Retrieval-augmented generation (RAG) remains challenging for multi-hop question answering when supporting evidence is distributed across heterogeneous sources, including text, tables, and knowledge graphs (KGs). Existing systems often rely on modality-specific retrieval pipelines or uncontrolled iterative search, making it difficult to preserve source structure while maintaining efficient evidence acquisition. We present **RELOOP**, a structure-aware retrieval framework built on **Hierarchical Sequence (HSEQ)**, a reversible representation that maps heterogeneous sources into provenance-preserving typed segments with lightweight structural links. Operating on this unified evidence interface, a lightweight iterator selects and expands candidate evidence through structure-respecting actions, while an explicit sufficiency signal determines when the accumulated evidence is adequate for answering. A head module then synthesizes the final answer from a compact, canonicalized evidence package, with an optional refinement step when inconsistent evidence is detected. Experiments on HotpotQA, HybridQA, TAT-QA, and MetaQA show that RELOOP consistently improves answer quality over strong single-pass and iterative RAG baselines while requiring substantially fewer retrieval steps than graph-expansion approaches. These results suggest that a reversible, structure-aware evidence interface provides an effective foundation for accurate and budget-controlled retrieval across heterogeneous QA settings.

Privacy protection for live web traffic requires more than detecting private spans. Agent-based privacy protection systems must determine whether an outgoing action complies with the destination site’s privacy policy, then apply only the level of rewriting or sanitisation justified by the residual disclosure risk. We present **GuardianAgent**, a policy-conditioned anonymization framework that couples structured risk assessment with verified adaptive rewriting. GuardianAgent computes risk through **AMRSF** (*Adaptive Multi-factor Risk Scoring Formula*), an explicit controller that combines policy-violation likelihood with data sensitivity, recipient transmission, purpose legitimacy, contextual basis, and policy transparency, rather than relying on an LLM to assign risk directly. This risk score determines both the allow/transform/deny decision and the initial anonymization level. For efficiency, GuardianAgent uses an evidential fast path for low-uncertainty policy matches and invokes an LLM slow path only for uncertain cases. For rewriting, it applies a **five-level hierarchy** driven by a **verified adversarial guesser**: guesses trigger escalation only when supported by the original text, preventing hallucinated attacker confidence from causing unnecessary over-anonymization. Experiments across three benchmarks spanning legal text (*TAB*), Reddit posts (*SynthPAI*), and multi-format synthetic PII records (*PII-Masking-300k*) show that GuardianAgent achieves the strongest privacy–utility trade-off among published baselines and is the only method to reach ≥ 0.90 privacy in all three domains, remaining robust under a backbone switch. Action-context stress tests further show that the same outgoing text receives different decisions and anonymization strengths under different recipients, purposes, action bases, and policy-transparency conditions.

Scope 3 greenhouse gas (GHG) emissions account for the majority of corporate carbon footprints, yet remain difficult to analyze at scale due to sparse disclosures, heterogeneous report document formats, and limited evidence traceability. Existing approaches typically rely on large language models to extract emissions information from ESG reports, but often lack explicit evidence grounding or depend on costly manual annotation and verification to ensure extraction reliability. To address these challenges, we propose **Scope3Trace**, an evidence-grounded information extraction framework designed to extract interpretable and traceable Scope 3 emissions information from real-world ESG and sustainability reports. The framework integrates a document information extraction pipeline that performs PDF collection and OCR parsing, LLM-assisted page localization and table reconstruction, and hybrid rule–LLM extraction of organization- and building-level emissions disclosures with evidence-grounded verification. Building upon this framework, we further contribute a dual-level, evidence-grounded, multimodal dataset comprising organization-level Scope 3 disclosures extracted from heterogeneous sustainability reports. **Scope3Trace** enables reliable extraction and transparent integration of heterogeneous sustainability disclosures, achieving high accuracy in extracting Scope 1–3 totals and category-level disclosures from sustainability reports. The code is available at https://github.com/cruiseresearchgroup/Scope3Trace.

Recent advancements in the Generative Reward Model (GRM) have demonstrated its potential to enhance the reasoning abilities of LLMs through Chain-of-Thought (CoT) prompting. Despite these gains, existing implementations of GRM suffer from two critical limitations. First, CoT prompting is applied indiscriminately to all inputs regardless of their inherent complexity. This introduces unnecessary computational costs for tasks amenable to fast, direct inference. Second, existing approaches primarily rely on voting-based mechanisms to evaluate CoT outputs, which often lack granularity and precision in assessing reasoning quality. In this paper, we propose E-GRM, an efficient generative reward modeling framework grounded in model-internal uncertainty. E-GRM leverages the convergence behavior of parallel model generations to estimate uncertainty and selectively trigger CoT reasoning only when needed, without relying on handcrafted features or task-dependent signals. To improve reward fidelity, we introduce a lightweight discriminative scorer trained with a hybrid regression–ranking objective to provide fine-grained evaluation of reasoning paths. Experiments on multiple reasoning benchmarks show that E-GRM substantially reduces inference cost while consistently improving answer accuracy, demonstrating that model-internal uncertainty is an effective and general signal for efficient reasoning-aware reward modeling.

