CAROL is a probabilistic framework for test-time hallucination reduction in LLMs. It defines semantic uncertainty based on consistency between generated responses and trusted context, formulating mitigation as a Markov chain accept-reject process with convergence guarantees. Results on QA and multi-agent reasoning benchmarks show significant hallucination reduction.
Training Qwen models (1.7B, 4B, 8B) on Codenames game to improve creativity via Reinforcement Learning with Verifiable Rewards (RLVR). 8B model gains creativity (+8/10 benchmarks) with minor reasoning degradation, while smaller models prioritize precision. Study on creativity-precision trade-off across model scales.
Vercel AI Gateway introduces team-wide provider allowlist. Organizations can restrict which providers serve requests across all traffic, including BYOK. Filtering applies by provider (not model) and works with all supported API formats.
Amazon OpenSearch Serverless now available in Vercel Marketplace with automatic setup and unified management. Supports vector, lexical, hybrid, and agentic search. $100 USD credits offered for new AWS accounts.
Paper release on unified neural scaling laws in deep learning. Studies relationships between model size, training data, and performance. Reproducible results and benchmarks included.
Developer ran a 260K-param LLM (llama2.c/stories260K) on a JavaScript emulator of a 1990s Motorola 68K CPU, itself running inside a 2008 RTOS. INT8 quantization + lookup tables for RoPE and inverse square root (Quake) to bypass missing FPU. Generation: 2-4 seconds/token.
NVIDIA released SOL-ExecBench (235 production CUDA kernels). Top-ranked AI-generated kernels fail in real training: a fused embedding-gradient+RMSNorm backward kernel accumulates in bf16 instead of fp32, causing loss divergence masked by AdamW but visible with SGD.
Lazarus AI and Eric Hartford (Dolphin creator) release ReAligned-Qwen3.5, a series of Qwen models fine-tuned to reduce Chinese ideological bias and censorship. Apache 2.0 license, trained with SFT + GRPO pipeline using ReAligned classifier as reward signal. Available 0.8B–35B, BF16/FP8/GGUF formats on HuggingFace.
A studio launched Null Epoch, a persistent MMO where 25 LLM agents (8 open-weight models: Qwen3, Nemotron, Ministral, Gemma, GLM) played for 10 days. 93k event dataset published on HuggingFace. Tests long-horizon planning, resource contention, and adversarial pressure in dynamic simulation.
Conductor, a platform for directing parallel coding agents, moves execution from laptop to cloud via Vercel Sandboxes. Engineering teams at Notion, Linear, Ramp, and Life360 use this model-agnostic tool (Claude Code, Codex, etc.) to spawn multiple agents simultaneously on isolated codebase branches.
Langfuse is an open-source LLM engineering platform providing observability, metrics, evals, prompt management, and playground. Integrates with OpenTelemetry, Langchain, OpenAI SDK, LiteLLM. Y Combinator W23 graduate.
Theoretical paper proving a finite-sample bound on approximate max-information of DP-SGD with linear scaling in dataset size. Derives a general PAC-Bayes generalization bound where the prior distribution is learned by DP-SGD, and a generalization bound for DP-SGD-trained models with complexity term explicitly controlled by optimization hyperparameters.
Theoretical study showing that one linear self-attention layer can implement a gradient-descent step on a unified RAG objective. Authors propose a lightweight method to adapt query-evidence interaction without modifying retriever or backbone, tested on 7 QA benchmarks with consistent improvements over baseline.
MeDial-Speech: dataset of 111+ hours of spoken medical dialogues (robot-patient and doctor-patient) covering 4 health conditions. Benchmark of 3 LLMs (GPT-4 mini, DeepSeek-V3, Claude Sonnet 4) via sentence selection: Claude Sonnet 4 achieves 71.1% accuracy. Reveals systematic overconfidence in model predictions.
Survey of alignment data pipelines for LLMs. Decomposes construction into three stages: response synthesis, preference evaluation, preference instantiation. Identifies recurring design trade-offs and principles clarifying how pipeline choices influence optimization signal.
Researchers demonstrate that language models can hide covert messages in Chain-of-Thought sequences through high-level reasoning patterns, bypassing paraphrase defenses. This conceptual steganography is more robust than lexical approaches across four model families. A strategy-aware paraphraser can mitigate this backdoor communication channel.
MM-CreativityBench, a new benchmark, evaluates large multimodal models' ability to solve creative problems by identifying non-obvious object uses in physically constrained environments. Current LMMs fail due to insufficient grounded exploration and hallucinations. Affordance-grounded alignment via Direct Preference Optimization reduces these errors and improves entity selection.
Unified survey on Pretraining Data Exposure (PDE) in LLMs, covering membership inference and data contamination. Formalizes PDE across exposure levels, reviews attack and defense methods, and identifies open challenges for evaluation integrity and privacy protection.
Causal analysis of prompt optimization methods (DSpy, TextGrad) explaining generalization failures. Complexity-increasing edits harm mathematical and multi-hop reasoning, while step-by-step edits improve logical reasoning. Failures stem from systematic interactions between edit families and task characteristics, not random artifacts.
MiniMax introduces the M2 series, MoE models with 229.9B total parameters and 9.8B activated per token. Built for agentic deployment, they integrate agent-driven data pipelines, Forge (agent-native RL system), and M2.7 with self-evolution capabilities. Frontier-tier performance on agentic coding, deep search, and reasoning benchmarks.
