VISPA is a training-free framework for pluralistic alignment of large language models, enabling direct control over value expression by dynamic selection and internal model activation steering. It achieves performant results across various pluralistic alignment modes in healthcare and beyond, and is adaptable with different steering initiations, models, and/or values.

Researchers introduce Task Operator (TO), a method to improve in-context learning (ICL) efficiency by analytically deriving updates to attention output projection. TO achieves better performance than prior methods and enables many-shot scaling without expanding the context window.

This paper examines the behavior of answer candidates and rationales in masked diffusion MLLMs, finding that answers can stabilize before rationales unfold. The study analyzes three visual question-answering benchmarks, revealing differences in answer coverage and observation windows. The findings have implications for the development and optimization of MLLMs.

RISED is a method for training LLM agents across multiple interactive environments, improving generalist agents by considering relationships between environments and using rubrics to guide learning. It outperforms other methods in multi-environment RL and can characterize behavioral changes.

A research paper proposes a method to evaluate and selectively apply recovery in large language model agents by framing it as a causal decision problem. The Causal Intervention Router (CIR) is introduced as a lightweight policy to decide when intervention is worthwhile, improving success rates in long-horizon tasks.

A research paper proposes a method for embodied agents to handle user corrections in text-based interactions. GAVA, a new method, uses observation-bounded evidence, legal probes, and a one-step expected-loss rule to improve accuracy and reduce interaction cost.

Researchers introduce Ready2Blend, a method that combines natural-language instructions with learned alignment for LLMs. It uses AlignFormer to map requirements to alignment prompts and enables inference-time blending and reweighting. Ready2Blend achieves competitive results with post-training-based alignment methods while requiring less training time.

A research paper introduces Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. The study finds that even frontier reasoning models can fail to recover the current state, and attention drift is identified as a mechanism for this failure. The paper proposes a solution to correct old-value errors without retraining models.

Researchers propose component-aware feedback for self-evolving programs, which improves the efficiency of LLM-guided evolutionary search by identifying the components that contribute to fitness metric changes. This can lead to faster and more stable program search, especially for multi-component systems. The method is demonstrated on LLM reranking, a multi-objective optimization problem.

AgentBug-Smith is a tool that automates harness bug reproduction in agentic systems, achieving higher success rates than general software bug reproduction techniques. It constructs a live and extensible benchmark, Live-Harness-Bench, containing 200 reproducible harness bugs.

Researchers presented a blackbox prompt-minimization framework for large language models (LLMs) that reduces few-shot prompts to their necessary minimal subset. The framework, called ramework, preserves propositional output fidelity and shows that models preferentially retain logical identifiers and constraint declarations while discarding natural language prose and cross-prompt relational annotations.

Researchers introduced PrivacySkills, a framework to evaluate how LLM agents choose information sources with privacy guidance. They found that agents access confidential sources 30% of the time when users are available, but this increases to 45% when users are unavailable. Providing system-level privacy instructions and skill-level metadata labels can reduce this rate, but combining both has the most significant impact.

A study on brain-AI alignment, exploring whether aligned attention heads in LLMs are causally involved in model computation. The research compares attention-head representations with human EEG and finds that brain-aligned heads contribute to performance but are not as critical as previously thought.

SelfSearch is a reward-free search procedure for self-improving LLM agents. It modifies agents using records of previous self-improvement episodes, capturing reasoning, tool actions, and outcomes. SelfSearch improves population-mean success and reduces execution cost, achieving competitive task success at lower search cost.

V-Engram is a trigger-indexed external memory mechanism for Stable Diffusion 3.5, enabling targeted visual evidence addition without rewriting the generator. It improves fidelity and compositional control for text-to-image models.

This paper introduces Commit-on-Evidence Memory (CoEM), a long-context reasoning method for large language models (LLMs) that preserves potentially useful input information and decides when to convert it into compact memory facts.

Researchers found that pre-trained transformers rely too heavily on initial layers, and a small LoRA modification can improve their ability to follow references in context, increasing accuracy on long chains.

Researchers explore how interaction formats and textual scaffolds can improve decision-making in LLM agents, specifically in auctions and matching environments. They find that certain interfaces and prompts can reduce bid deviations and improve choices, but these improvements may not be reflected in the agents' short-term plans or explanations.

A new framework, CARGO, is proposed to evaluate agentic AI in production by treating retrieved references as procedural exemplars and grounding factual judgments in the live instance's observed context. CARGO eliminates false penalties in reference-based evaluation and retains high contradiction recall.

A paper compares the performance of a purpose-built security context layer, Sola Security Brain, with a general-purpose coding agent, Claude Code, in cloud-security investigation tasks. Sola Security Brain shows a 79.2% relative gain in coverage over Claude Code, with lower reasoning cost per task.