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.
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.
Six modern tabular foundation models form a highly redundant ensemble (mean Q-statistic 0.961). On 153 OpenML classification tasks, the best ensemble (two-level cascade stacking) gains +0.18% accuracy at 253× compute cost. Friedman-Nemenyi analysis places three ensembles and the best single model in the same equivalence group. Greedy selection is recommended as practical default.
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.
Longitudinal study of 235M posts from 4M Bluesky users showing algorithmic feed exposure (News, Science, Blacksky) measurably shapes user language: semantic alignment, register formalization, psycholinguistic restructuring. Reposting is the strongest predictor of linguistic convergence across feeds.
D²Evo is an RL framework to enhance LLM reasoning through self-evolution. The method generates medium-difficulty training samples by mining anchors matched to model capability, then jointly optimizes a Questioner and Solver. Results: outperforms existing methods on mathematical reasoning benchmarks with <2K real examples.
Researchers train KinGPT (25M parameters) on chess data and demonstrate that high benchmark scores of chess-trained LLMs stem primarily from pattern-matching rather than genuine rule understanding. LLM-Modulo, a verifier-in-the-loop framework, improves RedPajama 3B from 1.2% to 21.2% best-move accuracy. Training code, datasets, and model checkpoints open-sourced.
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.
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).
ResDreamer, a hierarchical world model, reconstructs residuals at each layer to progressively abstract visual dynamics in a self-supervised manner. Without domain-specific knowledge, it achieves state-of-the-art sample and parameter efficiency for RL in 3D adversarial environments. Code released.
arXiv paper proposing a formal statistical framework to combine LLM and human evaluations. Uses a doubly robust estimator (missing data approach) to determine optimal sample sizes of human ratings needed for benchmark validation, based on LLM judgment predictability.
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.
Semantic Generative Tuning (SGT) aligns visual understanding and generation in unified multimodal models by using image segmentation as a generative proxy. High-level semantic tasks improve feature linear separability and visual-textual attention allocation, outperforming decoupled training approaches.
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.
Comparative study of tabular foundation models (TFMs) vs classical models on credit default prediction. On Home Credit and Lending Club datasets, context construction strategy (balanced vs uniform sampling) explains more AUC-ROC variance than model choice: +3-4 AUC points. With 5K-10K balanced examples, TFMs match classical GBDTs while improving default-class recall.
SAPO improves generative recommendation by aligning reinforcement learning optimization to individual reasoning steps. Instead of assigning a single advantage to the entire response, SAPO computes separate group-relative advantages for each reasoning step and SID token, stabilizing training and outperforming baselines across three real-world datasets.
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.
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.
Comparative study on surgical AI: multi-billion parameter Vision Language Models fail at neurosurgical tool detection despite extensive training. Scaling experiments show diminishing improvements. Obstacles persist across architectures, suggesting data and compute alone are insufficient.
MirrorBench is a benchmarking framework to evaluate user-proxy agents in conversational systems. It combines 6 metrics (MATTR, Yule's K, HD-D, GTEval, Pairwise Indistinguishability, Rubric-and-Reason) to measure realism of LLM-generated user utterances across 4 public datasets. Open-source code released.
AMARIS enhances rubric-based RL by integrating persistent evaluation memory. The system accumulates evaluation diagnostics over time, retrieves them via static and semantic search, and continuously adapts reward rubrics. Experiments show performance gains with ~5% time overhead.
Theoretical paper formulating multi-step reasoning as s-t connectivity on knowledge graphs. Shows phase transition: if pre-training knowledge is fragmented into small components, augmentation requires Ω(√n) queries; once density threshold is crossed forming a giant component, constant expected queries suffice.
EXG is an experience graph framework for self-evolving LLM-based agents. It organizes successes and failures into structured, relational representations, enabling real-time cross-task experience reuse and offline reuse as external memory. Tested on code generation and reasoning benchmarks, EXG outperforms reflection and memory-based baselines.
EvolveR is a framework enabling LLM agents to learn from their own experiences through a closed-loop lifecycle. It combines offline self-distillation (synthesizing interaction trajectories into reusable strategic principles) and online interaction (actively retrieving distilled principles to guide decisions). Tested on complex multi-hop QA benchmarks, it outperforms existing agentic baselines.
CheckSupport is an open-source system using locally-deployed LLMs to automate reporting checklist recommendation and completion for scientific manuscripts. Evaluated on peer-reviewed manuscripts, it achieves 90% accuracy for checklist recommendations and 88% for item-level completion, processing each manuscript in 12.5 seconds on CPU-only hardware.
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.
arXiv paper on supervised fine-tuning (SFT) effectiveness for LLMs. Authors show SFT primarily removes noise-like token interactions but rarely acquires reliable new ones. The denoising phase is extremely brief; continued fine-tuning introduces overfitted interactions. Implications for early stopping and LLM training.
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.
PAIR is an internal reward model for multi-step LLM training via GRPO. It combines a hidden-state probe (belief consistency) with a lightweight attention head to generate dense step-level reward signals without external model calls or ground-truth dependencies.
DuIVRS-2 is an LLM-based Interactive Voice Response system deployed at Baidu Maps for large-scale POI attribute acquisition. Using FSM-guided data augmentation, selective generation, and Chain-of-Thought mechanisms, the system processes 0.4 million calls daily with 83.9% Task Success Rate and 130ms latency.
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.
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.
SAME is an autoencoder for stereo music and general audio achieving 4096× temporal compression while maintaining reconstruction quality. The architecture combines a transformer backbone, semantic regularization, phase-aware reconstruction losses and improved discriminators. Two variants (SAME-L and SAME-S) are released in open-weights.
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.
FUNCanon breaks down long-horizon manipulation tasks into action sequences (actor-verb-object) and canonicalizes objects by functional affordances using VLM cues. FuncDiffuser, an object-centric and action-centric diffusion policy, learns on aligned data to generalize across object categories and enable cross-task behavior reuse.
FediLoRA introduces a federated LoRA fine-tuning framework for vision-language models (VLLMs) addressing imbalanced LoRA ranks from heterogeneous resources and missing modalities from user errors or device failures. The method combines simple averaging with structured editing, validated on general-domain and medical-domain benchmarks.
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.
Researchers demonstrate typographic attacks against household manipulation robots using CLIP. By placing adversarial stickers, they achieve 67.8% attack success rate on HomeRobot benchmark in Habitat simulation, causing physical grasping and transport errors of wrong objects.
AgroCoT is a VQA benchmark with 4,759 Chain-of-Thought samples designed to evaluate reasoning capabilities of Vision-Language Models in agriculture. Evaluation of 30 VLMs (proprietary and open-source) reveals significant gaps in zero-shot reasoning, highlighting the importance of CoT for precision farming applications.
MoleCode is an LLM-native molecular language representing molecules as explicit graphs with typed entities and direct relations, replacing implicit SMILES strings. Training-free, it improves frontier LLMs on molecular reasoning, editing and generation tasks, especially for unfamiliar molecules, topology-sensitive operations and larger structures.