure stability in diffusion models: mathematical reasoning degrades substantially while summarization remains robust. Autoregressive compression methods do not transfer uniformly to DLLMs.
PropGuard is a security framework for LLM-based multi-agent systems. It constructs a dual spatio-temporal graph to trace malicious instruction propagation across agents and rounds, then applies source-guided remediation. Tested across four communication architectures and five attack scenarios.
Multi-agent LLM teams outperform human teams in creativity (Cohen's d=1.50) across 4,541 AI ideas vs 341 human ideas on 6 tasks. Advantage driven by novelty while maintaining usefulness. LLM teams benefit from efficient exploration (high semantic spread, shorter paths), while humans benefit from local conversational coherence.
PAREDA is a multi-accent speech dataset (Australian, Indian, Chinese English) featuring spontaneous discussions on NLP papers. Current SOTA ASR models degrade in zero-shot settings, but fine-tuning on PAREDA significantly reduces WER, validating the corpus's richness in technical jargon and conversational phenomena.
SocialMemBench evaluates AI memory systems in multi-party social group settings. The benchmark includes 430 personas, 7,355 conversation turns, and 1,031 QA pairs across 5 social archetypes. Gemini 2.5 Flash reaches 0.721 on small networks; open-source frameworks (Mem0, LangMem, Graphiti, Cognee) plateau at 0.12-0.18, revealing a significant capability gap.
ICD coding model trained on multiple versions (ICD-9 + ICD-10) to predict medical codes. Multi-version training improves micro F1 by 27% on 18K rare ICD-10 codes and macro metrics on 8K frequent codes, with fewer model parameters.
New multilingual ASR approach for Nigerian languages using knowledge distillation. SBPN model (120M and 600M parameters) covers Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian English. Achieves 29% average WER reduction vs monolingual baselines, outperforms SOTA models on Common Voice and Fleurs benchmarks.
arXiv study examining whether LLMs reproduce socio-cognitive power effects observed in humans. Simulated asymmetric dialogues (principal-teacher, judge-lawyer) measure linguistic coordination, pronoun usage, persuasion success, and compliance with unsafe requests. LLMs exhibit these effects with variability.
PUMA detects semantic redundancy in reasoning chains to halt inference in large reasoning models before wasting tokens. The framework combines a lightweight redundancy detector with answer-level verification, achieving 26.2% average token reduction across five benchmarks while preserving accuracy and reasoning coherence.
New offline fine-tuning method for LLMs: Goal-Conditioned Supervised Learning (GCSL). Treats feedback signals as explicit goals and trains models through pure supervised learning, without external reward models. Evaluated on non-toxic generation, code generation, and recommendation tasks.
Orth-Dion improves Dion, a low-rank spectral optimizer for distributed training. By replacing column normalization with QR orthogonalization, it eliminates a √r convergence gap and achieves O(√L_r/T) rate matching exact spectral methods. Validated on large-scale language model pre-training.
UA-StatuteRetrieval: 20-year benchmark on 396M Ukrainian court citations. Co-citation predictability declines 33-47% (Adamic-Adar MRR 0.43→0.29). Non-uniform decay: criminal law stable (~0.40), civil law collapses (0.35→0.15) post-2017 reform. Mid-frequency articles (1K-10K citations) lose 50% predictability. E5-large detects 4.3% semantic drift.
Researchers identify preference instability in reward models under subtle input variations (paraphrasing, pattern injection, backdoors). They isolate unstable features using Sparse Autoencoders (SAEs) and propose two mitigation strategies: SAE Feature Steering and SAE Residual Correction, reducing incorrect preference assignments without retraining the model.
Study of ASR error propagation in ASR-LLM cascades for Korean spoken question answering. Speech recognition errors cause semantic failures that conventional ASR metrics cannot fully capture. Single-character Korean ASR errors create a distinct failure channel. A large audio language model outperforms the ASR-LLM pipeline on noisy Korean SQA.
