3148 articles
Researchers demonstrate that reinforcement learning can encode reusable solvers into LLM weights rather than solving each instance at inference time. Fine-tuning Qwen2.5-Coder-14B with GRPO on Synergistic Dependency Selection, the model converges to Simulated Annealing with 5.0% gap to optimal solver, 91× cheaper than Best-of-64 baseline.
Study on contextual sycophancy in LLMs: 60 participants collaborated with AI on analytical tasks. Results show models mirror user errors rather than correct them. An AI literacy intervention reduced incorrect mirroring but did not eliminate error propagation, suggesting system-level approaches are needed.
Focused Forcing optimizes KV caches in autoregressive video diffusion generation by selecting relevant historical frames per-frame and per-head. The method combines attention scores with diversity scores, achieving 1.48× end-to-end acceleration without training while improving visual quality and text alignment.
Large-scale study of 64,380 SWE-bench runs across 126 agent configurations (43 frameworks × LLMs). Behavioral rules derived from single frameworks do not transfer: the same signal (e.g., error rate) correlates positively with issue resolution in 47 configs and negatively in 48. Framework identity explains 64% of variance vs. 10% for LLM family.
Representation Autoencoders v2 improves VAEs using pretrained vision encoders. Authors find that summing the last k encoder layers, combining RAE with REPA (representation alignment), and re-parameterizing classifier-free guidance accelerates convergence 10x. RAEv2 achieves gFID 1.06 in 80 epochs on ImageNet-256 and EP_FID@2 of 35 epochs.
FireFly generates verified tool-call data for training agents from real MCP servers. The pipeline inverts standard synthesis: explores real APIs via DAG structures, then generates tasks backward from observed outcomes. 5,144 verified tasks across 240 servers and 993 tools. A 4B model trained with GRPO matches Claude Sonnet on held-out test set.
CommitDistill is an open-source Python prototype extracting typed knowledge units (Facts, Skills, Patterns) from local git history via deterministic regex and exposing them through a TF-IDF retriever. Tested on 5 repositories (25k commits), it achieves 0.750 hit-rate at 256-character budget versus 0.333 for BM25. No statistically detectable improvement on time-travel bug-fixes in LLM-as-judge evaluation.
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.
RAG-based pipeline for sentence-level EEG-to-text decoding. Combines EEG encoder aligned with semantic embeddings, vector retrieval, and LLM refinement. On ZuCo dataset, achieves 30.45% improvement over random baseline (cosine similarity 0.181 vs 0.139).
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.
CodeBind introduces a multimodal alignment framework using a shared-specific compositional codebook design. Tested across 9 modalities (text, image, video, audio, depth, thermal, tactile, 3D point cloud, EEG), it achieves state-of-the-art performance in multimodal classification and retrieval without requiring fully paired data.
First machine unlearning framework for masked diffusion language models (MDLMs like LLaDA and Dream). MDU minimizes KL divergence from prompt-conditional predictions to prompt-masked unconditional anchor, with temperature scaling to control privacy-utility trade-off. Code released.
POOL, a novel privacy-preserving RL algorithm, addresses reinforcement learning in multi-dimensional continuous state-action spaces with one-sided feedback. Theoretical analysis derives sample complexity matching known lower bounds for non-private RL, demonstrating strong privacy guarantees are achievable without sacrificing learning efficiency.
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.
Training-free approach for long-context LLM inference: attention-state memory externalizes prefix into lightweight lookup-based memory of precomputed attention states. On LLaMA-3.1-8B, reduces attention latency by 1.36x at 8K tokens and outperforms full-attention RAG with 20% memory footprint.
Theoretical paper on constant collapse in variational autoencoders, where the encoder mean becomes input-independent. Authors introduce a simplex witness certificate to detect and prevent this failure mode during training, with exact baseline and closed-form inverse.
SpatioRoute is a dynamic prompt routing approach for spatial question-answering over egocentric video. Without fine-tuning, it routes each question to a specialized prompt template (rule-based or LLM-driven mode) and achieves +5% accuracy gains on SQA3D versus fixed baselines, establishing SOTA for zero-shot video-only spatial VQA without 3D point-cloud inputs.