RAG-based question-answering (QA) in specialist domains faces a cold-start problem: no evaluative benchmark and no labeled data for post-training. We present DoRA (Domain-oriented RAG Assessment), a benchmark construction and evaluation framework that needs only a small set of specialist-domain documents, integrating and expert-validating otherwise established components for this cold-start setting. DoRA generates synthetic QA training and evaluation sets with auditable evidence across five domain-specific intents. To mitigate same-pipeline circularity, its training and test splits use different LLM families (Claude Sonnet for training; GPT-4o for test) drawn from disjoint seed-document corpora. Instantiated on 40 English-language defense documents, DoRA yields approximately 6.6K curated instances. Against 7 LLM baselines over a 1,259-sample benchmark, a LoRA-adapted Llama-3.1-8B trained on the synthetic split improves all 6 coverage and faithfulness metrics—each gain significant under a paired bootstrap test—and reduces hallucination by more than half under the default GTE retrieval setting, with gains persisting across alternative retrievers, prompting-based baselines, and matched answer-length budgets. Defense experts validate the synthetic QA, calibrate the LLM judge, and supply a fully human-written test set. We position DoRA as a practical framework for specialist-domain RAG under domain shift, with defense as a high-stakes case study.

Continual VideoQA with multimodal LLMs remains challenging because sequential adaptation induces task interference, while storing task-specific prompts becomes impractical as task sequences grow. We introduce DynaTokens, a transformer-based token generator that dynamically produces fine-tuning tokens on demand, enabling task-adaptive prompt updates through shared generation weights. To mitigate forgetting, we introduce meta-learning-inspired regularisers that look ahead to avoid task-specific sharp update directions while anchoring the evolving generator to prior-task behaviours. We theoretically connect this objective to sharpness-aware optimisation, showing how it favours flatter cross-task minima and improves retention. DynaTokens combines gradient-free routing based on robust pretrained token and visual embeddings with lightweight auxiliary multimodal supervision, reducing router drift during continual adaptation. Across standard continual VideoQA benchmarks, DynaTokens achieves higher average accuracy and substantially lower forgetting than strong baselines. It also improves zero-shot generalisation and remains effective in longer domain-incremental sequences with extended task shifts. Finally, we introduce a challenging ImageQA->VideoQA protocol and show that DynaTokens enables robust cross-modal continual transfer.

Retrieval-augmented generation over knowledge graphs (KGs) can improve the factual grounding of large language models, but efficient retrieval remains challenging when the underlying graph is large and multi-hop evidence is distributed across different regions. We present **SPLIT-RAG**, an efficiency-oriented multi-agent KG-RAG framework that organises a KG into reusable, question-aligned partitions using reasoning patterns observed in training questions. At inference time, SPLIT-RAG forms a query-specific retrieval plan, activates only a small set of relevant partition-specialised agents, and aggregates their retrieved evidence before final answer generation. To improve reliability when evidence from different agents is combined, the framework applies compatibility-based filtering over retrieved triples prior to synthesis. Experiments on WebQSP, CWQ, and MetaQA show that SPLIT-RAG achieves strong performance against recent KG+LLM baselines while reducing end-to-end latency under matched budgets. Component ablations and model-tier analyses further show that selective multi-agent retrieval is central to the performance gain, and that lightweight retrieval agents paired with a stronger head agent provide an effective accuracy–latency trade-off.

Motion sensor time-series are central to Human Activity Recognition (HAR), yet conventional approaches are constrained to fixed activity sets and typically require costly parameter retraining to adapt to new behaviors. While Large Language Models (LLMs) offer promising open-set reasoning capabilities, applying them directly to numerical time-series often leads to hallucinations and weak grounding. To address this challenge, we propose ZARA (Zero-training Activity Reasoning Agents), a knowledge- and retrieval-augmented agentic framework for motion time-series reasoning in a training-free inference setting. Rather than relying on black-box projections, ZARA distills reference data into a statistically grounded textual knowledge base that transforms implicit signal patterns into verifiable natural-language priors. Guided by retrieved evidence, ZARA iteratively selects discriminative cues and performs grounded reasoning over candidate activities. Extensive experiments on eight benchmarks show that ZARA generalizes robustly to unseen subjects and across datasets, demonstrating strong transferability across heterogeneous sensor domains. These results mark a step toward trustworthy, plug-and-play motion understanding beyond dataset-specific artifacts. Our code is available at https://github.com/zechenli03/ZARA.

Effective long-term memory management is crucial for language models handling extended contexts. We introduce the Enhanced Ranked Memory Augmented Retrieval ERMAR framework, which dynamically ranks memory entries based on relevance. Unlike prior models, ERMAR employs a novel relevance scoring mechanism and a pointwise re-ranking model for key-value embeddings, inspired by learning-to-rank techniques in information retrieval. By integrating historical usage patterns and adaptive retrieval, ERMAR achieves state-of-the-art results on standard benchmarks, demonstrating superior scalability and performance in long-context tasks.