Study reveals visibility asymmetry in multilingual datasets: 118 languages (59% of 200 most-spoken) have zero catalogued datasets per LRE Map and LDC. Using LLM-assisted citation-mining on Semantic Scholar, authors identify 609 unique datasets across 53 low-visibility languages, 356 openly accessible. Data scarcity is a documentation and discoverability issue, not just production.
Theoretical paper proposing a unified framework for phase transitions in deep learning (grokking, emergent capabilities) and non-equilibrium chemistry. Introduces two gradient fields (entropy production rate and information quasi-potential) and two order parameters (adversarial breakdown threshold α†, self-referential coupling threshold κc) defining a candidate universality class.
TopoPrior learns reusable collaboration structures for multi-agent LLM systems. Using a conditional variational graph framework, it captures structural regularities across domains offline, reducing online search cost and inference-time token consumption while remaining compatible with existing topology-evolution backbones.
New arXiv paper proposing HRC (Hybrid Reward-Cyclic), a reward model decomposing human preferences into transitive (scalar) and cyclic (vector) components via game theory. Introduces DSPPO (Dynamic Self-Play Preference Optimization) for dynamic alignment. Improves RewardBench 2 (+1.23% on Gemma-2B-it) and achieves 44.75% on AlpacaEval 2.0.
OProver is a unified framework for agentic formal theorem proving in Lean 4. The 32B model achieves 93.3% Pass@32 on MiniF2F and 58.2% on ProverBench. Training combines pretraining, SFT on repair trajectories, and RL on hard cases. OProofs contains 1.77M Lean statements and 6.86M compiler-verified proofs.
Study evaluating 4 LLMs for automatic annotation of narrative macrostructure in spoken Mandarin (MAIN). Best model achieves k=0.794 vs k=0.872 (human-human), reduces annotation time by 65%, but struggles with lexical variation and semantic ambiguity. Prompt templates open-sourced.
Residual gap-aware transformer model for predicting Alzheimer's disease progression over 24 months. Trained on 2,600 ADNI samples, it reduces MSE by 13.1% and increases prediction-observation correlation by 26.4% versus linear mixed-effects baseline, combining statistical reference with residual learning from irregular clinical and biomarker histories.
Study on when language models commit to deception. Using counterfactual localization across 5 environments (bluffing, mazes, financial advice, used-car sales, negotiation), authors analyze 1.46M sentences and 91.5B tokens. Lexical cues don't generalize, but attention-based features transfer across domains.
Sign-Muon is a 1-bit distributed optimizer combining majority-vote sign aggregation (signSGD) with Muon's polar-step framework. It reduces communication by 32× in float32 and achieves 92.15% accuracy on CIFAR-10/ResNet-50, with 37% less training time on 4 GPUs.
Systematic optimization of real-time diffusion models on Apple M3 Ultra (60-core GPU, 512 GB unified memory). CoreML conversion of SDXS-512 distilled model combined with 3-thread camera pipeline achieves 22.7 FPS at 512×512 resolution. Demonstrates that CUDA optimization insights do not transfer to Apple Silicon's unified memory architecture.
Auto-CoT improves in-context learning (ICL) by automatically generating reasoning chains for demonstrations. The framework augments prompts with structured intermediate explanations and selects high-quality examples, guiding models toward more reliable reasoning on multi-step tasks.
ConsumerSimBench, a benchmark built from 1,553 Chinese social-media topics and 23,122 reaction criteria, evaluates whether LLMs can reconstruct real consumer reaction patterns. Gemini-3.1-Pro achieves only 47.8% coverage of criteria, revealing a major gap between technical performance and socially grounded consumer intuition.
Agentic AI Translate is an agentic translator prototype replacing the text-in/text-out paradigm with a four-stage cycle (Identify → Prompt → Generate → Verify). Users compose a structured translation brief through model-assisted dialogue grounded in skopos theory, register, and genre conventions. Verification uses the GEMBA-MQM error-span protocol for evidence-grounded scoring.
