Mechanistic study of in-context learning (ICL) showing structured demonstrations induce concept inference in low-dimensional subspaces. On Llama-3-8B, a 68–73-dimensional subspace of 4096 restores 78.8% of clean–corrupted accuracy gap, while the complementary subspace has zero effect. Results confirmed on Qwen2.5-7B and cross-lingual rule tasks.
IC-Q algorithm for decentralized multi-agent workflow learning under interface constraints. Each agent observes only a local function of shared artifact and private state, with no centralized access to joint trajectories. Finite-sample convergence guarantee for neural Q-learning under decentralized partial observability.
DynaTrain is a distributed training system enabling sub-second online reconfiguration of multi-dimensional parallelism. Using a Virtual Parameter Space abstraction, it reconfigures a 70B dense model in 2s and a 235B MoE model in 4.36s, outperforming existing elastic systems by up to three orders of magnitude.
ReacTOD combines neuro-symbolic and ReAct for task-oriented dialogue. A bounded ReAct loop with symbolic validation iteratively corrects dialogue errors (93.1% self-correction rate), eliminating hallucinations and format errors. On MultiWOZ 2.1: gpt-oss-20B reaches 52.71% JGA (+14pp), Qwen3-8B 47.34%. On SGD: Claude-Opus 80.68%, Qwen3-32B 64.09%.
DecisionBench is a benchmark for evaluating emergent delegation in long-horizon multi-agent workflows. The substrate includes 11 models (7 vendor families), GAIA/tau-bench/BFCL tasks, and multi-axis metrics (quality, cost, latency, routing fidelity). Results show quality alone masks orchestration signals, and delivery channel dominates description content.
UCCI is an LLM cascade router using uncertainty calibration to reduce inference costs. Via isotonic regression, it maps token-level margin uncertainty to per-query error probability, then selects escalation threshold via cost minimization. On 75,000 NER queries with 4B/12B models, UCCI cuts costs by 31% while reducing calibration error from 0.12 to 0.03.
Study of 63 base models reveals hidden phase transition: below ~3.5B parameters, reasoning and truthfulness anticorrelate; above, they cooperate. Architecture, data curation, and training recipe independently shift this critical threshold. Width normalization eliminates anticorrelation; frontier models reach r=+0.72. Open-source steering tool and diagnostic dashboard released.
PRISM is a 10,372 instruction-code pair benchmark for evaluating programmatic video generation by LLMs. It proposes 4 metrics: code reliability, spatial coherence, visual complexity, and temporal density. Evaluation of 7 LLMs reveals a 41% execution-spatial gap: executable code does not guarantee spatially coherent output.
NVIDIA releases Nemotron-Labs-Diffusion, tri-mode model (AR, diffusion, self-speculation) in 3B/8B/14B sizes. Self-speculation combines diffusion drafting and AR verification with shared KV cache: 3× higher acceptance length vs Qwen3-8B-Eagle3, 2.2× speedup, 4× speedup on GB200 (1015 tok/sec with custom CUDA kernels).
BlendedNet++ is a dataset of 12,492 Blended Wing Body (BWB) aircraft geometries with RANS simulations for aerodynamic field prediction. Authors benchmark 5 deep learning architectures (Transolver best) and propose a generative inverse design pipeline using conditional diffusion models, validated by CFD with R² > 0.99.
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.
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.
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.
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.
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.
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).
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.
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).
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.
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.
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).
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).
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.
FinAuditing is a financial auditing benchmark built from 1,102 real XBRL instances (33k tokens average). It evaluates 13 LLMs on three tasks: Financial Semantic Matching, Financial Relationship Extraction, and Financial Mathematical Reasoning. Results reveal substantial gaps in concept retrieval and cross-document reasoning.
ProfBench is a benchmark of 7000+ response-criterion pairs evaluated by human experts in physics, chemistry, finance, and consulting. Authors propose robust LLM-judges reducing evaluation cost by 2-3 orders of magnitude. GPT-5-high achieves 65.9% performance, revealing significant gaps between proprietary and open-weight models.
arXiv paper shows Mixture-of-Experts (MoE) models outperform dense architectures under strictly equal resource constraints (identical total parameters, training compute, data budget). 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).
RLBFF combines human feedback and verifiable rewards for reward model training. The method extracts binary principles from natural language feedback (e.g., accuracy, code readability) and uses them as entailment tasks. Models achieve 86.2% on RM-Bench and 81.4% on JudgeBench (#1 as of September 2025). Qwen3-32B aligned with RLBFF matches o3-mini and DeepSeek R1 at <5% inference cost.
ProfBench introduces a benchmark of 7000+ response-criterion pairs evaluated by domain experts (Physics/Chemistry PhDs, Finance/Consulting MBAs). Top models like GPT-5-high achieve only 65.9% performance. Authors develop robust LLM-Judges reducing evaluation costs by 2-3 orders of magnitude.
Tongyi DeepResearch is a 30.5B-parameter agentic LLM (3.3B activated per token) designed for autonomous long-horizon research tasks. Trained via agentic mid-training and post-training with automatic data synthesis, it achieves SOTA on Humanity's Last Exam, BrowseComp, WebWalkerQA and other benchmarks. Model, framework and solutions are open-sourced.
Med-V1 is a family of 3-billion-parameter language models trained on synthetic data for biomedical evidence attribution and fact verification. It outperforms base models by +27% to +71% on five benchmarks and rivals GPT-5 while being far more efficient. The study quantifies hallucinations in LLM-generated answers under different citation instructions.
Researchers reveal a critical vulnerability in locate-then-edit model editing methods: parameter updates enable attackers to recover edited data via KSTER attack exploiting low-rank structure. A defense using subspace camouflage is proposed to obfuscate fingerprints without compromising editing utility.
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%.
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.
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.
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.
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.
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.
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 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 60-94% of variance. Model proposes recall is gated by signal-to-noise ratio scaling with concept frequency and model capacity.
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++.