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.
Theoretical investigation of loss of plasticity (LoP) in deep learning under non-stationary environments. Authors identify two primary mechanisms: activation saturation and representational redundancy creating traps in parameter space. Paradox: properties promoting static generalization (low-rank representations) worsen LoP in continual learning.
Critical audit of SAEBench, the de-facto standard evaluation suite for sparse autoencoders (SAEs). TPP and SCR metrics fail multiple reliability tests and should not be used. Other metrics show higher reseed noise and lower discriminability than assumed. Only sae-probes demonstrates acceptable reliability, but struggles to distinguish architecture variants.
OPERA is a retrieval-augmented generation (RAG) architecture coupling planning and execution via reinforcement learning. A Goal Planning Module decomposes complex questions into sub-goals, executed by a Reason-Execute Module with specialized components for reasoning and retrieval. Training uses MAPGRPO, a GRPO variant. Superior results on complex multi-hop benchmarks.
UniER is a unified benchmark for personalized exercise recommendation, comparing two paradigms: ILER (item-level) and PLER (path-level). The framework introduces Weighted Cognitive Gain (WCG) metric and evaluates 18 methods across 9 datasets. Results show systematic dominance of PLER and reveal ILER's pedagogical failures under extreme sparsity and noise.
A randomized controlled experiment shows LLM access significantly increases average performance, but gains are unevenly distributed. AI Interaction Competence (ability to elicit, filter, and verify outputs) predicts benefits, not GPA. A scaffolding intervention (conceptual maps) reduces outcome variance.
Study of 124 countries covering 99% of global GDP. Task-based automation exposure measure: 3.3% in South Sudan to 61.6% in China. Distinguishes labor-substituting vs labor-augmenting automation. AI more prevalent in substitution in low-income countries, augmentation in high-income. Women disproportionately exposed to substitution.
New credit assignment method for reinforcement learning with LLMs. IBPO (Implicit Behavior Policy Optimization) uses counterfactual trajectories to convert sparse terminal rewards into step-sensitive learning signals, reducing gradient variance and improving stability on mathematical and code reasoning benchmarks.
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.
D²Evo is an RL framework to enhance LLM reasoning. It addresses scarcity of medium-difficulty samples by mining anchors matched to model capability and training a Questioner to generate diverse questions at appropriate difficulty. Results: outperforms existing methods on math benchmarks with <2K real samples.
arXiv paper argues current LLM safety evaluations lack robustness due to small datasets, methodological inconsistencies, and unreliable setups. Systematically analyzes the evaluation pipeline—dataset curation, automated red-teaming, response generation, LLM judges—and proposes guidelines to reduce noise and improve comparability of attack/defense research.
ProRL is a programmatic reinforcement learning framework for combinatorial optimization (job shop scheduling). It generates interpretable policies as human-readable programs via a domain-specific language (DSL-S), exploring the program space through local search and Bayesian optimization. Outperforms classical heuristics and DRL baselines with minimal training episodes.
Novel data selection strategy for LLM alignment based on DPO implicit reward gap. Method selects harder examples (smaller reward gaps) and achieves superior performance with only 10% of original data across multiple benchmarks.
QSTRBench is a benchmark evaluating LLMs' ability to reason with qualitative spatial and temporal reasoning (QSTR). It covers 9 calculi (Point Algebra, Allen's Interval Algebra, RCC-5/8/22, etc.) with composition tables, converse relations, and conceptual neighbourhoods. Tested models outperform guessing but none answer all questions correctly. RCC-22 proves most difficult.
Study reveals an 'unlearnability' phenomenon in Reinforcement Learning with Verifiable Reward (RLVR) for LLMs. Some hard examples remain unlearnable even with correct rollouts. Cross-example gradient analysis shows fundamental representation flaws: low gradient similarity and ungeneralizable reasoning patterns. Data augmentation fails to improve gradient similarity.
EmoMind decodes affective captions directly from brain fMRI signals. The system first retrieves a neutral scene description from brain-decoded visual features, then rewrites it using a continuous 34-dimensional emotion vector extracted from the same fMRI recording. Evaluated on two independent emotion fMRI datasets, EmoMind outperforms GPT-4 with discrete emotion labels across all validation axes.
OCCAM is a framework for explaining black-box image classifier decisions through causal visual concepts. It discovers concepts in open-set manner, localizes them via text-guided segmentation, and measures causal contribution through object-level interventions. OCCAM aggregates interventional evidence to induce a structured ontology revealing concept dependencies and systematic model biases.
LLM cascade systems, designed to balance efficiency and performance by routing complex queries to powerful models, are vulnerable to targeted adversarial attacks. A novel attack exploits lightweight models and internal decision mechanisms to simultaneously degrade accuracy and cost-efficiency.