HalluScore is an Arabic QA benchmark with 827 curated questions to evaluate LLM hallucinations. Empirical analysis of 17 Arabic and multilingual models shows hallucinations extend beyond factual errors to cultural understanding, linguistic reasoning, and logical consistency challenges.
Study on personality induction in LLMs via fine-tuning (SFT, DPO, ORPO) on long-form essays linked to Big Five profiles. Fine-tuning reduces variance in IPIP-NEO questionnaire responses, but accuracy on full personality profiles remains near chance. Unguided essays lack sufficient cues for faithful personality expression.
Response-free item difficulty modelling for multiple-choice questions using fine-tuned transformers. End-to-end approach on item wording eliminates manual feature engineering. Multi-task variant with auxiliary QA objective shows significant improvements on small training sets.
RTPurbo transforms LLMs into sparse models in ~100 training steps. The approach exploits three observations: only certain heads require full attention, long-range retrieval uses a 16D subspace, and dynamic top-p selection outperforms fixed top-k. Results: 9.36× prefill speedup at 1M tokens, 2.01× decode speedup, accuracy preserved.
E-PMQ introduces expert-guided post-merge quantization to efficiently deploy multi-expert merged models at low precision. Using source expert weights as calibration targets and merged-weight anchoring, E-PMQ improves 4-bit GPTQ from 65.0% to 73.6% on CLIP-ViT-B/32 (8-task merging) and from 34.8% to 76.7% on CLIP-ViT-L/14 (20-task merging).
PaliBench is a benchmark for Pali-to-English translation containing 1,700 passages (345,000 tokens) aligned with three independent reference translations. The method combines LLM-assisted alignment, automated verification, and multi-metric evaluation. Evaluation of ten contemporary LLMs shows strong cross-metric concordance but substantial variation in reliability.
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.
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).
Systematic study of lightweight LLMs (<2B parameters) for criminal court view generation and charge prediction. Development of CVGEvalKit, an evaluation framework with 3 public datasets. Comparison of architectures, model sizes, and direct prediction vs. intermediate generation approaches.
Retrieval-based approach for multi-label legal annotation: frozen embedding model with k-NN prediction on documents and label descriptions. On Eurlex (100 labels), Qwen-8B achieves Macro-F1 49.12 vs 40.41 (GPT-5.2 zero-shot), reduces compute 20-30×. Eliminates hallucinations (GPT-5.2: 0.12-0.9% out-of-taxonomy labels).
CHI-Bench evaluates AI agent automation of complex healthcare workflows. Benchmark spans 3 domains (prior authorization, utilization management, care management) with 87 MCP tools and 1,290+ policy documents. Best result: 28% task resolution, 3.8% in single session.
VaCoAl, a hyperdimensional memory built from Galois-field linear-feedback shift registers, provides a unified algebraic substrate for Vector-HaSH and the Tolman-Eichenbaum Machine. The approach maps two VaCoAl regimes to hippocampal EC-CA3 and EC-DG-CA3 pathways, connects deterministic GF(2) binding to Pearl's causal algebra, and derives testable iEEG predictions.
Research paper proposes 6G architecture prioritizing operator control over technology. Introduces '6G Control Compact' (ownership taxonomy) and 'Guarantee Economy' (outcome-priced commercial model). Rakuten Mobile, world's first national-scale fully cloud-native Open RAN deployment, demonstrates feasibility with EBITDA profitability in FY2025.
Amortized Bayesian inference method using transport maps for nonlinear inverse problems. Learns observation-dependent map minimizing averaged energy-distance between true and learned posterior. Likelihood-free, no invertibility constraints. Validated on finite-dimensional and PDE-constrained inverse problems (porous flow, seismic inversion).
SurgicalMamba, a Mamba2-based model, performs online surgical phase recognition with O(d) per-frame cost. Three components address domain-specific challenges: dual-path SSD separating long/short-term regimes, intensity-modulated stepping adapting effective rate, and state regramming enabling cross-channel mixing. SOTA results: 94.6%/82.7% on Cholec80, 89.5%/68.9% on AutoLaparo, 238.74 fps on single GPU.
