Researchers have identified that LLMs suffer from a bias toward providing immediate, often incorrect answers in ambiguous situations instead of asking clarifying questions. They introduced BACE (Belief-Aware Critic with Epistemic tracking), a lightweight method that runs on top of a frozen language model to strategically decide when to request clarification.
A new method called FlagRot optimizes key-value (KV) cache quantization for long-context models. By selectively rotating key segments based on attention error rather than applying global rotation, it reduces accuracy loss in 2-bit or 4-bit caches.
Researchers introduced SLATE, a benchmark for evaluating long-horizon, large-scale API selection in enterprise software environments. Unlike current benchmarks, it forces models to navigate roughly 1,000 tools to complete complex tasks.
New research shows that reasoning history can be rewritten into compact state records before new tool results are introduced. This method effectively reduces future reasoning tokens and total agent spending by eliminating redundant thoughts without requiring model retraining.
Researchers have developed a method to calculate the theoretical upper bound for Multi-Draft Speculative Decoding (MDSD) efficiency by solving the dual of the optimal transport problem. The study reveals that current verification algorithms often fall short of this theoretical limit, and that sampling without replacement significantly improves acceptance rates compared to sampling with replacement.
Research shows that identical persona instructions produce different behavioral outcomes in LLMs when switching from English to Korean. While buyers followed instructions similarly, sellers became significantly more accommodating and polite in Korean, leading to higher agreement rates and lower prices.
Researchers have introduced Circuit Oracle, a multi-agent LLM system designed to interpret language model computations by analyzing attribution graphs. The tool uses tool calling to answer natural-language queries about target model behavior and safety.
Apple researchers have unveiled an online training infrastructure for mobile UI agents that utilizes a two-part architecture to split navigational logic from visual interpretation. By employing step-level feedback provided by an LLM-as-a-Judge and a novel value estimation technique known as GiGPPO, these agents demonstrate significantly better adaptability than static models trained on pre-recorded sequences.
Recent research shows that many attention heads in large language models are redundant and specifically contribute to verbatim memorisation of training data. By selectively pruning these heads at inference time, researchers achieved a 91.5% reduction in memorisation events while simultaneously improving performance on standard benchmarks.
Recent research reveals that language models fail to adapt to evolving user preferences, instead defaulting to 'exploit-first' behavior. While models can identify the correct strategy in multiple-choice scenarios, they show a 35-46% drop in calibration during open-ended generation, persistently over-relying on stale historical data.
Researchers have proposed LOTUS, a framework for hallucination detection that uses multilinear low-rank decomposition to extract signals from hidden LLM layers. By analyzing peak activations and cross-layer patterns, the method identifies incorrect responses more effectively than existing pooling-based techniques.
Researchers have developed TraceQuery, a sparse decoding framework that reduces the computational cost of diffusion large language models (dLLMs). By analyzing denoising trajectories, the method selectively recomputes only necessary token representations rather than performing redundant operations, maintaining high accuracy with fewer computations.
Researchers introduced Recursive Semantic Divergence (RSD), an unsupervised metric that measures and penalizes inconsistency in LLM agent trajectories. By applying RSD as a reward signal during fine-tuning, models exhibit significantly more stable behavior and higher success rates across tool-use benchmarks.
Recent research on Qwen3.5-27B demonstrates that while activation interventions can force a model to produce a specific target output, they often fail to instill correct underlying reasoning. Researchers found that gains in target selection collapse when facts are recombined, suggesting that forced outputs do not reliably represent valid tool-use condition logic.
IBM Research released vLLM-Hook MIA, a plugin for the open-source vLLM inference engine that lets developers observe and intervene in a model's internal states at runtime — for example to decline malicious prompts, flag hallucinated facts, or adjust tone — without disabling vLLM's fast graph-replay execution mode. According to IBM's technical report, MIA is up to 20x faster than the earlier vLLM-Hook and up to 14x faster than methods that insert code directly into the model. In tests, steering an 8B Llama or 1.5B Qwen model added only about 2% latency, and a 70B Llama model added about 6%.
Allen Institute for AI (Ai2) released FairStress, a benchmark that tests demographic fairness in LLMs not just by their final decisions but by how easily those decisions change under persuasive pressure. Testing open and frontier models including Olmo, Llama, Qwen, Gemma, Claude, and GPT found that models often abandon correct, well-supported decisions more readily when doing so favors a minority-coded candidate — a pattern the authors call 'anti-stereotypical over-correction.' The benchmark covers 300 scenarios across 29 domains (hiring, lending, healthcare, education, housing, criminal justice), expanded via demographic and persuasion variations into over 13 million test cases, plus a smaller 50K-item 'FairStress-Core' subset for faster evaluation.
