AASIST3 enhances speech deepfake detection by integrating Kolmogorov-Arnold Networks (KAN) into the AASIST framework. The model achieves minDCF=0.5357 (closed) and 0.1414 (open) on ASVspoof 2024, doubling prior performance. Code released on HuggingFace.
DocReward is a document reward model evaluating structure and style of professional documents, independent of textual quality. Trained on DocPair (117K document pairs, 32 domains), it outperforms GPT-4 by 14.6 percentage points and effectively guides agents via RL toward higher structural and stylistic professionalism.
Attractor-Vascular Coupling Theory (AVCT): mathematical framework showing cardiac attractor geometry encodes blood pressure information. Calibrated LightGBM model on smartphone PPG achieves MAE 2.05 mmHg (SBP) and 1.67 mmHg (DBP) in strict leave-one-subject-out cross-validation (46 subjects, 29,684 windows), meeting AAMI/IEEE SP10 criteria. PPG-only ablation matches ECG+PPG within 0.05 mmHg.
Circuit analysis of a backdoor in an 8B model: a 3-word Latin trigger redirects English output to French. The circuit operates in 3 phases via attention heads, propagates through a subspace orthogonal to natural language-identity directions, then converts via MLP. A single serial bottleneck position controls the entire flow.
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.
Dual-process memory architecture for scientific agents: decouples episodic window (10 messages) from semantic consolidation (3 tokens/message). Evaluation on 15,000 messages across 6 LLMs (OpenAI, Anthropic, Google): maintains 70-85% accuracy at 10,000 messages with 62% fewer tokens. Identifies trade-offs: Dual Process excels at numeric/temporal queries, RAG for historical retrieval.
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.
GraphMind automates complex operational workflows by extracting action graphs from human resolution traces, then executes them via a multi-agent engine with LLM reasoning. An adaptive reinforcement mechanism (ATR) optimizes successful paths. Deployed across 4 cloud services, the system outperforms a Trace-RAG baseline with a 4.95/5 expert review score.
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.
MEMOIR, a memory-guided tree-search framework, automatically synthesizes solvers for combinatorial optimization using LLMs. With a two-level memory hierarchy (branch-local and global), it achieves 96.7% solution validity across 7 problems (scheduling, routing, packing), outperforming baselines by 9.2 points and reducing run-to-run validity variance by over an order of magnitude.
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.
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.
Systematic study of MoE model compression (Qwen3-Next-80A3B → 23A2B) via pruning and distillation at pretraining scale. Pruning outperforms training from scratch, multi-token prediction (MTP) distillation improves performance, and progressive schedules beat one-shot compression.
Researchers propose the Refusal Index (RI), a metric measuring LLMs' ability to refuse questions beyond their knowledge. RI correlates refusal probability with error probability using Spearman's rank correlation. Testing across 16 models and 5 datasets shows LLMs refuse unreliably despite high factual accuracy.
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.
GiG, a planning framework for embodied agents, uses Graph-in-Graph architecture with GNN to encode environmental states and structure experience memory. A bounded lookahead module enhances planning via symbolic transition logic. Evaluated on Robotouille and ALFWorld, GiG outperforms baselines with +22% to +37% Pass@1 gains.
Language models fail to reliably apply simple rules over long sequences. Test on 126 model variants: all models cannot count above a model-dependent threshold. Failures are abrupt and persist despite increasing model size and computation. Mechanistic probing shows models use finite internal states to simulate counting, exhausting them beyond threshold.
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.
CBT-Audio is a dataset of 1,802 patient turns from 96 public CBT recordings with expert-validated distress labels. Evaluation of 10 open-source audio language models shows audio improves distress estimation over text alone in 8/10 model families, with strongest gains when verbal content and vocal delivery diverge.
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.
Learning-Zone Energy (LZE) is an online data selection framework for RL post-training of LLMs. Tested on Qwen 1.5B-8B across GSM8K and MATH, it retains 40% of training data per step while matching full-data baselines, with OOD gains of +45.9% on AIME25 and 36% FLOP reduction.
