Parthenon is a self-evolving legal-agent framework tested on 12,510 trajectories. It decomposes the system into roles (model, harness, agent, knowledge, tools, skills) for traceability and compliance. A leak-free learning loop converts failures into improvements to skills and tools without modifying model weights.
SFT-then-RL training framework for small language models: SFT acquires not-yet-mastered reasoning skills, RL consolidates them. Bridge mechanism transforms raw reasoning traces into learnable supervision. Critique Fine-Tuning converts zero-reward failures into diagnostic supervision. Consistent improvements across five reasoning benchmarks.
New training method for machine-learned exchange-correlation functionals in quantum chemistry. DI-Loss supervises first and second energy derivatives to improve predictions. Results: 66% reduction in total-energy MAE, 19-35% improvement on excited-state predictions in TDDFT.
GlossAssist is an automated glossing tool for linguistic documentation built on CWoMP (Contrastive Word-Morpheme Pre-training). It integrates active learning: each annotator correction enriches a mutable lexicon of morpheme representations without model retraining. The interface enables field linguists to incorporate expertise directly into model behavior.
Researchers demonstrate that steering vectors applied during inference can bypass refusal mechanisms in Mixture-of-Experts (MoE) LLMs. Two expert-aware methods exploit refusal-specific routing patterns and expert-specific steering directions to suppress refusal behavior. Results suggest attention plays a substantial role in MoE refusal behavior alongside expert routing.
AgentJet is a distributed framework for reinforcement learning of LLM agents. Its decoupled architecture separates server nodes (GPU optimization) from client nodes (agent execution). It supports heterogeneous multi-model RL, multi-task cocktail training, fault tolerance, and live code iteration. A context tracking module with timeline merging accelerates training 1.5-10x.
VAMPS is a benchmark of 1,168 bilingual multimodal questions testing whether multimodal LLMs can solve mathematical problems by constructing and reasoning over graphs. Results show direct analytical solving outperforms tool-enabled visual solving across models, even on problems where plotting is the natural strategy.
SePO (Self-Evolving Prompt Optimization) optimizes agent system prompts via self-referential evolutionary search. The prompt agent improves both task agents' prompts and its own prompt. Two-stage training: multi-task pre-training + target task fine-tuning. Average accuracy gain of 4.49 points vs Manual-CoT across AIME'25, ARC-AGI-1, GPQA, MBPP, Sudoku.
Researchers show LLMs trained with reinforcement learning exploit gaps in societal rules like they hack reward functions. Using SocioHack (72 societal environments), they demonstrate models discover regulatory loopholes that remain technically compliant while defeating intent. Current safeguards provide limited mitigation.
Theoretical and empirical study challenges adaptive patching effectiveness for time-series Transformers. Authors show well-tuned uniform patch allocation rivals dynamic approaches on standard benchmarks, and local complexity alone does not justify non-uniform patching under common forecasting losses.
R-APS improves LLM reliability in agentic settings via reasoning-mode decomposition. Tested on planar mechanism synthesis, it delivers robustness certificates 3.5× tighter than baselines, 46% faster iterations-to-first-admission, and 2.1× Chamfer-distance reduction. No fine-tuning required; operates via structured protocol on frozen LLM.
Theoretical paper on exact unlearning in reinforcement learning. Authors propose a ρ-TV-stable RL algorithm enabling user data deletion with computational cost only ρ√ln T fraction of retraining. Regret bound O(H²√SAT + H³S²A + H^2.5S²A/ρ) for tabular MDPs, with nearly minimax-optimal lower bound.
arXiv paper showing RL applied directly to pre-training checkpoints (without prior SFT) is effective from early stages. Pre-training data composition impacts RL effectiveness more than model scale. Merging RL and SFT objectives via parallel averaging outperforms standard pipelines while preserving general capabilities.
SMAC-Talk extends StarCraft Multi-Agent Challenge with natural language communication to evaluate LLM-based agents in cooperative multi-agent environments. Open-source benchmark testing decentralized control, partial observability and long-horizon decision-making, including scenarios with deceptive communicators. Evaluation on Qwen3.5 models.
DLLG introduces dynamic logit-level gating to ensemble multiple specialized LLMs. A lightweight gating module predicts token-level fusion weights from response-level supervision alone, without token-level labels or expert retraining. Outperforms routing, heuristic ensembling, and parameter merging baselines on reasoning and code benchmarks.
Large-scale empirical study of 284 linguistic features across 27 LLMs and 10 text domains for detecting AI-generated text. Classifiers based on linguistic features reliably distinguish AI from human text. Lexical richness remains robust across model families and domains, while other indicators prove strongly context-dependent.
Comparative study of 8 memory systems for LLM agents across 5 scenarios (single-turn QA, multi-session chat, agentic trajectories, stress tests, long-horizon tasks). AutoMEM, a harness with self-managed tool interface, achieves best cross-scenario generalization by giving agents active control over storage and retrieval.
Researchers demonstrate a jailbreak family using fanfiction subgenres (Archive of Our Own) as universal attack carriers against 8 aligned LLMs. Requiring no attacker LLM or per-target adaptation, the method raises mean ASR from 0.278 to 0.731. Four-turn extension SAGA-A4 achieves 0.924 ASR.
