COSMO-Agent, a tool-augmented RL framework, trains LLMs to orchestrate iterative CAD-CAE processes. The system learns to generate parametric geometry, solve simulations, and revise designs under multiple constraints. Industry-aligned dataset covering 25 component categories. Trained small LLMs outperform large open-source and closed-source models in feasibility and stability.
ConTact explicitly decomposes antibody CDR design into three stages: surface complementarity fingerprints, CDR-antigen contact prediction, and contact-gated feature injection. On CHIMERA-Bench, the model achieves 7% RMSD improvement, 10% F1 gain in epitope awareness, and 0.38 AAR sequence recovery over baselines.
ScenePilot generates critical scenarios for autonomous driving testing via multi-objective reinforcement learning. The framework combines RSS-derived physical feasibility with an AV-risk predictor to target boundary-band scenarios: physically solvable yet causing failures. Results: +6.2 percentage points collision rate on SafeBench while preserving physical validity.
ArabDiscrim is a corpus of 293K Arabic Facebook posts (2014-2024) on racism and discrimination. It includes 200 curated terms with morphological families (13+ inflections), 20 discrimination axes, and native engagement signals (reactions, shares, comments). Released under restricted research-use license for ethical compliance.
Token-weighted DPO (TwDPO) and AttentionPO introduce preference optimization that weights tokens by importance. AttentionPO uses the model's own attention to estimate weights without a separate reward model. Results: improvements on AlpacaEval, MT-Bench, ArenaHard.
ACC converts agent trajectories (search, software engineering, database querying) into long-context QA pairs for SFT training. Removes tool response masking and creates explicit supervision over distant dependencies. Qwen3-30B-A3B achieves +18.1 on MRCR and +7.6 on GraphWalks, comparable to Qwen3-235B.
Study across 6,620 runs showing Claude Haiku compresses 10 English intensity modifiers into 5 distinct outputs. System state context dominates lexical effect (explained variance: 0.782 vs 0.079). Near operational boundaries, model exhibits three modes: small adjustments for weak words, abstention for strong words, ceiling-pushing for 'drastically'.
OGCaReBench is a retrieval-focused benchmark evaluating LLMs on off-guideline clinical questions extracted from published medical case reports. GPT-5.2 achieves 56% without retrieval, 82% with retrieved medical articles. Specialized models reach only 42%.
RankJudge is a benchmark generator for evaluating LLMs-as-judges on multi-turn conversations grounded in reference documents. The system creates conversation pairs with a single flaw injected into one turn, enabling unambiguous labeling. Evaluation of 21 frontier LLM judges ranked via Bradley-Terry model across machine learning, biomedicine, and finance domains.
Sem-Detect detects AI-generated peer reviews by analyzing textual features and semantic analysis at claim level. The method compares a target review against multiple AI-generated reviews of the same paper, exploiting AI model convergence versus human reviewer diversity. On 20,000+ ICLR/NeurIPS reviews, Sem-Detect improves strongest baseline by 25.5% in TPR@0.1% FPR.
arXiv paper on data scaling laws: progressive coverage of a latent predictive contribution spectrum (via suffix-automaton representation) strongly correlates with empirical scaling exponent. Across 12 real corpora, log K(N) shows near-linear relationship with log N (R²≈0.96), suggesting training advances an effective frontier through a predictive state spectrum.
DeepWeb-Bench is a deep research benchmark evaluating 9 frontier models on tasks requiring massive evidence collection, cross-source reconciliation, and long-horizon multi-step derivation. Errors stem primarily from derivation and calibration (>70%), not retrieval (12-14%). Strong and weak models fail differently: incomplete derivation vs hallucinated precision.
SpecHop accelerates multi-hop retrieval agents by maintaining multiple speculative threads with faster but less reliable tools, asynchronously verifying predictions and committing/rolling back branches. The framework reduces latency by up to 40% while preserving accuracy and final trajectory, approaching oracle latency gains with sufficient active threads.
PALS is an LLM inference optimization system integrated into vLLM that treats GPU power caps as a controllable parameter. By combining offline power-performance models with feedback-driven control, it improves energy efficiency by up to 26.3% and reduces QoS violations by 4x to 7x across dense and mixture-of-experts models.