Google introduced TorchTPU, a native PyTorch backend for Cloud TPUs (torch.device("tpu")), and co-developed a fork of Meta's TorchTitan trainer adapted for TPU training. Previously, running PyTorch training on TPUs meant rewriting code in JAX or relying on the older torch_xla path; now only minimal device-initialization changes are needed. The fork provides tested recipes for pre-training large dense models (Llama 3 70B/405B, Qwen 3 8B/32B), Mixture-of-Experts models (Qwen 3 235B), and diffusion transformers (FLUX.1 12B), plus a 'Graph Trainer' that captures an entire training step as one XLA program for better compiler optimization.
A new study questions the common belief that whether an AI agent falls for a prompt injection is settled the moment it reads the malicious text. The researchers built a branching-rollout technique on AgentDojo benchmarks: they replayed an agent's run up to the point it encounters an injected instruction, then let the rest of the episode play out many different ways to measure how much of the final outcome was already locked in versus still open. For a Qwen3 agent running without extended reasoning, 91% of the outcome was already determined at that moment; switching on reasoning ('thinking') cut that figure to 62%, showing the agent's compliance decision often keeps forming well after it first sees the attack. Enabling reasoning also cut Qwen3's susceptibility to successful attacks, but made two other tested models more vulnerable, and benchmark scores for reasoning agents flipped on a rerun as often as 35% of the time.
AWS released a push-based NIXL connector for vLLM that changes how key-value (KV) cache data moves between prefill and decode instances in disaggregated LLM serving. Instead of the decode instance pulling KV blocks after prefill finishes, the decode instance registers its target memory in advance and the prefill instance writes the KV data directly once ready, moving coordination work off the critical path. Benchmarked on a 1-prefill/1-decode Qwen3-32B deployment using EC2 P5en instances with EFA, push mode delivered up to 3x better time-to-first-token (TTFT) and up to 30% better time-per-output-token (TPOT) than the default pull mode, especially under high request rates.
Google Research describes new work on grounding AI agents in physical lab reality, using a digital-twin framework called VERID to let agents safely plan and execute liquid-handling robot protocols. Combined with closed-loop molecular design cycles for aptamers and proteins, the agents iteratively learned from wet-lab results, improving hit rates over multiple rounds and discovering single-digit nanomolar protein binders for a cancer-related target.
NVIDIA, working with IIT Madras's AI4Bharat lab, is releasing Nemotron-pretraining-Indic-v1, a roughly 2.8 trillion token synthetic pretraining corpus covering 22 Indian languages, along with the open GPU-accelerated pipeline used to generate it. The corpus includes native-script data for all 22 languages and romanized text for twelve, built by translating curated English source material (web text, math, code, Q&A) using NeMo Data Designer and quality-checked with an MQM-inspired rubric. NVIDIA already used a 10-billion-token Indic subset for continued pretraining of Nemotron 3.5 Super VL, which showed measurable gains on Indic benchmarks.
Apple's MLX team published benchmarks showing large open-weight models — including a near-trillion-parameter MoE model called Muse Spark Open — running on Mac Studio hardware, both on a single machine and distributed across up to four Mac Studios using a new RDMA-over-Thunderbolt driver. The new M5 Ultra chip, with dedicated GPU Neural Accelerators and 1.2 TB/s memory bandwidth, delivers up to 4.6x faster time-to-first-token and 1.6x faster generation speed versus the previous M3 Ultra. For models too large to fit on one machine, MLX now supports pipeline and tensor parallelization across multiple Macs connected via Thunderbolt RDMA, letting a four-Mac-Studio cluster generate over 36 tokens per second on an 816B-parameter model.
AWS detailed three kernel-level optimizations for Mixture-of-Experts (MoE) model serving on Hopper GPUs, implemented in collaboration with the vLLM community and shipped in Amazon SageMaker AI's LMI container. The changes target the inefficiencies of fused MoE computation during low-batch decode, which previously left Tensor Cores underutilized. Combined, the optimizations deliver up to 10% higher throughput and up to 9% faster time-to-first-token (TTFT).
IBM Research used a coordinated multi-agent 'agent factory' to explore hardware design optimizations for a quantum error-correction decoder, comparing proprietary and open frontier models under the same 100-hypothesis search budget. The open-model configuration, run on-premises via IBM's llm-d distributed inference framework, achieved a 52x speedup over baseline for about $26 in inference cost, versus $299 for a proprietary-model configuration that reached a higher 204x speedup. The study highlights that cache-aware routing and efficient serving (98% prefix-cache hit rate, ~$0.11 per million tokens) made the open-model approach dramatically cheaper while still finding valid, compact hardware designs.