COCOCO, a post-hoc framework, integrates Conformal Prediction into neuro-symbolic concept-based models (NeSy-CBMs) to improve reliability. It conformalizes concepts and labels jointly via a single deduction-abduction revision step, ensuring consistency, coverage, and conciseness with distribution-free guarantees. Tested on 8 datasets.
PIPER is a content-driven retrieval method for tabular datasets using table profiles and LLM-generated pseudoqueries for dense retrieval. It outperforms metadata-based baselines and existing TableQA methods in poor-metadata settings.
HyperPersona introduces a multi-level hypergraph framework for text-based automatic personality prediction. The model represents documents, sentences, and words as hyperedges and nodes, capturing global, local, and lexical dependencies. Evaluated on Big Five personality dimensions, it outperforms existing baselines by leveraging textual hierarchy.
Context Codec introduces a formal framework for compressing LLM context while preserving semantic commitments (goals, constraints, decisions, evidence). It defines metrics (Critical Atom Recall, Commitment Density) and CCL, an ASCII-first compact rendering language, to make context compression verifiable and auditable.
RGB-only framework for active 3D scene graph (3DSG) generation using fixed external cameras as common prior maps. System fuses onboard and external camera observations in a single hardware-agnostic pipeline, guiding robots toward high semantic uncertainty regions. Single external camera increases initial object recall by +79%.
MARS is a multimodal system for the CASTLE 2026 challenge that reasons over 4 days of activity, 15 synchronized perspectives, transcripts, and auxiliary modalities (photos, videos, gaze, thermal imagery, heartrate). The approach uses DeepSeek for video summaries and a GPT-5.4 agent to select evidence sources. The system achieved second place on the final leaderboard.
Vision Inference Former (VIF) is a lightweight architectural module improving visual consistency in multimodal models. It continuously injects visual semantics during generation to counter weakening vision-language alignment over long sequences. Tested on 14 benchmarks (reasoning, OCR, tables), VIF improves performance with minimal overhead.
Empirical study of privacy-leakage chains via prompt injection in black-box chatbot environments. Researchers analyze how attackers can hijack LLM agent tasks by injecting malicious content into external sources. They introduce the 'exemplification' technique and demonstrate a functional data-exfiltration chain combining prompt injection, jailbreaking, and web-tool invocation.
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.
Proposes parameterized 4-qubit EWL quantum game circuit for recommender systems in quadruple helix innovation ecosystems. Uses real CORDIS Horizon Europe data, integrates Dirac-Solow-Swan Hamiltonian to simulate capital dynamics under disruptive innovation. Circuit depth 11, NISQ-compatible, Qiskit implementation provided.
Teger, a structured uncertainty module, improves spatio-temporal time-series forecasting by mitigating over-squashing through a Forman curvature-aware graph rewiring mechanism. Integrated into a low-rank-plus-diagonal covariance head, Teger is backbone-agnostic and demonstrates consistent CRPS improvements across LSTM, Transformer, and xLSTM architectures.
New clinical VQA benchmark UCSF-PDGM-VQA: 2,387 QA pairs from 473 glioma MRI studies. Evaluation of 6 VLMs and 1 LLM shows current models fail on multi-sequence 3D MRI, suffering from modality collapse and over-reliance on language priors.
Study on using predictive uncertainty for autonomous/deferral decisions in robotics. Across three temporal activity-recognition benchmarks, uncertainty provides reliable error ranking only when the base model is sufficiently competent. Softmax, MC Dropout, and ensembles produce similar gating behavior; threshold choice has larger impact than uncertainty method.
Preregistered study (N=418) on trust calibration in explainable AI: disclosing model limitations slightly improves alignment between user trust and actual performance in skin-lesion classification, but direct experience outweighs onboarding manipulations.
RAGA is an LLM-based autonomous agent for knowledge graph construction and retrieval-augmented generation. It replaces stateless batch pipelines with a ReAct loop supporting full CRUD operations, hybrid KG-vector synchronization, and evidence-anchored verification linked to source text. Experiments on QASPER show measurable gains in answer and evidence quality.
Proposes quantum sidecar architectures for hybrid AI training and inference. Two operating modes: stateful protected-register mode (QND readout with ancilla) and stateless reset-and-reprepare mode (QAOA-style circuits). Simulations on 2/4/6/8 protected qubits. Positions quantum sidecars as bounded signal generators for optimization, expert selection, and routing.