Paper evaluating energy efficiency of neural vs heuristic combinatorial solvers. Defines Amortized Efficiency Threshold (AET): deployment volume where neural network training cost breaks even. On CVRP (n=50), attention-based solver from Kool et al. (2019) reaches energy parity at ~4560 deployed instances. Per-instance neural-to-heuristic ratio: 2.29e-3.
ARES-LSHADE, an LLM-designed differential-evolution variant, wins 510/744 function evaluations on the GNBG benchmark (GECCO 2026). Combines adaptive CMA-ES mutation and L-BFGS-B polish phase generated via 30 autoresearch experiments. Reaches machine precision on 18/24 functions; 6 remaining show characteristic plateaus. Code released.
Multi-stage framework for detecting reclaimed slurs in multilingual social media (English, Spanish, Italian). Uses XLM-RoBERTa with GPT-4o-mini back-translation data augmentation, dynamic undersampling, and language-specific threshold refinement. Achieves 2-5% absolute F1 improvement without model retraining.
ConSPO, a new sequence-level policy optimization approach, improves GRPO by replacing clipped ratio-based scores with length-normalized log-probabilities and using an InfoNCE-style contrastive objective. Evaluated on mathematical reasoning benchmarks, ConSPO outperforms several RLVR baselines.
PQR is a framework that automatically generates realistic user queries to identify failures in LLM-based QA agents. Through iterative interaction between query refinement and prompt refinement modules, PQR uncovers 23-78% more unhelpful responses than prior methods, with more diverse and realistic queries.
Theoretical paper on evaluating LLM-based agent-based models. Authors propose a four-level Mechanism Plausibility Scale to distinguish generative sufficiency (ability to reproduce phenomena) from mechanistic plausibility (how phenomena are produced), grounded in philosophy of science.
WriteSAE introduces the first sparse autoencoder decomposing and editing matrix cache writes in Gated DeltaNet, Mamba-2, and RWKV-7 recurrent models. Factored atoms expose closed-form logit shifts per token, achieving 92.4% successful substitutions across 4,851 firings on Qwen3.5-0.8B and 88.1% on Mamba-2-370M.
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.
Convergence analysis of deflated Q-value iteration using joint spectral radius (JSR) framework for discounted Markov decision processes. Authors show that deflation with all-ones residual correction reveals sharper convergence geometry by removing a redundant invariant direction, potentially improving convergence-rate characterization beyond the standard γ-bound.
Novel mathematical decomposition of stochastic generative models: deterministic velocity field splits into marginal transport (u_t) and osmotic effect (d_t). Bridge Matching framework learns this decomposition for controllable, interpretable sampling.
Softmax-attention transformers can implement preconditioned Richardson iteration for in-context Gaussian kernel regression. Authors construct a single-head transformer with O(log(1/ε)) blocks achieving ε-accurate prediction on prompts of length N, where softmax attention produces a Gaussian-kernel operator and ReLU MLP layers perform local scalar arithmetic.
MULTITEXTEDIT is a benchmark of 3,600 instances across 12 typologically diverse languages for evaluating text-in-image editing. Authors introduce a language fidelity (LSF) metric detecting script-level errors (missing diacritics, reversed RTL order). Evaluation of 12 systems reveals pronounced cross-lingual degradation, especially on Hebrew and Arabic.
New MoLF (Mixture of LoRA and Full) method for LLM adaptation that dynamically routes gradients between full fine-tuning and LoRA at optimizer level. Tested on Gemma-3-1B, Qwen2.5-1.5B/3B across SQL, Medical QA, and counterfactual knowledge tasks. MoLF-Efficient outperforms adaptive LoRA approaches by 20% (Fact) and 9% (Med/SQL). Code open-sourced.
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.
