RELIC is a framework that enables agents to learn and transfer composable skills in multi-agent planning scenarios without sharing internal code. Instead, agents abstract successful behaviors into general principles, which can be instantiated by others with different interfaces. Skill refinement occurs via private LLM-guided search, and a trusted orchestrator evaluates updates based on overall team performance. This approach allows for coordination and skill transfer while preserving privacy and accommodating heterogeneous agent designs.
Why it matters: RELIC introduces a new paradigm for privacy-preserving coordination in multi-agent systems, addressing a key challenge in scenarios where agents cannot share internal implementations.
Researchers introduce Logical Graph Uncertainty (LGU), a framework that models logical relationships among large language model (LLM) outputs to enhance uncertainty estimation. LGU aggregates probability along entailment chains and penalizes mutual incompatibility, leading to improved performance over semantic entropy baselines by up to +7.1% AUROC and +3.5% AUARC across multiple question-answering benchmarks.
Why it matters: Improved uncertainty quantification is critical for the safe deployment of LLMs, especially in scenarios where outputs are logically compatible but semantically diverse.
A new preprint proposes diversity-oriented fine-tuning strategies—using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO)—to improve semantic-entropy-based hallucination detection in large language models. By encouraging more varied generations, these methods reduce cases where models repeatedly produce identical incorrect answers, which often evade detection. Experiments show that the approach improves detection effectiveness, matching or surpassing current state-of-the-art methods.
Why it matters: This work introduces a practical fine-tuning approach that directly enhances the reliability of hallucination detection in large language models, addressing a key challenge in their deployment.
A new framework, MCF-MOE, is introduced to enhance expert selection in Mixture-of-Experts (MoE) models by fusing cross-layer semantic information and local token-level interactions. The approach addresses context incompleteness in existing MoE routers, leading to more consistent and context-aware expert routing. Experimental results on language modeling and understanding tasks show that MCF-MOE achieves improved routing consistency and better downstream performance compared to strong MoE baselines.
Why it matters: This work offers a practical advance in MoE model design, potentially enabling more efficient and stable large-scale Transformer systems by improving expert routing consistency.
Researchers introduce MSCE, a training-free framework that organizes agent experience into reusable procedural skills and declarative environmental knowledge. MSCE employs reflection-weighted value backfilling to guide the evolution of memory and skills, enabling agents to autonomously crystallize and refine capabilities from experience. Experimental results on EvoAgentBench and LoCoMo show that MSCE significantly outperforms state-of-the-art skill-augmented and memory-driven agent baselines, with strong cross-domain transfer and lifelong learning abilities.
Why it matters: This work advances the autonomous evolution of LLM agent capabilities without retraining, supporting more robust lifelong learning and cross-domain adaptability.
A new preprint investigates how language models and humans process next-word prediction by analyzing EEG recordings during reading. The study finds that only surprisal, not top-1 prediction accuracy, correlates with brain signals associated with language processing. Additionally, increasing model size does not necessarily improve alignment with human cognitive patterns.
Why it matters: The findings suggest that simply scaling up language models does not guarantee more human-like cognitive processing, highlighting the need for new approaches to achieve cognitive plausibility.
TopoTuner is a topology-guided fine-tuning framework for large language models that selectively freezes attention projection matrices based on topological changes measured by Wasserstein distances between persistence diagrams. The method achieves competitive performance with full fine-tuning while updating only 1-2% of parameters, and outperforms LoRA in 7 out of 9 model-dataset settings. TopoTuner also reduces training time by 20.4% compared to full fine-tuning and 5.5% compared to LoRA, and enables the transfer of reusable freezing profiles across tasks.
Why it matters: TopoTuner offers a novel approach to efficient LLM adaptation by enabling reusable fine-tuning profiles, potentially reducing computational costs and improving transferability across tasks.
AgentBrew introduces a method for transferring knowledge from a strong teacher LLM to a weaker student agent by distilling interactive experiences into an external memory, without requiring weight updates or teacher access at test time. The approach uses failure-triggered teacher interventions and student-aware synthesis to generate actionable, environment-validated guidance. Evaluations across coding, math, and tool-use tasks indicate that this method produces capable and deployable LLM agents.
Why it matters: This approach could make it easier and more cost-effective to deploy capable LLM agents by eliminating the need for continual model retraining or ongoing teacher involvement.
Researchers introduce RIMS, a three-stage preference optimization framework designed to enhance small language models in retrieval-augmented generation tasks. RIMS generates synthetic chain-of-thought preference data, employs a differentiable soft aggregation mechanism to better utilize preference signals, and applies preference optimization to improve robustness against noisy evidence. Experiments on four multi-hop question answering benchmarks demonstrate that RIMS consistently outperforms state-of-the-art baselines in both Exact Match and F1 scores under noisy retrieval conditions.
Why it matters: This work advances the reliability and effectiveness of small language models for retrieval-augmented generation, particularly in resource-constrained environments.
A new preprint investigates whether the neurons responsible for arithmetic in large language models (LLMs) are form-invariant across symbolic arithmetic, natural language word problems, and Python code. Using attribution and activation patching on Llama-3 models, the researchers identified a compact set of neurons shared across all three formats, which are both necessary and sufficient for late-layer arithmetic computation. Transferring activations from successful to failed executions recovered over 97% of incorrect predictions for addition and subtraction, suggesting that cross-format failures are due to activation states rather than distinct neural circuits.
Why it matters: This work provides mechanistic evidence that LLMs use a shared neural circuit for arithmetic across different input formats, informing future efforts to improve model robustness and interpretability.