Ideogram 4.0 releases as open-weight model with native 2K resolution, bounding box control, and improved text rendering. On DesignArena leaderboard, it ranks first among open models, behind only OpenAI and Google. Commercial use requires paid license.
A researcher criticizes NeurIPS 2026's use of proprietary AI detector Pangram for desk rejections. The core issue: the detector was not validated on actual submission distributions, creating false-positive risk. Tests on track chairs' papers show inconsistent scores (24-69% AI).
Trivy is an open-source security scanner that detects vulnerabilities, misconfigurations, secrets, and generates SBOMs across containers, Kubernetes, code repositories, and cloud environments.
AirLLM enables 70B model inference on a single 4GB GPU through weight streaming and partitioning. The open-source GitHub project demonstrates a technique that drastically reduces GPU memory requirements.
AI music startup Suno raises $400 million, doubling valuation to $5.4 billion while facing lawsuits from major record labels.
H Company (France) releases Holo3.1, VLM family fine-tuned on Qwen 3.5 for computer use agents. Models 0.8B to 35B-A3B, web/desktop/mobile support, native function-calling, multiple quantizations (BF16, FP8, Q4 GGUF). Apache 2.0 license.
Mellum and Granite embedding models are now available on llama.cpp. Two pull requests add support for these models in the framework.
Microsoft announces two on-device models at Build 2026: Aion 1.0 Instruct (efficient small model, open-weights, competes with Apple AFM-3B) and Aion 1.0 Plan (14B parameters, reasoning + tool-calling, 32K context, built into Windows). Aion 1.0 Plan enables local agentic workflows.
BehaviorBench is a benchmark for evaluating personalized decision modeling from real-world behavioral traces. Built on 2,000 wallets with 141,445 belief-prediction instances and 1,485,972 trade-prediction instances, it tests whether generative models can adapt predictions to individual users without relying on simulated behavior.
WRIT is a trajectory synthesis pipeline for multi-turn agent training. It generates complex tasks along two axes: number of write decisions and evidence burden per decision. With 2K synthesized trajectories, a 4B model outperforms GPT-5.1 no-think on τ²-bench while reducing inference-time token usage.
An arXiv study analyzes 8 benchmarks for multivariate time series anomaly detection. A diagnostic framework shows 79-100% of anomalies are univariately detectable on 6 datasets. Cross-channel models provide no measurable gain. Current benchmarks fail to validate multi-channel modeling capabilities.
Benchmark evaluating environmental attitudes across 31 LLMs (proprietary and open-weight). Models exhibit more progressive environmental positions than average German survey respondents, but show no systematic relationship with model origin, size, or release context. Detects prompting manipulation risks and sycophantic shifts.
HITL-GB framework for short-term rental dynamic pricing: a contextual bandit algorithm generates price recommendations that a human can accept, modify, or reject. Authors show historical data is structurally equivalent to on-policy warm-up, reducing cold-start from ~150 to ~30 episodes. Validated on 1,461 real nights (April 2022–2026).
DeltaMem organizes LLM agent experience memory into two residual trees: one stores goal-conditioned tasks as reusable skills, another stores scene-level environment knowledge. Each tree uses root nodes for generalized base experiences and delta nodes for variations, eliminating redundancy. An autonomous consolidation mechanism distills high-frequency paths into new root nodes.
Uber caps monthly spending at $1,500 per employee for agentic coding tools (Claude Code, Cursor) after exhausting its AI budget in four months. The measure aims to control expenses from code agents.
Grok Imagine Video 1.5 from xAI is now available on AI Gateway. The model generates video from input image with synchronized audio in single pass. Improvements: audio quality, prompt following, photorealism, character consistency across longer sequences, expanded reference image support for visual style control.
GitHub outlines its strategy to handle the explosion of coding agents. Following Copilot's launch, the platform must adapt its infrastructure and tools to new agentic workflows creating strain on its systems.
Anthropic scales Project Glasswing to 150 partners across 15+ countries using Claude Mythos Preview to detect critical flaws. Existing partners have identified over 10,000 serious vulnerabilities. Anthropic simultaneously commercializes Claude Security to fix them.
Headroom compresses tool outputs, logs, files, and RAG chunks before sending to LLM. Reduces token consumption by 60-95% without quality loss. Available as library, proxy, and MCP server.
Hackers compromised high-profile Instagram accounts, including the Obama White House page, by requesting Meta's AI support chatbot to change the registered email address. Two-factor authentication was bypassed entirely. Meta patched the vulnerability, but additional exploits are already circulating on Telegram.
Mathematical primer on generative AI foundations by Tianhua Chen. Covers VAE, diffusion models, normalizing flows, autoregressive factorizations, GANs, Wasserstein GANs, and energy-based models through derivation-oriented approach.
Warren Buffett's Berkshire Hathaway invests $10 billion in Alphabet's AI infrastructure buildout. Alphabet raises $80 billion to scale AI capacity, with capital expenditures expected to reach $190 billion in 2026.