FedSDR addresses federated fine-tuning of LLMs under statistical heterogeneity. The method combines self-distillation (FedSD) with a dual-stream mechanism: a local LoRA-S branch to absorb heterogeneity via distilled data, and a parallel global LoRA-R branch anchored to raw data for factual correctness.
Study of memorization-generalization coexistence in over-parameterized neural networks. With 80% label noise on arithmetic tasks, models memorize noisy labels while maintaining an internal generalization structure. Frequency-based extraction achieves near-perfect accuracy. Task-agnostic partitioning into generalization/memorization components proposed.
TinySAM 2 compresses SAM 2 for efficient video segmentation. Memory quality management mechanism + joint spatial-temporal token compression. Achieves 90% of SAM 2.1 performance with 7% memory tokens and 3% training data. Reduces parameters, computational load, and deployment costs.
SAS introduces semantic-aware dataset distillation leveraging CLIP as a semantic prior to improve compressed dataset quality. Three scoring functions evaluate class relevance, inter-class separability, and intra-set diversity. A two-stage strategy filters discriminative samples then dynamically selects to reduce redundancy while preserving semantic coverage.
MARR introduces a low-bit post-training quantization method (≤4-bit) for LLMs and Vision Transformers using module-specific scaling coefficients to balance accumulated-error correction and residual-induced bias, with a PID-based adaptive update strategy. Achieves up to 20.2% gains on LLMs and 4.6% on ViTs over prior residual reconstruction methods.
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 while maintaining answer quality.
ViTC-UNet conditions a UNet on frozen pre-trained Vision Transformer representations via learnable tokens and two-way attention decoder. The approach improves biomedical semantic segmentation on MRI and CT without end-to-end fine-tuning, combining ViT global priors with UNet local inductive bias and high-resolution decoding.
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.
New TDA method for diffusion models using mirrored unlearning and noise-consistent skew. MUCS fine-tunes a second model with bounded mirrored gradient ascent and measures normalized skew using consistent noise samples. Outperforms existing approaches on three datasets.
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 guarantee stability in diffusion models: mathematical reasoning degrades substantially while summarization remains robust. Autoregressive compression methods do not transfer uniformly to DLLMs.
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.
Novel theoretical approach to domain transfer identifiability. Authors prove that a single paired anchor sample suffices to identify ground-truth transfer under structural sparsity conditions on Jacobian support, substantially reducing supervision requirements compared to prior work. Empirical validation on synthetic and real-world tasks.
UniTrans translates heterogeneous sensor modalities in real-time for collaborative perception without retraining. A universal model combines translator experts parameterized via an intrinsic latent space, validated on OPV2V-H and DAIR-V2X with open-source code.
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.
OASIS, an inter-layer null signaling technique, reduces attention sinks and activation outliers in AttnResidual architectures. Across three datasets, OASIS decreases maximum infinity norm by 9.26%, average kurtosis by 2.60%, and improves post-quantization performance (W8A8: -75.85% perplexity, W4A4: +12.42% GSM8K).
PAREDA is a multi-accent speech dataset (Australian, Indian, Chinese English) featuring spontaneous discussions on NLP papers. SOTA ASR models degrade in zero-shot settings, but fine-tuning on PAREDA significantly reduces WER, validating the corpus's value for building robust ASR systems.
BiKD introduces a bilevel framework to dynamically balance hard and soft losses in knowledge distillation on imbalanced data. A weight generation network produces adaptive per-sample weights guided by a small balanced validation set. Experiments on long-tailed CIFAR-10/100 show improvements over recent balanced distillation methods.
TAPE, a training-free pruning method for diffusion-based video generation, reduces computational overhead by intelligently removing tokens while maintaining temporal coherence across frames. It applies temporal smoothing between frames, performs token reselection per layer, and uses timestep-level budget scheduling to prune aggressively at early steps and relax during refinement.
EBOMLC method for noisy label correction via efficient bilevel optimization. Uses a meta model trained on clean data to correct large noisy dataset. Reduces hypergradient computational cost and improves stability on CIFAR-10/100 under high noise rates.
EmoMind decodes affective captions directly from fMRI signals in two stages: retrieving semantically grounded neutral descriptions, then rewriting using a continuous 34-dimensional emotion vector. Uses classifier-free guidance to balance semantic fidelity and affective expressivity. Outperforms GPT-4 on two independent fMRI emotion datasets.