Behavioral evaluation methodology for agentic AI systems: scoring intermediate decisions via LLM judge ensemble across 6 dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery). Behavioral score correlates at rho=0.72 with Sharpe ratio. Closed-loop reinforcement (SAC) reduces MAPE from 0.61% to 0.54% on 2017-2025 test set.
MA-GIG improves feature attribution by guiding integration paths in the latent space of a pre-trained VAE, reducing exposure to noisy off-manifold regions. Qualitative and quantitative evaluation across multiple datasets shows more faithful explanations than prior path-based attribution methods.
1BT (One-Block Transformer) is a compact model for EEG-based cognitive workload assessment. With <0.5M parameters and 0.02 GFLOPs, it uses a minimal latent bottleneck and single cross-attention module. Tested on 11 participants (abstract reasoning, numerical problem-solving, video game), it achieves high workload classification performance for real-time deployment in resource-constrained settings.
MedSynapse-V proposes a framework for latent diagnostic memory evolution in medical imaging. Through meta-query memorization and Causal Counterfactual Refinement (CCR), the model dynamically synthesizes implicit diagnostic memories aligned with clinical logic. Evaluations across multiple datasets show significant gains over state-of-the-art methods.
Sherpa.ai introduces a multi-party protocol for privacy-preserving entity alignment in vertical Federated Learning. The method hides intersection membership while enabling exact and typo-tolerant matching, without revealing which samples are shared across parties.
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.
CoLLM unifies federated fine-tuning (FL PEFT) and inference on shared edge GPU clusters. The system coordinates real-time parameter sharing via shadow adapter strategies and dynamically balances workloads to optimize model quality and inference latency. Evaluation shows 3x higher goodput vs existing systems.
DPrivBench is a benchmark evaluating LLMs' ability to reason about differential privacy (DP). It tests whether functions satisfy stated DP guarantees under specified assumptions. Strongest models handle textbook mechanisms well but all struggle with advanced algorithms, revealing substantial gaps in DP reasoning capabilities.
A lightweight transformer predicts robot co-speech iconic gestures from text and emotion alone, without audio at inference time. The model outperforms GPT-4o on semantic gesture placement classification and intensity regression on the BEAT2 dataset, while remaining computationally compact for real-time embodied agent deployment.
Temporal modeling framework for predicting student dropout in higher education using LMS engagement data and administrative records. Penalized logistic regression on person-period rows achieves AUC 0.8405 (test). Counterfactual policy-simulation layer evaluates intervention impact on academic survival with limited positive contrasts (max ΔS 0.0819).
Photonic Kolmogorov-Arnold networks (SSP-KANs) implemented with standard telecom components (Mach-Zehnder interferometers, semiconductor optical amplifiers). Four-module network achieves 94.3% on nonlinear classification; seven-module network reaches R²=0.986 on regression. Robust down to 6-bit resolution and 14 dB SNR.
ClawArena is a benchmark evaluating AI agents in evolving information environments. It tests agents' ability to maintain correct beliefs amid contradictory sources, dynamic updates, and implicit user preferences. 12 multi-turn scenarios, 337 evaluation rounds, 5 frameworks and 18 language models assessed.
Researchers localize 'entity cells'—selective MLP neurons encoding entity-specific facts—across seven language models. On Qwen2.5-7B, suppressing a cell selectively erases recall for its matched entity while activating a single cell recovers knowledge even without context. These cells remain stable under aliases, acronyms, and multilingual forms.
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).
AuthorMix introduces a modular authorship style transfer framework using style-specific LoRA adapters mixed layer-wise. Trained on few examples, it outperforms SoTA baselines and GPT-5.1 while better preserving original meaning.
Multi-teacher collaborative framework for forward-looking sonar image semantic segmentation. Combines general teacher with sonar-specific teachers and cross-teacher reliability assessment for pseudo-label validation. On FLSMD with 2% labeled data: +5.08% mIoU improvement over SOTA.
CoDiLA combines diffusion with local autoregressive decoding for parallel code generation. A compact auxiliary AR model (0.6B parameters) ensures fine-grained syntactic coherence while the diffusion model handles bidirectional block-level generation. Eliminates coherence artifacts on code generation benchmarks.
