Memory has become a standard substrate for self-evolving agents, yet retaining experience is not the same as learning how to evolve through it. Existing memory agents can store trajectories, retrieve reflections, or accumulate skills, but often lack the holistic competence to select useful experience, act on it, write reusable knowledge, and maintain a growing repository. We introduce OPD-Evolver, a slow-fast co-evolution framework that cultivates such an agent evolver through on-policy self-distillation. In the fast loop, OPD-Evolver interacts with a four-level memory hierarchy to read, use, write, and maintain experience for rapid test-time evolution. In the slow loop, outcome-calibrated memory attribution and privileged hindsight distill these four abilities into the deployable policy. Across multi-domain benchmarks, OPD-Evolver surpasses memory systems such as ReasoningBank by up to 11.5 points, and training-based methods such as Complementary RL by ~6.45 on average. Further analysis shows that OPD-Evolver internalizes high-value experience and memory management, enabling OPD-Evolver-9B to challenge giant counterparts such as Qwen3.5-397B-A17B and Step-3.5-Flash, pointing beyond memory-augmented agents toward genuinely qualified agent evolvers.

Vision-Language Models (VLMs) are increasingly applied to embodied task settings, yet it remains unclear whether their task-progress judgments are grounded in physical state comparison or in input-order shortcuts. We study a diagnostic two-frame setting: given two robot manipulation states and a task goal, the model must decide which state is closer to completion, or whether the two states are effectively equal. We observe a strong Forward-Backward Performance Gap: models often answer correctly in the natural forward order, but fail when the same two physical states are shown with their image slots swapped. This reversal is not intended to mimic the dominant deployment order of robot execution; rather, it is a counterfactual diagnostic for testing whether the selected physical state remains consistent under input-slot inversion. To reduce this order dependence, we introduce Equal-State supervision as a zero-point calibration signal, training models to output equal when no task-relevant progress difference is present instead of defaulting to a directional answer. From real manipulation videos, we construct the Embodied Causal-Temporal Reasoning (ECTR) dataset and propose Tri-Stage Weighted Alignment (TSWA), a staged fine-tuning procedure that strengthens spatial description, task-conditioned comparison, and answer calibration in sequence. On Qwen2.5-VL-7B, TSWA reduces GAP from 60.50% to 5.14% and raises overall accuracy from 44.66% to 80.19%. This comes with a small decrease in forward accuracy, but large gains in backward accuracy, equal-state recognition, and pairwise consistency, suggesting that TSWA replaces order-dependent correctness with more calibrated state comparison.

Multimodal Large Language Models (MLLMs) achieve strong vision-language reasoning but incur large KV caches and high decoding latency with long visual contexts. Existing compression methods rely on observation window attention for stable token importance estimation, yet this aggregation can dilute sparse critical evidence and discard answer-relevant tokens under aggressive compression. We identify last query attention as a complementary signal for recovering such evidence, though its irrelevant signals may introduce additional noise. We propose BACON, a plug-and-play method that calibrates observation window attention with last query evidence while suppressing noise through intra-layer coherence and inter-layer persistence. Across diverse benchmarks, models, budgets, and compression methods, BACON improves multimodal KV-cache compression by 7.5% on average under the most aggressive budget, with gains up to 30.9%.

Touch is a key modality for embodied agents to understand the physical world. Although recent work has incorporated tactile signals into lan- guage systems for tactile Commonsense Rea- soning, scaling such systems to realistic open- world settings remains challenging due to two key bottlenecks: (1) current tactile reasoning datasets remain limited in format and scale, pro- viding insufficient supervision for reasoning from tactile observations to physical common- sense and hindering the learning of transferable tactile commonsense; (2) Tactile signals are inherently redundant and action-specific, yet existing methods often overlook these proper- ties, resulting in inefficient representations with limited semantic expressiveness. To address these limitations, we propose TouchThinker, a tactile-language framework that scales tac- tile commonsense reasoning to the open world from both data and representation perspec- tives. First, we construct TouchThinker-1M, a million-scale, multi-source tactile reasoning dataset covering 410 objects, 8 scenarios, and 7 sensor types, providing a solid data founda- tion for open-world generalization. We fur- ther introduce TouchThinker-Bench, an open- world benchmark with more realistic and di- verse tasks. Then, we further introduce action- aware tactile modeling and a two-stage training paradigm to improve tactile representation effi- ciency and enable efficient reasoning. Experi- mental results demonstrate that TouchThinker achieves competitive performance against state- of-the-art models across multiple datasets. Our code and dataset will be released.