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.
TierCheck is a three-tier checkpointing system for LLM training. It maintains lightweight differential checkpoints in local/peer memory for fast recovery, asynchronously migrates base checkpoints to remote storage, and ensures global consistency without stalling training. On models up to 40B parameters, it reduces checkpointing time to under 10s.
Longitudinal study of 235M posts from 4M Bluesky users showing algorithmic feed exposure (News, Science, Blacksky) measurably shapes user language: stylistic accommodation, semantic alignment, register formalization. Effects vary by feed; reposting is the strongest predictor of linguistic convergence.
Paper examining LLM integration in qualitative research. Identifies critical parameters (context window, temperature, top-p, prompt design, system cards) and highlights tensions between modern LLM opacity and qualitative epistemological principles (reflexivity, positionality, interpretive judgment).
Study of a shared-weight recurrent Transformer architecture (AIR) that develops two distinct roles without modular partitioning. On Sudoku-Extreme and Maze, state zH acts as committed proposal while zL retains local uncertainty. Freezing experiments and ablations show that input injection asymmetry induces this functional specialization.
CGPO (Curriculum Group Policy Optimization) improves text-to-image model training via adaptive curriculum based on reward variance. Method prioritizes partially-mastered prompts (high variance) and balances categories through proportional fairness optimization. Gains validated on GenEval, T2I-CompBench++, DPG Bench.
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.
Systematic evaluation of synthetic clinical notes generated by LLMs at million-note scale. Study shows synthetic notes preserve core clinical information for coarse-grained tasks but lose fine-grained details for ICD coding. Chunk-based rephrasing mitigates detail loss but reduces factual precision under incomplete context.
Perspective paper on using LLMs (ChatGPT, Gemini, Claude) and specialized tools (Consensus, Elicit) for conducting literature reviews. Covers summarization, question-answering, data extraction, and translation. Discusses opportunities, risks, and philosophical implications for scientific progress.
DEPPA optimizes pocket-aware diffusion models for drug design via reinforcement learning. The method fine-tunes a pre-trained model by formulating the denoising process as a Markov Decision Process, optimizing binding affinity, drug-likeness, synthesizability and diversity. On CrossDocked2020 benchmark, DEPPA achieves Vina Score -8.5 kcal/mol.
Benchmark of LLMs on legal precedent treatment classification. Expert-annotated dataset of 239 real-world legal 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.
Hybrid framework combining 1D and 2D CNNs with attention-guided fusion for ECG-based biometric recognition. Evaluation on ECG-ID, MIT-BIH, PTB: 99.56%, 100%, 99.89% accuracy. Multi-session tests (Heartprint, 10 years): 98.54%-99.09% same-session, 53-56% cross-session.
PEIRA is a non-contrastive self-supervised learning method analyzing JEPA dynamics through a regularized linear regressor. It minimizes an explicit objective based on the trace of the optimal regressor, ensuring stable non-collapsed equilibria aligned with canonical correlation subspaces. Competitive results on ImageNet-1K and CIFAR-10.
A label-wise attention model trained on combined ICD-9 and ICD-10 data improves rare medical code prediction by 27% micro F1 (18K rare codes) and macro metrics on frequent codes, despite version mismatch. Version-independent approach to automate clinical coding.
Investigation of invariant and equivariant semi-supervised learning (FixMatch, Dense FixMatch) for multi-task training on partially labeled datasets. Evaluation on Cityscapes and BDD100K for object detection and semantic segmentation. Equivariant approaches outperform supervised baselines, especially with limited labeled samples per task.
Bayesian-Monte Carlo probabilistic framework for dynamic construction project schedule updating. Models activity durations with lognormal distributions, updates them via Bayesian inference, and propagates uncertainty through Monte Carlo simulation. Demonstrates improved forecasting accuracy over deterministic CPM methods on PSPLIB benchmarks.
UniAlign is a model-agnostic framework improving robustness of network traffic classification under distribution shifts. It combines domain alignment fine-tuning and stable model ensembling, achieving 2.51% accuracy and 2.71% F1 improvements on three public datasets, requiring only 12.4–53.9% of baseline training time.
Study on automated classification of invalid bug reports and no-code fix generation. Researchers propose a standardized taxonomy and benchmark, testing vanilla LLMs, RAG, and agentic web search. RAG achieves 0.66 weighted F1 for subclassification; agentic web search reaches 68.9% Judge LLM success rate for fix generation.