AutoRPA distills decision logic from LLM agents (ReAct paradigm) into robust RPA functions using a translator-builder pipeline and retrieval-augmented generation. On repetitive GUI tasks, generated functions reduce token usage by 82–96% while maintaining performance.
Anthropic leases Colossus 1, xAI's supercomputer, for $1.25B/month through May 2029 ($40B+ total). The contract caps Anthropic's access at 11% of cluster capacity, restricting the company to a fraction of available resources.
WebVoyager benchmark on 720 browser agent tasks: MiniMax M2.5 costs 2.3× less per successful task than Gemini 2.5 Flash. GLM-5 achieves 57.1% accuracy, Kimi K2.5 shows 0% parse retry rate. Open-weight models outperform Gemini not through intelligence but reliability. True cost exceeds per-token pricing once retries compound.
Tencent releases Hy-MT2, a multilingual translation model family in three sizes (1.8B, 7B, 30B-MoE) supporting 33 languages. The 1.8B model compressed to 440 MB via 1.25-bit quantization outperforms commercial APIs from Microsoft and Doubao. The 7B and 30B variants exceed DeepSeek-V4-Pro and Kimi K2.6 performance. Includes IFMTBench benchmark and WMT26 partnership.
Masked diffusion language models (MDLMs) outperform autoregressive LLMs as world models for agentic RL. Fine-tuned SDAR-8B and WeDLM-8B achieve 4x gains on BLEU-1/ROUGE-L/MAUVE. GRPO training yields +15% absolute task-success on ScienceWorld, ALFWorld, AppWorld with Qwen3, Mistral, LFM2.5 in zero-shot transfer.
FusionCell predicts standard-cell performance by fusing routed layout geometry (DeiT encoder) and netlist topology (graph transformer). Trained on 19.5k 7nm cells (ASAP7), the model achieves 0.92% MAPE on delay/power metrics and accelerates characterization by orders of magnitude versus circuit simulation.
TabPFN-MT extends Prior-Data Fitted Networks to multitask in-context learning for tabular data. Trained on multi-target synthetics, the model captures inter-task dependencies and reduces inference from O(T) to O(1) forward passes. On 344 datasets (<1000 samples), it achieves rank 4.89 in multitask accuracy, competitive with single-task ensembles.
Study of 6,233 MedGPTs and 10 open-source models deployed on the web. 25-30% show low factual accuracy, 33.6-54.3% violate operational thresholds, 57% of Action-enabled models lack privacy disclosures. Authors introduce MedGPT-HEval for hallucination detection and release HAA-MedGPT, a structured dataset.
Corpus-centric diagnostic framework for analyzing biomedical named entity recognition (NER) and entity linking (EL) benchmarks. Applied to 9 corpora, reveals substantially different properties can mask apparently identical tasks. Open-source code and interactive dashboard provided.
CSA (Conformal Selective Acting) is a deployment wrapper for RLVR-fine-tuned LLMs guaranteeing per-round risk control without pooling across deployments. Tested on 480 specialist streams and 10,300 Expert-Iteration rounds with LoRA, CSA maintains a Ville e-process per threshold and achieves selective-risk bound R_T^act ≤ α+O(N_T^{-1/2}) with anytime pathwise validity.
FlowLM converts pre-trained diffusion language models into flow matching models via efficient fine-tuning. By realigning curved diffusion trajectories into straight-line flows, FlowLM achieves high-quality few-step text generation rivaling 2,000-step diffusion sampling. Performance saturation is reached with half the training epochs compared to training from scratch.
arXiv paper showing that real-data scaling laws are governed by progressive coverage of a latent predictive contribution spectrum, beyond token-frequency tails alone. Using suffix-automaton representation, authors define a global-KL spectrum and demonstrate strong correlation (R²≈0.96) between spectrum slope and empirical scaling exponent across 12 corpora.
Mix-Quant introduces phase-aware quantization for agentic LLMs: FP4 during prefilling (3x speedup) and BF16 during decoding. This approach alleviates the computational bottleneck in agentic workflows while maintaining task performance on long-context and multi-turn benchmarks.