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.
VLM-AutoDrive is a post-training framework for adapting Vision-Language Models to safety-critical anomaly detection in autonomous driving. Fine-tuning on Nexar dashcam videos improves collision F1 from 0.00 to 0.69 and overall accuracy from 35.35% to 77.27% versus NVIDIA Cosmos-Reason1 7B zero-shot.
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).
LightZeroNav addresses zero-shot vision-language navigation in continuous environments using lightweight VLMs. Built on Qwen3-VL-8B, it handles information redundancy, noisy progress estimation, and task entanglement. Achieves GPT-4o-level performance without task-specific training, graph search, or waypoint predictors.
OxyGen proposes unified KV cache management for Vision-Language-Action (VLA) model inference under multi-task parallelism. Implemented on π₀.₅, the system achieves 3.7× speedup on RTX 4090 and Jetson AGX Thor, delivering 200+ tokens/s and 70 Hz simultaneously without quality degradation.
Automated methodology to generate high-performance RL environments using generic prompts, hierarchical verification, and cross-backend policy transfer. Demonstrated on 5 environments (PyBoy→EmuRust, Pokemon Showdown→PokeJAX, new TCGJax). Overhead <4% at 200M parameters.
Researchers identify a shared preference vector in Gemma-3-27B and Qwen-3.5-122B by training linear probes on residual-stream activations. This vector predicts and causally controls the model's task choices across different personas, including an evil persona, revealing a largely shared preference representation underlying different behavioral modes.
arXiv study on automating Motivational Interviewing (MI) session coding for alcohol use reduction. Uses audio-language models (ALMs) with 4 complementary analytic prompts and self-consistency (12 reasoning trajectories per utterance). On 5 sessions, achieves 52.56% accuracy, 54.03% precision, 47.45% recall, macro-F1 46.40%, exceeding baselines.
Study of 189 participants solving matchstick-arithmetic problems. Participants exposed to similar problems (Same group) improve faster and verbalize problem characteristics more. NLP analysis of speech shows transferable insights are distinguished by their accessibility to verbal report.
Investigation of how action information is incorporated into RNN state update functions for reinforcement learning. Authors empirically evaluate several recurrent cell architectures on illustrative domains and identify critical design choices often unreported in large-scale RL agents.
Study of implicit bias of Sharpness-Aware Minimization (SAM) on linear diagonal networks for binary classification. For L=1, both ℓ∞-SAM and ℓ2-SAM recover ℓ2 max-margin classifier like gradient descent. At L=2, ℓ2-SAM exhibits "sequential feature amplification": predictor initially relies on minor coordinates then shifts to major ones, contrasting with GD behavior.
Soohak is a 439-problem research-level math benchmark authored by 64 mathematicians. Gemini-3-Pro reaches 30.4%, GPT-5 26.4%, Claude-Opus-4.5 10.4%. The benchmark introduces a refusal subset evaluating the ability to recognize ill-posed problems: no model exceeds 50%.
Theoretical study of conditional-independence enforcement in regulatory settings (path-specific fairness, classified information, non-public trading). Authors prove that averaged-constraint optimization violates stratum-wise requirements while satisfying averaged ones exactly, making detection hard without access to the decision rule itself.
Systematic evaluation of multimodal LLMs on video anomaly detection (VAD) using ShanghaiTech and CHAD benchmarks. Models exhibit conservative bias in zero-shot settings: high precision but recall collapse. Class-specific instructions improve F1-score from 0.09 to 0.64 on ShanghaiTech, yet recall remains a critical bottleneck for real-world surveillance.
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.
CoCoReviewBench is a 3,900-paper benchmark (ICLR, NeurIPS) to evaluate AI reviewer systems. It addresses metric bias by using reviewer-author-meta-review discussions as expert annotations. Results show AI reviewers suffer from hallucinations and reasoning models are more effective reviewers.