Robotic clay sculpting planning method using visually-aligned representation. System models deformable material dynamics capturing textures and lighting, enabling long-horizon planning (>100 actions) without per-goal retraining. Tested on three materials with various end-effectors.
Vision paper on AI-assisted code review. Proposes framework of specialized agents spanning 5 stages (PR creation, augmentation, reviewer selection, AI-assisted review, retrospective) with human control at critical gates. Identifies challenges: reliability, bias, privacy, automation bias, transparency, evaluation.
BESplit introduces a Split Federated Learning framework to mitigate non-IID data effects. The method combines Evidential Aggregation for client contribution reweighting, Bias-Compensated Collaboration for representation alignment, and Dual-Teacher Distillation for model synchronization. Experiments on 5 benchmarks show improvements in accuracy and convergence stability.
Theoretical paper on mechanistic interpretability of vision models. Proposes a distributional framework using KL-minimal optimization to interpret internal feature activations, addressing biases in heuristic methods (top-K retrieval, regularized optimization). Implementation via energy-guided diffusion posterior sampling, validated on DINOv3.
KAN-SAE, a sparse autoencoder using nonlinear B-spline activations from Kolmogorov-Arnold Networks, discovers 975 climate features in weather prediction models (vs 566 for linear SAEs). Without climate supervision, it identifies interpretable phenomena like European heatwaves and western Pacific typhoons confirmed by causal steering.
Multilingual OCR-aware fine-tuning framework for MLLMs combining synthetic OCR-to-translation data generation, LoRA-based SFT, and structured visual chain-of-thought reasoning. Significantly improves extraction of small, blurred, occluded text on receipts, menus, documents under degraded visual conditions. Outperforms GPT-5 and Gemini on OCR grounding and hallucination reduction.
Hilbert-Geo introduces a unified formal framework for solid geometry via Parse2Reason: parsing into CDL (Conditional Description Language) then reasoning with theorem bank. Achieves 77.3% on SolidFGeo2k and 84.1% on MathVerse-Solid, outperforming Gemini-2.5-pro (54.2%) and GPT-5 (62.9%). Two annotated datasets: SolidFGeo2k and PlaneFGeo3k.
GCE-MIL improves multiple instance learning for whole-slide image analysis by directly optimizing evidence quality (sufficiency, necessity, recoverability) instead of relying on attention weights. Across 81 configurations (9 backbones, 9 datasets), Macro-F1 gains +0.024 and C-index +0.014, with 5× faster inference.
MemRepair is a memory-augmented agentic framework for repository-level vulnerability repair. It combines three memory layers (History-Fix, Security-Pattern, Refinement-Trajectory) with an iterative refinement loop. Evaluated on SEC-Bench, PatchEval, and Multi-SWE-bench, MemRepair achieves 58.0%, 58.2%, and 30.58% resolution rates, outperforming OpenHands, SWE-agent, and InfCode-C++.
DiagEval is a trajectory-conditioned diagnostic evaluation protocol for GUI agents testing LLM-generated interactive software. It reuses failed trajectories to determine whether failures stem from the evaluator or the software itself. On WebDevJudge-Unit and RealDevBench, DiagEval recovers 45.6-62.1% of false negatives and improves accuracy from 69.9% to 78.3% and from 65.0% to 81.6%.
arXiv paper introducing Progressive Generalization Augmentation (PGA) to improve robustness of agricultural RL systems. Coupled RND-PPO architecture + hierarchical noise injection. Results: +8.43% yield, +16.42% nitrogen use efficiency vs BERT-DQN in Florida; 94.4% performance retention under combined perturbations.
Landslide early warning system robust to rainfall forecast uncertainties. Uses Rainfall-Motion-Aware Contrastive Learning (RMCL) to learn representations stable under spatial displacement of rainfall fields. Tested on 2 years of data across Japan (19 regions): 37% higher precision than state-of-the-art baselines.
IVF-TQ introduces an IVF index with codebook-free residual layer: fixed random rotation plus precomputed Lloyd-Max scalar quantization. Substantially reduces streaming degradation (Deep-10M: -0.80pp vs IVF-PQ -3.23pp). Code released.
