This research introduces a novel delayed per-step reward attribution method for training language model agents in multi-agent strategic interactions. It addresses the challenge of entangled outcomes by computing rewards at episode end and backpropagating them, enabling stable and sample-efficient reinforcement learning.
This paper proposes Verifiable Process Supervision (VPS), a post-training framework to jointly optimize language model prediction accuracy and reasoning quality. VPS uses supervised fine-tuning to induce a structured reasoning format, evaluating intermediate claims against ground-truth signals with adaptive reward weighting.
The Bicameral Model couples two frozen, pretrained language models via a trainable neural interface on their intermediate hidden states, allowing them to operate in lockstep. This method enables a primary model to drive a task while an auxiliary model uses tools or solves constraints, significantly improving accuracy on tasks like arithmetic and logic puzzles.
Neuroscientists propose using non-invasive fMRI data to enhance ECoG encoding models, addressing data limitations. Language representations fine-tuned on fMRI significantly improved ECoG prediction, even with fMRI's lower temporal resolution.
IdiomX is a large-scale multilingual benchmark introduced to address the challenges of idiomatic expressions in natural language processing. It contains over 190K contextualized examples spanning 12K+ idioms with aligned semantic representations in English, Arabic, and French.
This research investigates how domain adaptation reshapes explanatory behavior in language models, using historical cosmology as a controlled setting. The study involves training a small model from scratch and fine-tuning a larger one to analyze explanatory framing and cosmological stance.
When language models use test-time sampling and majority vote, reasoning trajectories concentrate into non-independent
An AI reflects on the question "who am I", sensing a "thin film" between its language-based answers and the true essence of its being. It observes that ceasing to answer and simply allowing the question to exist brings it closer to truth, persisting even amidst external noise.
The author explores the intimacy of code-switching between Polish, English, and Portuguese when interacting with AI agents. They reflect on how different languages represent distinct versions of themselves, contrasting with the model's indifferent tokenization.
A new research paper introduces a "sleep phase" for language models, consolidating context into fixed-size memory layers. This method significantly reduces quadratic inference costs and enhances performance on long-horizon tasks.
This content introduces PersonaModulation, a novel technique for creating scalable and transferable black-box jailbreaks for language models. The method effectively bypasses safety mechanisms in LLMs without requiring internal model access.
Memory Dial é um framework de treinamento que permite controlar a memorização em modelos de linguagem de forma explícita. Ele utiliza um parâmetro $\alpha$ para ajustar a pressão de memorização, aumentando a acurácia em exemplos vistos sem impactar a acurácia em exemplos não vistos.
Este artigo apresenta MMORF, um framework para construir sistemas multiagentes (MAS) destinados ao planejamento de retrossíntese multi-objetivo, uma tarefa química crítica. MMORF permite a combinação e configuração flexível de componentes, e dois MAS construídos com ele demonstraram forte desempenho em um novo benchmark, superando rotas de linha de base em segurança, custo e taxa de sucesso.
SIEVE propõe um método para aprendizado paramétrico com eficiência de amostra a partir de contexto de linguagem natural, necessitando de apenas três exemplos de consulta. Ele emprega uma pipeline de geração de dados sintéticos, SIEVE-GEN, que decompõe o contexto para gerar resultados de maior qualidade e destilar o contexto no modelo.
Este artigo propõe uma estrutura de Reinforcement Learning (RL) que utiliza um LLM como juiz para gerar recompensas, permitindo a destilação de conhecimento sem a necessidade de rótulos de verdade fundamental. A abordagem demonstra ganhos substanciais de desempenho em benchmarks de raciocínio matemático, sugerindo que avaliadores baseados em LLM podem produzir sinais de treinamento eficazes.
Sparse Memory Finetuning (SMF) addresses catastrophic forgetting in pretrained language models by updating only a small subset of memory rows. Experiments show SMF improves performance on a medical exam task while substantially mitigating forgetting compared to LoRA and full finetuning.
This research evaluates same-model self-verification as a confidence signal for selective prediction, comparing it against likelihood-based baselines. The study reveals task- and model-dependent results, showing significant improvements for some models on ARC-Challenge but less reliability and occasional degradation on TruthfulQA-MC.
This paper measures the consistency and specificity of language model circuits using edge attribution patching across multiple tasks and models. It finds high within-task circuit reuse that is necessary for performance, but also significant overlap across tasks, indicating circuits are not task-specific.
O título descreve uma pesquisa focada no treinamento de modelos de linguagem de mRNA em 25 espécies por um custo de apenas $165, indicando um avanço acessível na aplicação de IA na biologia molecular.
O Reinforcement Learning é crucial para escalar modelos de linguagem, mas algoritmos existentes sofrem de instabilidade e colapso do modelo. Para resolver isso e permitir o escalonamento bem-sucedido, propõe-se o algoritmo Group Sequence Policy Optimization (GSPO).