Open Source→Official→arXiv Computation and Language
OpenLanguageModel (OLM) is an open-source PyTorch library designed for building and pretraining small language models with readable and modular code. It offers 27 presets across nine model families, supports CPU, single-GPU, and multi-GPU execution, and demonstrates 90.6% four-GPU weak-scaling efficiency on a 348M-parameter workload. OLM is MIT-licensed and available via PyPI, GitHub, and its documentation site.
Why it matters: OLM lowers the barrier for education and research in language model pretraining by providing transparent, modular code that is accessible for both teaching and experimentation.
LaCache is a training-free acceleration framework for diffusion-based large language models (DLLMs) that reduces redundant computation by caching intermediate results, including embeddings, pre-attention states, and softmax statistics. It also introduces per-group FP8 quantization for feed-forward network layers to further improve efficiency. Experiments show LaCache achieves about 1.3x speedup on its own and up to 40.2x when combined with other acceleration methods, without compromising task accuracy.
Why it matters: Improving inference efficiency in diffusion LLMs can make these models more practical for real-world deployment by reducing computational costs without sacrificing output quality.
A new preprint evaluates nine open-weight language models ranging from 135M to 3B parameters on a structured, multiple-choice benchmark tailored for local deployment. The study finds that Qwen Coder 3B achieves 75.67% strict accuracy, and parameter-efficient fine-tuning can boost performance by up to 26.85 points. The results suggest that, with careful benchmarking and adaptation, sub-3B models can serve as effective local experts for structured, niche tasks.
Why it matters: This work challenges the prevailing view that only large-scale models are practical for real-world AI applications, showing that smaller models can be viable for specialized local use cases.
Researchers introduce a new framework that formalizes task alignment as a partially observable Markov decision process (POMDP), enabling models to infer latent user intent from ambiguous interactions. Experiments show that current language models recover the intended task only 22-32% of the time under ambiguity, compared to 48% for humans. While post-training methods improve model performance, they still fall short of human abilities in resolving ambiguous requests.
Why it matters: This work exposes a significant shortfall in current language models' ability to handle ambiguous, real-world user tasks, highlighting a key challenge for developing more reliable AI assistants.
Researchers present agrepl, a command-line framework designed for deterministic replay of AI agent executions. By intercepting all external interactions through a man-in-the-middle proxy and replaying them in a fully isolated environment, agrepl achieves perfect replay fidelity and a median per-step latency reduction of 98.3%. The tool is open-sourced under the MIT license and aims to address reproducibility challenges in LLM-based agent systems.
Why it matters: This work provides a practical solution to the reproducibility and debugging challenges in non-deterministic AI agent systems, facilitating more reliable development and testing.
A new study introduces masked diffusion language models (MDLMs) as world models for reinforcement learning (RL), addressing limitations of autoregressive models such as left-to-right bias. The researchers curated over 239,000 trajectories across nine environments and demonstrated that MDLMs provide greater coherence, groundedness, and rollout diversity than much larger autoregressive language models. Their plug-and-play GRPO training framework led to up to 47% absolute improvement in zero-shot transfer to out-of-distribution environments, without environment-specific fine-tuning.
Why it matters: This work advances scalable and steerable world modeling for RL agents, potentially reducing reliance on hand-crafted training environments and improving generalization.
Researchers have identified a significant blind spot in neural fact-checkers: they often fail to recognize when quantities are canonically equivalent (such as 95°C and 368.15 K), leading to potential misinterpretation of scientific claims. The team introduces Symbolic Augmentation, a training-time framework that generates label-preserving augmented data using symbolic reasoning, which raises robustness to canonical-equivalent rewrites from 36.5% to 98.2% and slightly improves in-distribution accuracy. The approach also transfers well to external benchmarks and demonstrates that training-time augmentation is a more effective integration point for symbolic and neural methods than feature-level fusion.
Why it matters: This work addresses a critical failure mode in AI fact-checking that could silently invert scientific claims, and demonstrates a principled integration of symbolic and neural methods.
SelKV is a training-free framework for compressing key-value (KV) caches in large language models (LLMs). It uses a soft cosine gate to adaptively merge or drop tokens based on value-vector similarity, and introduces an attention-ratio compensation mechanism to correct softmax imbalances caused by merging. On LongBench, SelKV retains only 25% of the KV cache while achieving strong performance, near-lossless quality on grouped-query attention (GQA) models, and a 3.3x decoding speedup at 100k tokens.
Why it matters: SelKV addresses the memory and speed bottlenecks of KV caches in LLMs, enabling longer context lengths and faster decoding without significant quality loss.
A new preprint identifies a structured confound in Reinforcement Learning from Human Feedback (RLHF): the emotional or psychological state of human raters during annotation can systematically bias preference labels, especially under stressful conditions. These state-dependent shifts are distinct from random noise and can propagate through reward modeling and policy optimization. The authors propose an audit framework with falsifiable predictions and outline a pilot study plan to detect such bias in publicly available instruction-tuned models.
Why it matters: This work draws attention to a novel, testable source of bias in RLHF that could undermine the reliability of preference-based AI alignment methods.
Researchers introduce Exact Network Surgery, a formal method for inserting residual blocks into live computational graphs while preserving the network's function exactly and ensuring that new parameters are immediately trainable. The method is validated on the NeuroDSL platform, demonstrating bit-exact function preservation, predictable gradient behavior, and constant bookkeeping cost during model expansion.
Why it matters: This work enables neural networks to be expanded dynamically without retraining or loss of function, potentially reducing computational overhead and increasing architectural flexibility.