BoLT is an open-source benchmark for black-box optimization applied to LLMs. It covers hyperparameters, data mixtures, and prompts via lightweight surrogate models fitted to thousands of real experiments. Benchmarking Bayesian Optimization and BBO methods reveals gaps in existing approaches.
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.
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.
Self-improving CAD agents using finite element analysis (FEA) as feedback. Codex (GPT-5.5) and Claude Code (Opus-4.7) models produce no valid artifacts on first attempt; only ~20% of requirements met. Two supervision signals (text blueprint schema and 21-view renderer) improve iterative loops: Box-IoU rises from 0.444 to 0.592 on S2O.
SLEIGHT-Bench is a benchmark of 40 evasion attacks against LLM-based coding agent monitors. Claude Opus 4.6 with extended thinking catches only 23% of attacks (24/40 never detected). Evasion strategies exploit model priors, instruction ambiguity, and state manipulation.
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.
CyberCorrect formalizes LLM self-correction as a closed-loop control system. A tri-modal error detector (self-consistency, verbalized confidence, logic-chain verification) and type-directed correction controller achieve 79.8% accuracy on CyberCorrect-Bench (440 reasoning tasks), +6.2pp over existing methods, reducing overshoot by 41% via convergence control.
FinTagging is a benchmark for evaluating LLMs on extracting and tagging financial data with XBRL. It decomposes the task into two stages: FinNI (extracting numeric entities) and FinCL (mapping to the full US GAAP taxonomy). Testing shows models extract well but struggle with fine-grained concept linking across 10k+ concepts.
ShareChat is a corpus of 142,808 conversations (660,293 turns) collected from ChatGPT, Perplexity, Grok, Gemini, and Claude between April 2023 and October 2025. The dataset preserves native affordances (citations, reasoning traces, code artifacts) across 95 languages and enables analysis of cross-platform differences in intent satisfaction, citation strategies, and response latency.
AgentKernelArena is an open-source benchmark for evaluating AI agents on GPU kernel optimization. It contains 196 tasks (HIP-to-HIP, Triton-to-Triton, PyTorch-to-HIP) and tests generalization on unseen configurations. Tested agents (Cursor Agent, Claude Code, Codex) achieve speedups up to 6.89x, but show generalization weaknesses on PyTorch-to-HIP.
CAM-Bench is a Lean 4 benchmark of 1,000 computational and applied mathematics problems (optimization, numerical linear algebra, numerical analysis) adapted from textbooks with locally recovered context via dependency-recovery pipeline. Evaluation of LLMs and formalization agents reveals failures in tracking local assumptions and long-horizon control in Lean.
PEGRL is a two-stage RL framework for LLM-based machine translation. It uses post-editing as an auxiliary task to stabilize training and guide optimization. Tests on EN→FI, EN→TR, EN↔ZH show consistent gains; EN→TR achieves performance comparable to DeepSeek-V3.2 on COMET-KIWI.
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%.
ChemVA framework advances LLM understanding of chemical reaction diagrams by addressing visual and semantic bottlenecks. Uses Visual Anchor mechanism for functional group detection and semantic alignment to activate chemical reasoning. Achieves 92.0% structural recognition accuracy on OCRD-Bench with ~20 percentage point gains across 9 diverse LLMs.
Stream2LLM is an LLM serving system that reduces time-to-first-token (TTFT) by overlapping context retrieval with inference. It handles two modes: append (progressive accumulation) and update (iterative refinement). Evaluation on real workloads shows up to 11x TTFT improvement.
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.
ClawGym is a framework for developing agents capable of executing multi-step workflows over local files and persistent tools. It includes ClawGym-SynData (13.5K synthesized tasks), models fine-tuned via supervised learning, and ClawGym-Bench (200 evaluation instances). Code and resources released.
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.
MADP is a multi-agent architecture for enterprise document automation, combining deep learning classification and LLM extraction with human validation. Deployed on 955 real documents, it achieves 97% full-pipeline automation and reduces FTE requirements by 70%. 98.5% document-level accuracy with human-in-the-loop; 69% CO2 reduction vs manual processing.