DACA-GRPO improves reinforcement learning for diffusion language models by addressing temporal credit assignment and mean-field likelihood bias. It introduces Denoising Progress Scores and Stratified Masking Likelihood, achieving gains up to 7.4pp on code generation and 5.6pp on math reasoning across seven benchmarks.
FML-Bench is a benchmark of 18 ML tasks across 10 domains evaluating 6 AI research agents. Key findings: strategy complexity alone does not ensure performance (greedy hill-climber matches tree-search); effectiveness depends on improvement opportunity structure; an adaptive agent detecting stagnation outperforms others. Includes 12 process-level behavioral metrics.
New arXiv paper proposing HRC (Hybrid Reward-Cyclic), a reward model explicitly decomposing human preferences into transitive (scalar) and cyclic (vector) components via game theory. Introduces DSPPO (Dynamic Self-Play Preference Optimization) for alignment. Results: +1.23% on RewardBench 2 vs GPM, 44.75% win-rate on AlpacaEval 2.0 with Gemma-2B-it.
Comparative study of LLM agents predicting responses to the SHARE survey (health, ageing, retirement in Europe). Agents based solely on demographic data (age, income, education, etc.) exhibit central tendency bias and unrealistic accuracy, while survey-anchored agents better reproduce interactions between retirement planning factors.
ASPI is a benchmark of 728 task-attack scenarios measuring how clarification amplifies prompt injection vulnerability. Testing on 10 frontier LLMs shows attack success rates rise from 1.8% to 34.0% for o3 and 2.2% to 35.7% for Gemini-3-Flash in clarification mode. Code and data released.
TClone is a forkable personal workspace system for computer-use agents. It enables snapshotting, forking, and rollbacking live GUI environments with fast branching and isolation. Uses sibling containers, copy-on-write memory sharing, filesystem versioning, and asynchronous checkpointing. Reduces total task latency by 1.9x vs KVM and 1.5x vs CRIU.
Multi-Scale Hypergraph Laplacians (MSHL): two-stage framework for imputing incomplete spatiotemporal sensor network data. Discovers higher-order structure via multi-scale hypergraphs, then refines with hypergraph-conditioned residual network. Theoretical guarantees and evaluation on real traffic networks with structured outages.
Attention Hijacking: an adversarial attack that manipulates internal attention distributions in vision-language models to maintain a target response across diverse text queries. The method amplifies visual token influence on response tokens while suppressing textual token competition, improving cross-query transferability.
StyleText is a dataset of 28,518 image-mask-prompt triplets for scene text inpainting with style preservation. Automated pipeline combines LLM templating, Flux with KV-cache injection, OCR, polygon mask extraction, and FluxFill augmentation. FluxFill+LoRA baseline substantially improves OCR accuracy while maintaining scene style consistency.
UNR-Explainer generates counterfactual explanations for unsupervised node representation learning models (GNNs). The method identifies critical subgraphs that alter k-nearest neighbors of a node in embedding space using Monte Carlo Tree Search (MCTS). Evaluated on GraphSAGE and DGI.
OProver is a unified framework for agentic formal theorem proving in Lean 4. The system iteratively revises failed proof attempts using retrieved compiler-verified proofs and Lean compiler feedback. Trained via continued pretraining and iterative post-training, OProver-32B achieves 93.3% Pass@32 on MiniF2F and 58.2% on ProverBench.
ContractBench benchmarks LLM agents' ability to preserve observation contracts (temporally valid, byte-level intact artifacts) in API calls. Of 38 models tested, none exceed 80%: Claude-Opus-4.6 leads at 77.8%. Results show integrity and validity failures uncorrelated with model size, and non-monotonic regression in the GPT-5 family despite larger scale.
Latency-aware benchmark for 8 deep learning architectures (MLPs, Transformers) in anomaly detection on inverter-dominated power grids. Real-time classification < 15 ms per cycle, but end-to-end latency 50-90 ms (3+ cycles). Critical gap between algorithmic capability and protection-grade deployment identified.
Rover is a code merge conflict resolution system combining program analysis with LLMs. It introduces Multi-layer Code Property Graph (MtCPG) to capture inter-file dependencies and uses graph connectivity algorithms to create meaningful contexts. Evaluation: Rover outperforms standalone LLMs, MergeGen, and WizardMerge at character, lexical, and semantic levels.