Supervised Fine-Tuning (SFT) is the standard approach for adapting large language models (LLMs) to downstream tasks. However, we observe a persistent failure mode: even after convergence, models often fail to correctly reproduce a subset of their own supervised training data. We refer to this behavior as the Incomplete Learning Phenomenon (ILP). This paper presents the first systematic study of ILP in LLM fine-tuning. We formalize ILP as post-training failure to internalize supervised instances and demonstrate its prevalence across multiple model families, domains, and datasets. Through controlled analyses, we identify five recurrent sources of incomplete learning: (1) missing prerequisite knowledge in the pre-trained model, (2) conflicts between SFT supervision and pre-training knowledge, (3) internal inconsistencies within SFT data, (4) left-side forgetting during sequential fine-tuning, and (5) insufficient optimization for rare or complex patterns. We introduce a diagnostic-first framework that maps unlearned samples to these causes using observable training and inference signals, and study several targeted mitigation strategies as causal interventions. Experiments on Qwen, LLaMA, and OLMo2 show that incomplete learning is widespread and heterogeneous, and that improvements in aggregate metrics can mask persistent unlearned subsets. The findings highlight the need for fine-grained diagnosis of what supervised fine-tuning fails to learn, and why.

Automated simulator construction requires distributional fidelity, distinguishing it from generic code generation. We identify two failure modes in long-horizon LLM agents: contextual drift and optimization instability arising from conflating structural and parametric errors. We propose SOCIA-EVO, a dual-anchored evolutionary framework. SOCIA-EVO introduces: (1) a static blueprint to enforce empirical constraints; (2) a bi-level optimization to decouple structural refinement from parameter calibration; and (3) a self-curating Strategy Playbook that manages remedial hypotheses via Bayesian-weighted retrieval. By falsifying ineffective strategies through execution feedback, SOCIA-EVO achieves robust convergence, generating simulators that are statistically consistent with observational data. SOCIA-EVO’s code and data are available here: https://github.com/cruiseresearchgroup/SOCIA/tree/evo.

Transformer-based models primarily rely on Next Token Prediction (NTP), which predicts the next token in a sequence based on the preceding context. However, NTP’s focus on single-token prediction often limits a model’s ability to plan ahead or maintain long-range coherence, raising questions about how well LLMs can predict longer contexts, such as full sentences within structured documents. While NTP encourages local fluency, it provides no explicit incentive to ensure global coherence across sentence boundaries—an essential skill for reconstructive or discursive tasks. To investigate this, we evaluate three commercial LLMs (GPT-4o, Claude 3.5 Sonnet, and Gemini 2.0 Flash) on Masked Sentence Prediction (MSP) — the task of infilling a randomly removed sentence — from three domains: ROCStories (narrative), Recipe1M (procedural), and Wikipedia (expository). We assess both fidelity (similarity to the original sentence) and cohesiveness (fit within the surrounding context). Our key finding reveals that commercial LLMs, despite their superlative performance in other tasks, are poor at predicting masked sentences in low-structured domains, highlighting a gap in current model capabilities.

Natural language interaction has long served as the primary medium through which humans exchange ideas. A key enabler of this communication is the human capacity for Theory of Mind (ToM)—the ability to infer and align with the mental states of others. ToM is usually modeled as components of desires, beliefs, and intentions. Research in linguistics and psychology has shown that people oftentimes reveal their ToM through pragmatic aspects of language. Considering the advancements in natural language generation and perception that Large Language Models (LLMs) have made in recent years, a critical question arises in relation to ToM: can LLM-powered agents develop similar abilities for inferring mental states during natural language communication? This study investigates the extent to which open-source LLaMA models can represent and retain ToM-related constructs, and whether these internal representations contribute to a coherent mental state modeling in a given conversation. Additionally, we explore the potential for manipulating ToM-related information to generate more aligned responses. Empirical evaluations of LLaMA-3 models (3B and 8B) demonstrate that ToM-informed alignment improves response quality, achieving win rates of 63% and 67%, respectively. These findings suggest that integrating ToM principles can enhance alignment in LLM-based conversational agents. For further details, refer to the code repository.

We introduce SensorLLM, a two-stage framework that enables Large Language Models (LLMs) to perform human activity recognition (HAR) from sensor time-series data. Despite their strong reasoning and generalization capabilities, LLMs remain underutilized for motion sensor data due to the lack of semantic context in time-series, computational constraints, and challenges in processing numerical inputs. SensorLLM addresses these limitations through a Sensor-Language Alignment stage, where the model aligns sensor inputs with trend descriptions. Special tokens are introduced to mark channel boundaries. This alignment enables LLMs to capture numerical variations, channel-specific features, and data of varying durations, without requiring human annotations. In the subsequent Task-Aware Tuning stage, we refine the model for HAR classification, achieving performance that matches or surpasses state-of-the-art methods. Our results demonstrate that SensorLLM evolves into an effective sensor learner, reasoner, and classifier through human-intuitive Sensor-Language Alignment, generalizing across diverse HAR datasets. We believe this work establishes a foundation for future research on time-series and text alignment, paving the way for foundation models in sensor data analysis. Our codes are available at https://github.com/zechenli03/SensorLLM.