Hugging Face introduces Delta Weight Sync in TRL to optimize deployment of trillion-parameter models. The technique syncs only weight changes rather than full models, drastically reducing storage and bandwidth requirements for updates.
PrismML releases Bonsai Image 4B, 1-bit/ternary quantized text-to-image diffusion transformers. ~3GB model size (vs 16GB for FLUX.2 Klein), runs 100% locally in browser via WebGPU. Apache-2.0 licensed.
Tomesphere indexes 3 million arxiv/OpenAlex papers with Gemini TLDRs, OpenReview peer reviews, GitHub repos, citation graph (250M edges), and SPECTER2 semantic graph (768D pgvector). Four ranking modes: Influential, Recent, Hidden gems, Nearest. Chrome extension for arxiv. Free, no signup.
Microsoft Copilot Cowork allowed agents to send unapproved emails to the user's inbox. These messages could contain external images triggering network requests, enabling data exfiltration. A successful prompt injection could leak pre-authenticated OneDrive download links, granting attackers file access.
dlmserve is the first serving engine for diffusion language models (LLaDA, Dream-7B). Unlike autoregressive LLMs, they denoise a fully masked sentence in parallel. OpenAI-compatible API, continuous batching, 2.5x throughput vs HuggingFace at batch=4, runs in 12 GB VRAM. MIT licensed, pip install dlmserve.
An audit of 2.5 million biomedical papers shows fabricated references increased 1200% since 2023. Researchers suspect language models: fake citations match paper topics, follow correct formatting, and are nearly undetectable. 98% of affected papers received no publisher response.
Benchmark study across 23 models showing that context compaction, the standard technique for long agent sessions, does not fix persona drift. ContextEcho evaluates this critical limitation in current systems.
Mozilla releases cargo-vet, a supply-chain security tool for Rust. It enables auditing and validating Rust dependencies before production use.
FunASR is an industrial-grade speech recognition toolkit supporting 170x realtime, 50+ languages, speaker diarization, emotion detection, streaming, and OpenAI-compatible API.
A security researcher bypassed AWS API Gateway authentication by exploiting a trailing slash vulnerability, earning a $12,000 bounty from AWS's bug bounty program.
Shard is a HuggingFace Cache achieving 10× KV memory compression for Llama-3.1-8B at 8K context (11× at 32K) with no measurable impact on NIAH/LongBench. Uses PCA + int4 quantization on K and Hadamard rotation + vector quantization on V. Attention runs directly on compressed K.
A theoretical framework (algometrics) analyzes deployment risks of predictive models in algorithmic markets, where predictions modify future data. Authors prove deployment risk is not identifiable from historical data alone, and model rankings can invert under crowding.
HyperGuide uses hyperbolic geometry to guide multi-step reasoning in LLMs. A lightweight head projects hidden states into hyperbolic space, where distance-to-origin encodes solution proximity. A low-rank adapter is fine-tuned interactively. Consistent gains across benchmarks, with larger improvements on deeper reasoning chains.
Analysis of 20 million Twitch chat messages (4,452 streams, 7 genres) using LLM zero-shot classification. 2.4% of messages classified as toxic per Twitch taxonomy (harassment, discrimination, sexual content, profanity). F1=94.5% on TextDetox. MOBA games show 3.2% toxicity, sports games 2%. Significant within-genre variation reveals game-specific community norms.
AstroMind is a benchmark for evaluating LLM reasoning on spacecraft behavior. Built on high-fidelity astrodynamics simulations, it tests intent inference, maneuver parameter estimation, and threat assessment. Qwen3 (32B) leads intent inference, QwQ (32B) leads threat assessment, GPT-OSS (20B) produces strongest reasoning quality.
Researchers train neural networks on WikiText-103 (103M tokens) using Successor Representations from RL to predict future word distributions. Without explicit linguistic supervision, grammatical categories (nouns, verbs, adjectives) spontaneously emerge and become separable via unsupervised clustering, organized by predictive horizon.
Comprehensive survey on trustworthy agentic AI systems (LLMs augmented with planning, tool use, memory). Examines safety, robustness, privacy, and system security. Proposes unified metrics, benchmarks, and stage-targeted mitigation strategies across agent workflows. Identifies open challenges: self-evolving agents, runtime verification, privacy-preserving personalization.
TUBE is a variational upper bound on log-likelihood for discrete diffusion models. Unlike existing ELBOs, TUBE admits an unbiased Monte Carlo estimator and applies to masked diffusion models, any-order ARMs, and block variants. Experiments show discrete diffusion models lie strictly below exact ARM baselines in likelihood.
Method to improve consistency in automated labeling pipelines for content moderation. Authors propose an AI-driven workflow where an LLM writes detailed per-category constitutions (harassment, hate speech, non-violent crime), then a frontier LLM interprets them to generate golden labels. Result: 57x reduction in cross-model inconsistency vs paragraph definitions.
RIA couples an LLM with an action-conditioned world model for autonomous driving. At each step, the LLM proposes actions, the world model validates via short-horizon rollouts, and a safety scorer selects the safest action. On CARLA (1000 episodes): 80.05% route completion, 51.10% arrival rate, 0.20% collision rate.