Study of instruction-following vs. pattern-completion conflict across 13 LLMs. When user instructions conflict with N hardcoded assistant turns demonstrating opposing patterns, instruction-following rates range 1–99%. Transition is universal but model-dependent. Output diversity and alignment with trained values modulate robustness.
SCRIBE is a diagnostic framework for Indic ASR that decomposes errors into categories (lexical, punctuation, numerals, domain entities) instead of WER. Sandhi-tolerant alignment with domain vocabulary injection. Open-weight rich transcription models released for Hindi, Malayalam, and Kannada.
arXiv study on Chain-of-Thought (CoT) impact on gender bias in LLMs. Researchers combine benchmark evaluation, mechanistic interpretability, and reasoning chain analysis. Finding: CoT does not consistently reduce bias gaps; observed improvements stem from memorization rather than genuine understanding, with gender bias remaining embedded in hidden representations.
Geometry-Lite is a compact safety probe analyzing hidden-state geometry across LLM layers (1.2B–70B). It maps layer-wise margins via centroid, local-neighborhood, and supervised linear-boundary readouts, showing that unsafe-prompt detection relies primarily on persistent margin geometry rather than layer-to-layer motion signals.
GROW adapts GRPO (Group Relative Policy Optimization) for VLM agents by decomposing trajectories into state-action samples to avoid excessively long contexts. Tested on 800+ Minecraft tasks, the method achieves SOTA in multi-turn RL for open-world agents.
OmniISR proposes a unified framework for centralized and federated learning via intermediate supervision and regularization. The framework uses mutual information to align internal covariate shifts and negative entropy to regularize overconfident predictions. O(1/sqrt(T)) convergence guaranteed theoretically; CL-FL gap reduced by 22.60% in experiments.
MTR-Suite is a framework for auditing and synthesizing conversational retrieval benchmarks. It includes MTR-Eval (LLM-based auditor), MTR-Pipeline (multi-agent system generating high-fidelity dialogues at 1/400th human cost), and MTR-Bench (general-domain benchmark). Code and data publicly available.
Direct sign-to-sign translation between ASL, CSL, DGS using MBART model trained on synthetic pairs via back-translation. Direct S2S outperforms cascaded baseline (20% lower DTW-aligned MPJPE, 50% higher BLEU-4) with 2.3× speedup.
HRM-Text replaces standard Transformers with a Hierarchical Recurrent Model decoupling slow strategic and fast execution layers. A 1B model trained on 40B tokens and $1,500 achieves 60.7% MMLU, 81.9% ARC-C, 82.2% DROP, 84.5% GSM8K, 56.2% MATH — 100-900x fewer tokens and 96-432x less compute than baselines.
Two-stage pipeline for captioning cultural images in Indigenous languages: Qwen2.5-VL generates Spanish intermediate caption, then Gemini 2.5 Flash produces target-language caption via retrieval-augmented prompting. Achieves 164.1% (Bribri), 131.7% (Guaraní), 122.6% (Orizaba Nahuatl) improvements over baseline. Overall winner of AmericasNLP 2026 shared task.
Expert study (45 scientists, 469 hours) evaluating 2,960 criticisms from 82 Nature papers. GPT-5.2 outperforms top human reviewer (60.0% vs 48.2%), but AI shows 16 recurring weaknesses (limited subfield knowledge, poor long-context handling). AI reviewers complement rather than replace humans.
DIVE compresses LLM embeddings through lightweight adapters using self-limiting triplet loss and NT-Xent contrastive loss. Outperforms Matryoshka-Adaptor, Search-Adaptor, and SMEC across 6 BEIR datasets at all compression ratios. 14M-parameter open-source implementation.
New d_NTP metric evaluates task vector quality in ICL by measuring alignment of next-token probability distributions. Linear Task Vector (LTV) method minimizes d_NTP via closed-form linear regression, improves accuracy by 9.2% across 8 benchmarks and 5 LLMs, reduces inference latency. Task vectors transferable across model scales (+6.4% for smaller model).