HINT-SD proposes targeted self-distillation for training long-horizon LLM agents. The method uses full-trajectory hindsight to identify failure-relevant actions and applies feedback-conditioned distillation only on targeted action spans. On BFCL v3 and AppWorld, it improves over dense per-turn feedback baselines by up to 18.80% while achieving 2.26× lower time per training step.
Analysis of 6.8M solutions across 6.7K tasks in 9 domains (materials, structural biology, AI, computational biomedicine, data science, theoretical CS, F1, wheel building). Three universal patterns: heavy-tailed waiting times, sublinear record accumulation, temporal correlation of breakthroughs. Minimal model unifies radical innovation and incremental refinement.
arXiv paper proposing formal framework for combining LLM and human evaluations. Uses doubly robust estimator (missing data literature) to determine optimal number of human reviews needed. Shifts LLM role from substitutive to auxiliary in two-stage sampling design.
First machine unlearning framework for masked diffusion language models (LLaDA, Dream). MDU minimizes KL divergence from prompt-conditional to prompt-masked unconditional distribution at each masked position, with temperature scaling for privacy-utility trade-off. Code released.
EvoSynth, an autonomous multi-agent framework, optimizes jailbreak attacks in executable code space rather than prompt space. The system iteratively evolves and self-corrects code-based attack algorithms. Results: 85.5% Attack Success Rate against Claude-Sonnet-4.5, 95.9% average ASR across evaluated targets.
Comprehensive study of AI-assisted research systems through April 2026. LLMs excel at structured, retrieval-grounded, and tool-mediated tasks but remain fragile for genuinely novel ideas and scientific judgment. End-to-end autonomous systems have not yet consistently reached major-venue acceptance standards. Human-governed collaboration is the most credible deployment paradigm.
F³A is a training-free router for visual token pruning in vision-language models. It selects relevant visual tokens via question-conditioned cues without extra LLM forward passes, reducing inference costs while maintaining performance across model scales.
Infini-News indexes 1.35B CC-News articles (August 2016–present) with metadata extraction, language detection (GlotLID, lingua, CommonLingua), and geographic attribution (83.4% coverage). Infini-gram suffix-array indexes enable sub-second full-text pattern search across the entire archive.
TTG (Token Games) is an evaluation framework where language models challenge each other by creating programming puzzles. The system uses pairwise duels and Elo ratings to compare 10 frontier models. Results match existing benchmarks (Humanity's Last Exam) for under $200 USD without human puzzle curation.
HTSC-2025 is an open-source benchmark of high-temperature superconducting materials discovered 2023-2025 (X₂YH₆ systems, MXH₃ perovskites, M₃XH₈, BCN-doped cage structures, 2D honeycomb). Addresses the lack of standardized datasets for fair comparison of AI algorithms predicting critical transition temperatures.
Theoretical paper on transformer training in mean-field regime (infinite depth and width). Authors model training as controlling a neural PDE (vs ODE for ResNets), establish well-posedness of forward pass, derive explicit formulas for Wasserstein gradients, and prove gradient flow convergence to global minima under NTK injectivity conditions.
DiVT (Disentangled Visual Tokenization) clusters patch embeddings into coherent semantic units for MLLMs, creating discrete meaningful visual tokens instead of continuous streams. Adapts token budget to image complexity, reducing memory and latency while improving LLM compatibility.
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.
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).
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.
StreamPro introduces StreamPro-Bench, a benchmark evaluating proactive video streaming understanding across three dimensions: perception, temporal reasoning, and proactive agency. The framework proposes CB-Stream Loss to address supervision imbalance and applies GRPO with multi-grained rewards. Results: 41.5 on StreamPro-Bench vs 10.4 previously, 78.9 on StreamingBench-RTVU.
Comparative study of 11 LLMs (GPT-4o, Gemma-3, Phi-4, etc.) across 10 common workplace tasks. GPT-4o delivers superior performance but at higher cost and environmental footprint; smaller models (Gemma-3, Phi-4) achieve strong results with better efficiency. Advocates task-aware sufficiency assessments over performance-maximizing benchmarks.
HASP converts textual skills for LLM agents into executable Program Functions that actively intervene in the agent loop at failure-prone states. The framework achieves 25% improvement on web-search vs ReAct and 30.4% gain on math/coding vs Search-R1 through inference-time intervention, post-training, or self-improvement mechanisms.
New IH-GRPO algorithm decouples tool invocation from execution to enhance LLM mathematical reasoning. Achieves 1.87–2.53% improvements on mathematical benchmarks with Qwen3 (1.7B–8B). Code released.
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.
AMARIS introduces persistent evaluation memory to improve rubrics in LLM RL fine-tuning. The system accumulates evaluation diagnostics over time, uses static and dynamic retrieval to contextualize rubric modifications, and adds ~5% time overhead. Experiments show consistent gains across closed and open-ended domains.
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.