Bagging-Based Model Merging for Robust General Text Embeddings

General-purpose text embedding models underpin a wide range of NLP and information retrieval applications, and are typically trained on large-scale multi-task corpora to encourage broad generalization. However, it remains unclear how different multi-task training strategies compare in practice, and how to efficiently adapt embedding models as new domains and data types continually emerge. In this work, we present a systematic study of multi-task training for text embeddings from two perspectives: data scheduling and model merging. We compare batch-level shuffling, sequential training variants, two-stage training, and multiple merging granularities, and find that simple batch-level shuffling consistently yields the strongest overall performance, suggesting that task conflicts are limited and training datasets are largely complementary. Despite its effectiveness, batch-level shuffling exhibits two practical limitations: suboptimal out-of-domain (OOD) generalization and poor suitability for incremental learning due to expensive full retraining. To address these issues, we propose Bagging-based rObust mOdel Merging (\modelname), which trains multiple embedding models on sampled subsets and merges them into a single model, improving robustness while retaining single-model inference efficiency. Moreover, \modelname naturally supports efficient incremental updates by training lightweight update models on new data with a small historical subset and merging them into the existing model. Experiments across diverse embedding benchmarks demonstrate that \modelname consistently improves both in-domain and OOD performance over full-corpus batch-level shuffling, while substantially reducing training cost in incremental learning settings.

PersonaAct: Simulating Short-Video Users with Personalized Agents for Counterfactual Filter Bubble Auditing

Short-video platforms rely on personalized recommendation, raising concerns about filter bubbles that narrow content exposure. Auditing such phenomena at scale is challenging because real user studies are costly and privacy-sensitive, and existing simulators fail to reproduce realistic behaviors due to their reliance on textual signals and weak personalization. We propose PersonaAct, a framework for simulating short-video users with persona-conditioned multimodal agents trained on real behavioral traces for auditing filter bubbles in breadth and depth. PersonaAct synthesizes interpretable personas through automated interviews combining behavioral analysis with structured questioning, then trains agents on multimodal observations using supervised fine-tuning and reinforcement learning. We deploy trained agents for filter bubble auditing and evaluate bubble breadth via content diversity and bubble depth via escape potential. The evaluation demonstrates substantial improvements in fidelity over generic LLM baselines, enabling realistic behavior reproduction. Results reveal significant content narrowing over interaction. However, we find that Bilibili demonstrates the strongest escape potential. We release the first open multimodal short-video dataset and code to support reproducible auditing of recommender systems.

D-Models and E-Models: Diversity-Stability Trade-offs in the Sampling Behavior of Large Language Models

The predictive probability of the next token (P_token) in large language models (LLMs) is inextricably linked to the probability of relevance for the next piece of information, the purchase probability of the next product, and the execution probability of the next action-all of which fall under the scope of the task-level target distribution (P_task). While LLMs are known to generate samples that approximate real-world distributions, whether their fine-grained sampling probabilities faithfully align with task requirements remains an open question. Through controlled distribution-sampling simulations, we uncover a striking dichotomy in LLM behavior, distinguishing two model types: D-models (e.g. Qwen-2.5), whose P_token exhibits large step-to-step variability and poor alignment with P_task; and E-models (e.g. Mistral-Small), whose P_token is more stable and better aligned with P_task. We further evaluate these two model types in downstream tasks such as code generation and recommendation, revealing systematic trade-offs between diversity and stability that shape task outcomes. Finally, we analyze the internal properties of both model families to probe their underlying mechanisms. These findings offer foundational insights into the probabilistic sampling behavior of LLMs and provide practical guidance on when to favor D- versus E-models. For web-scale applications, including recommendation, search, and conversational agents, our results inform model selection and configuration to balance diversity with reliability under real-world uncertainty, providing a better level of interpretation.

Gated Differentiable Working Memory for Long-Context Language Modeling

Long contexts challenge transformers: attention scores dilute across thousands of tokens, critical information is often lost in the middle, and models struggle to adapt to novel patterns at inference time. Recent work on test-time adaptation addresses this by maintaining a form of working memory -- transient parameters updated on the current context -- but existing approaches rely on uniform write policies that waste computation on low-utility regions and suffer from high gradient variance across semantically heterogeneous contexts. In this work, we reframe test-time adaptation as a budget-constrained memory consolidation problem, focusing on which parts of the context should be consolidated into working memory under limited computation. We propose Gdwm (Gated Differentiable Working Memory), a framework that introduces a write controller to gate the consolidation process. The controller estimates Contextual Utility, an information-theoretic measure of long-range contextual dependence, and allocates gradient steps accordingly while maintaining global coverage. Experiments on ZeroSCROLLS and LongBench v2 demonstrate that Gdwm achieves comparable or superior performance with 4$\times$ fewer gradient steps than uniform baselines, establishing a new efficiency-performance Pareto frontier for test-time adaptation.

Do We Always Need Query-Level Workflows? Rethinking Agentic Workflow Generation for Multi-Agent Systems

Multi-Agent Systems (MAS) built on large language models typically solve complex tasks by coordinating multiple agents through workflows. Existing approaches generates workflows either at task level or query level, but their relative costs and benefits remain unclear. After rethinking and empirical analyses, we show that query-level workflow generation is not always necessary, since a small set of top-K best task-level workflows together already covers equivalent or even more queries. We further find that exhaustive execution-based task-level evaluation is both extremely token-costly and frequently unreliable. Inspired by the idea of self-evolution and generative reward modeling, we propose a low-cost task-level generation framework \textbf{SCALE}, which means \underline{\textbf{S}}elf prediction of the optimizer with few shot \underline{\textbf{CAL}}ibration for \underline{\textbf{E}}valuation instead of full validation execution. Extensive experiments demonstrate that \textbf{SCALE} maintains competitive performance, with an average degradation of just 0.61\% compared to existing approach across multiple datasets, while cutting overall token usage by up to 83\%.

Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents

Memory enables Large Language Model (LLM) agents to perceive, store, and use information from past dialogues, which is essential for personalization. However, existing methods fail to properly model the temporal dimension of memory in two aspects: 1) Temporal inaccuracy: memories are organized by dialogue time rather than their actual occurrence time; 2) Temporal fragmentation: existing methods focus on point-wise memory, losing durative information that captures persistent states and evolving patterns. To address these limitations, we propose Temporal Semantic Memory (TSM), a memory framework that models semantic time for point-wise memory and supports the construction and utilization of durative memory. During memory construction, it first builds a semantic timeline rather than a dialogue one. Then, it consolidates temporally continuous and semantically related information into a durative memory. During memory utilization, it incorporates the query's temporal intent on the semantic timeline, enabling the retrieval of temporally appropriate durative memories and providing time-valid, duration-consistent context to support response generation. Experiments on LongMemEval and LoCoMo show that TSM consistently outperforms existing methods and achieves up to 12.2% absolute improvement in accuracy, demonstrating the effectiveness of the proposed method.

MI-PRUN: Optimize Large Language Model Pruning via Mutual Information

Large Language Models (LLMs) have become indispensable across various domains, but this comes at the cost of substantial computational and memory resources. Model pruning addresses this by removing redundant components from models. In particular, block pruning can achieve significant compression and inference acceleration. However, existing block pruning methods are often unstable and struggle to attain globally optimal solutions. In this paper, we propose a mutual information based pruning method MI-PRUN for LLMs. Specifically, we leverages mutual information to identify redundant blocks by evaluating transitions in hidden states. Additionally, we incorporate the Data Processing Inequality (DPI) to reveal the relationship between the importance of entire contiguous blocks and that of individual blocks. Moreover, we develop the Fast-Block-Select algorithm, which iteratively updates block combinations to achieve a globally optimal solution while significantly improving the efficiency. Extensive experiments across various models and datasets demonstrate the stability and effectiveness of our method.

Circular Reasoning: Understanding Self-Reinforcing Loops in Large Reasoning Models

Despite the success of test-time scaling, Large Reasoning Models (LRMs) frequently encounter repetitive loops that lead to computational waste and inference failure. In this paper, we identify a distinct failure mode termed Circular Reasoning. Unlike traditional model degeneration, this phenomenon manifests as a self-reinforcing trap where generated content acts as a logical premise for its own recurrence, compelling the reiteration of preceding text. To systematically analyze this phenomenon, we introduce LoopBench, a dataset designed to capture two distinct loop typologies: numerical loops and statement loops. Mechanistically, we characterize circular reasoning as a state collapse exhibiting distinct boundaries, where semantic repetition precedes textual repetition. We reveal that reasoning impasses trigger the loop onset, which subsequently persists as an inescapable cycle driven by a self-reinforcing V-shaped attention mechanism. Guided by these findings, we employ the Cumulative Sum (CUSUM) algorithm to capture these precursors for early loop prediction. Experiments across diverse LRMs validate its accuracy and elucidate the stability of long-chain reasoning.

Projecting Out the Malice: A Global Subspace Approach to LLM Detoxification

Large language models (LLMs) exhibit exceptional performance but pose inherent risks of generating toxic content, restricting their safe deployment. While traditional methods (e.g., alignment) adjust output preferences, they fail to eliminate underlying toxic regions in parameters, leaving models vulnerable to adversarial attacks. Prior mechanistic studies characterize toxic regions as "toxic vectors" or "layer-wise subspaces", yet our analysis identifies critical limitations: i) Removed toxic vectors can be reconstructed via linear combinations of non-toxic vectors, demanding targeting of entire toxic subspace; ii) Contrastive objective over limited samples inject noise into layer-wise subspaces, hindering stable extraction. These highlight the challenge of identifying robust toxic subspace and removing them. Therefore, we propose GLOSS (GLobal tOxic Subspace Suppression), a lightweight method that mitigates toxicity by identifying and eliminating this global subspace from FFN parameters. Experiments on LLMs (e.g., Qwen3) show GLOSS achieves SOTA detoxification while preserving general capabilities without requiring large-scale retraining. WARNING: This paper contains context which is toxic in nature.

BaseCal: Unsupervised Confidence Calibration via Base Model Signals

Reliable confidence is essential for trusting the outputs of LLMs, yet widely deployed post-trained LLMs (PoLLMs) typically compromise this trust with severe overconfidence. In contrast, we observe that their corresponding base LLMs often remain well-calibrated. This naturally motivates us to calibrate PoLLM confidence using the base LLM as a reference. This work proposes two ways to achieve this. A straightforward solution, BaseCal-ReEval, evaluates PoLLM's responses by feeding them into the base LLM to get average probabilities as confidence. While effective, this approach introduces additional inference overhead. To address this, we propose BaseCal-Proj, which trains a lightweight projection to map the final-layer hidden states of PoLLMs back to those of their base LLMs. These projected states are then processed by the base LLM's output layer to derive base-calibrated confidence for PoLLM's responses. Notably, BaseCal is an unsupervised, plug-and-play solution that operates without human labels or LLM modifications. Experiments across five datasets and three LLM families demonstrate the effectiveness of BaseCal, reducing Expected Calibration Error (ECE) by an average of 42.90\% compared to the best unsupervised baselines.