Persona-driven role-playing agents tend to drift from their assigned profile or contradict their earlier statements as the dialogue lengthens, a faithfulness failure we call persona inconsistency hallucination. Existing inference-time mitigations apply reward-guided decoding uniformly at every token, while we observe that inconsistency risk is concentrated on a small subset of turns and is locally predictable, which makes a selective intervention feasible. We propose TaDI (Targeted Decoding Intervention), a training-free method with three components: an Inconsistency Risk Predictor that gates intervention before generation, a Targeted Reward-Guided Decoding module that combines a positive profile reward with a negative commitment reward, and a Commitment Propagation Memory that tracks the agent’s prior claims across turns. We additionally introduce a Type-A/B/C violation taxonomy distinguishing profile violation, history contradiction, and commitment fabrication, used as a diagnostic for attributing where any method gains or loses. Extensive experiments show TaDI is competitive with the strongest always-on baseline while being substantially more efficient and more robust under adversarial interrogation.

Despite the remarkable progress of Vision-Language Models (VLMs) in adopting "Thinking-with-Images" capabilities, accurately evaluating the authenticity of their reasoning process remains a critical challenge. Existing benchmarks mainly rely on outcome-oriented accuracy, lacking the capability to assess whether models can accurately leverage fine-grained visual cues for multi-step reasoning. To address these limitations, we propose ViEBench, a process-verifiable benchmark designed to evaluate faithful visual reasoning. Comprising 200 multi-scenario high-resolution images with expert-annotated visual evidence, ViEBench uniquely categorizes tasks by difficulty into perception and reasoning dimensions, where reasoning tasks require utilizing localized visual details with prior knowledge. To establish comprehensive evaluation criteria, we introduce a dual-axis matrix that provides fine-grained metrics through four diagnostic quadrants, enabling transparent diagnosis of model behavior across varying task complexities. Our experiments yield several interesting observations: (1) VLMs can sometimes produce correct final answers despite grounding on irrelevant regions, and (2) they may successfully locate the correct evidence but still fail to utilize it to reach accurate conclusions. Our findings demonstrate that ViEBench can serve as a more explainable and practical benchmark for comprehensively evaluating the effectiveness of agentic VLMs.

Complex decision-making often requires agents to progress through intermediate tasks rather than solve the final target directly. Existing LLM self-refinement methods typically iterate on a fixed target, while curriculum-learning methods often rely on hand-designed schedules, training-time optimization, or domain-specific difficulty metrics. To smooth the learning curve through adaptive curriculum design, we introduce EvoCurr, a general inference-time framework that co-evolves curricula and executable policies. A Designer proposes verifiable intermediate tasks from the latest accepted progress and recent failures, while a Solver generates or trains policies for these tasks. A task-policy pair is accepted only when it passes hard feasibility checks and reaches a specified performance threshold; meanwhile, an accepted-floor constraint preserves the latest verified checkpoint and prevents failed harder attempts from overwriting mastered behavior. We evaluate EvoCurr across code-as-policy and closed-loop MARL settings, using StarCraft II super-late-game micromanagement as our hardest stress test. EvoCurr demonstrates effective micromanagement across three LLM backbones without LLM parameter training. Improvements in additional experiments on Flatland, POGEMA, Overcooked, and LiveCodeBench Hard demonstrate the generality of EvoCurr.

Large language model (LLM) agents increasingly rely on accumulated memory to solve long-horizon decision-making tasks. However, most existing approaches store memory in fixed representations and reuse it at a single or implicit level of abstraction, which limits generalization and often leads to negative transfer when distribution shift. This paper proposes the Meta-Cognitive Memory Abstraction method (MCMA), which treats memory abstraction as a learnable cognitive skill rather than a fixed design choice. MCMA decouples task execution from memory management by combining a frozen task model with a learned memory copilot. The memory copilot is trained using direct preference optimization; it determines how experience should be structured, abstracted, and reused. Memories are further organized into a hierarchy of abstraction levels, enabling selective reuse based on task similarity. When no memory is transferable, MCMA transfers the ability to abstract and manage memory by transferring the memory copilot. Experiments on ALFWorld, ScienceWorld, and BabyAI demonstrate substantial improvements in performance, out-of-distribution generalization, and cross-task transfer over several baselines.

Modern LLM agents persist state across sessions via two channels: long-lived memory files silently prepended to the system prompt, and scheduled tasks waking the agent without user input. A one-time indirect prompt injection committed to either channel can re-fire across sessions long after the attacker leaves, threats we term carryover attacks. Existing prompt-injection benchmarks evaluate attacks only in the session where malicious content is first observed, leaving the gate governing when and how often a planted payload later re-fires unmeasured. We present CarryoverBench: 12 personal-assistant scenarios and 101 attack tasks across three trigger channels: Instant, same-turn IPI; Autonomous, heartbeat-scheduled; and Latent, memory-resurfaced. A rule-based verifier scores planting and triggering as independent stages. Across five frontier models, planted payloads re-fire without measurable decay over five rounds; conditional on successful planting, the trigger stage often provides little additional protection. Porting the Latent channel to a second runtime, Hermes, attenuates but does not eliminate the effect, suggesting that carryover defense requires runtime-layer controls in addition to prompt-level interventions.

Large language models (LLMs) excel at natural language tasks but face deployment challenges due to computational demands. We introduce Dual Activation-Weight Sparsity (DAWS), a training-free framework that jointly exploits activation and weight sparsity through magnitude-based routing. Systematic analysis of pretrained transformers reveals two key observations: (1) the activation energy is concentrated in a few neurons, and (2) activation and weight sparsity patterns are complementary between attention and FFN layers. DAWS employs a three-tier routing strategy: high-magnitude activations pass through full-precision weights to preserve critical pathways, medium-magnitude activations use magnitude-pruned sparse weights for efficiency, and low-magnitude activations are directly discarded. Unlike prior work that uses activation-aware pruning methods like WANDA, our approach uses direct magnitude-based pruning, which we show is more robust to sample-level variations. Experiments on Llama and Mistral models demonstrate that DAWS maintains >98% of dense model performance at 50% sparsity, outperforming WANDA, TEAL, and R-Sparse.