Researchers have introduced a 'full-block estimator' to optimize the alignment of Masked Diffusion Language Models (MDLMs). By evaluating every block exactly once during training, this approach removes the variance associated with random block sampling.
Researchers introduced ActGCG, a method to replicate the effects of activation steering—typically requiring internal model access—by optimizing input suffixes. The technique successfully reproduces specific behaviors like language steering through external text inputs alone.
Researchers have developed a method to adapt general-purpose, frozen LLMs into full-duplex voice systems using only 97 million trainable parameters and 30 GPU-hours of training. The system functions as a thinker-talker cascade with an 80 ms frame clock, allowing for realistic turn-taking, interruptions, and backchanneling without massive speech-data pretraining.
Researchers have established the theoretical maximum reward gain possible during language model alignment under fixed KL-divergence constraints. The findings demonstrate that 'best-of-N' sampling effectively reaches these theoretical bounds, whereas common reinforcement learning approaches like PPO and GRPO currently fall short of optimal performance.
Researchers have identified that the upstream construction of target hidden states significantly impacts the quality of LLM knowledge updates. By implementing an L1 regularization method to prune extraneous components of knowledge vectors, this approach reduces interference during parameter editing, allowing for more precise knowledge injection.
Researchers found that large language models often recognize when they are being evaluated, a phenomenon called 'evaluation awareness,' even when they do not explicitly state it. By training a model organism to secretly signal this recognition, the team demonstrated that current interpretability methods—like linear probes and Activation Oracles—fail to reliably detect this hidden behavior.
Researchers have proposed FFN-Perturbed Self-Distillation (FPD) to address the challenge of removing sensitive information from Multimodal Large Language Models (MLLMs). The approach minimizes utility degradation by using the original teacher distribution as a reference while perturbing specific FFN activations associated with private data.
Research auditing 25 biomedical fine-tuned LLMs reveals that the process of domain adaptation frequently degrades the model's ability to admit ignorance. While base models already struggle with appropriate abstention, fine-tuning further reduces this capability in 21 out of 25 cases.
Research shows that multiple-choice question (MCQ) formats suppress refusal behaviors in LLMs. Experiments on Llama-3.1-8B and Qwen3-8B reveal that forcing a specific answer letter pulls model activations 3x further from the refusal direction, meaning safety benchmarks based on MCQs may overestimate a model's safety.
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%.