context-conditioned generation

This study addresses the growing demand in digital applications for high-quality, semantically aligned, and contextually relevant AI-generated sound effects that exhibit both diversity and controllability. It presents a systematic review of sound effect generation models from the past five years, encompassing approaches driven by text, visual, audio, and multimodal inputs. By analyzing 30 peer-reviewed papers sourced from Google Scholar, IEEE Xplore, and ACM Digital Library, this work offers the first comprehensive comparison of how different input modalities influence generation performance. The review highlights advances in audio fidelity, semantic alignment, and temporal coherence, while identifying persistent challenges such as temporal synchronization and perceptual consistency. Furthermore, it underscores a significant gap between current objective evaluation metrics and human perceptual judgments.

This study addresses the frequent inefficiencies in human-AI collaboration caused by incomplete contextual information, which often leads to excessive iteration and suboptimal output quality. To mitigate this, the authors propose a structured context construction framework that integrates a five-role context package—comprising authority, exemplars, constraints, evaluation criteria, and metadata—within a four-stage workflow encompassing review, design, construction, and audit. Notably, this work pioneers the incorporation of information theory and reliability engineering principles into context quality assessment, yielding a reusable and auditable collaboration framework. Empirical results from 200 interaction trials demonstrate that the approach reduces the average number of iterations from 3.8 to 2.0, increases first-pass success rates from 32% to 55%, and achieves a final task success rate of 91.5%.

This work proposes Meta Context Engineering (MCE), a novel framework that overcomes the structural biases and limited optimization space inherent in existing context engineering approaches, which rely on manually designed, fixed pipelines. MCE introduces a bilevel agent co-evolution paradigm that dynamically optimizes both the structure and content of prompts through agent crossover, skill history tracing, execution feedback, and programmable context representations. By moving beyond static heuristics, the method significantly enhances context adaptability, transferability, and training efficiency. Empirical evaluations across five tasks demonstrate consistent improvements, yielding relative performance gains ranging from 5.6% to 53.8% (average 16.9%) over state-of-the-art agent-based context engineering methods.

This paper addresses system synthesis under partial observability by introducing “Synthesis with Guided Environments” (SGE): a novel paradigm wherein the system, while satisfying a global LTL specification, may issue procedural directives to the environment (e.g., “assign hidden input x to output y”) to actively coordinate task completion. Unlike classical synthesis—which treats the environment as purely adversarial—SGE models the environment as a programmatically controllable collaborator. To formalize this, the authors propose an extended LTL specification framework, design a memory-bounded synthesis algorithm, and conduct theoretical analysis of distributed synthesis. Their results fully settle the decidability of SGE, establish strict improvements over classical synthesis in both state complexity and memory requirements, and prove an exact equivalence between SGE and partially observable distributed synthesis.

Autoregressive models suffer from high generation latency due to sequential, token-by-token dependency, limiting inference efficiency. This paper proposes a context-aware pipelined parallel decoding architecture: the output sequence is partitioned into multiple subsequences, each generating one new token synchronously per step, with a lightweight context synchronization mechanism ensuring coherence. The method introduces no additional parameters or KV cache overhead, preserving autoregressive modeling capability and generation quality while enabling multi-token parallelism. Experiments on question answering, summarization, and keyword generation show 1.8–2.3× speedup in inference latency, with negligible degradation (<0.5 points) in BLEU and ROUGE scores and virtually unchanged memory footprint. The core contribution is the first integration of a structured pipelined mechanism into autoregressive decoding—achieving high-quality parallel generation without any memory overhead.

Existing bundle generation methods struggle to simultaneously satisfy fixed-size constraints and accurately model user intent, resulting in poor interpretability and limited intelligence. This paper proposes an intent-generation joint modeling paradigm for personalized bundle recommendation: leveraging large language models (LLMs), it integrates retrieval-augmented generation (RAG) with adaptive in-context learning, where similar historical dialogues are retrieved to provide customized demonstrations. We introduce a novel dual-task cooperative self-correction mechanism and error-aware automatic feedback supervision, enabling end-to-end joint optimization of intent inference and bundle generation without labeled data. Evaluated on three real-world datasets, our method significantly outperforms state-of-the-art approaches—improving bundle relevance by 12.7% and intent consistency by 19.3%. It effectively mitigates LLM hallucination, enhancing both generation reliability and interpretability.

This work addresses the challenge that large language models struggle to effectively extract and apply novel knowledge from complex task contexts during in-context learning. To overcome this limitation, the authors propose a high-fidelity chain-of-thought synthesis mechanism that enhances the model’s understanding and utilization of task-specific context through the generation of high-quality synthetic reasoning paths. The proposed approach substantially improves model performance on context-dependent tasks, achieving a significant increase in average solve rate—from 17.2% to a markedly higher level—on the CL-Bench benchmark. This advancement effectively narrows the performance gap between state-of-the-art models on such challenging tasks.

Current approaches to controlling the outputs of large language models rely on manually defined context-sensitive constraints, which are difficult to scale and require significant expertise. This work proposes a novel two-stage framework that integrates context-sensitive grammar learning with large language model generation for the first time: it first automatically collects diverse model outputs through syntactic exploration to learn constraints, then enforces the learned rules during text generation. The method guarantees output validity without any human intervention and achieves a 100% constraint compliance rate using only a 1B-parameter model—outperforming both larger-scale models and state-of-the-art reasoning systems in constrained generation tasks.

AI systems exhibit insufficient understanding of human context and intent. Method: This study systematically defines “context engineering” for the first time and traces its two-decade evolution, integrating philosophical inquiry, historical analysis, human–computer interaction (HCI) design principles, and social relational theory to develop a capability-driven, stage-based HCI theoretical model and propose a foundational theoretical framework for context engineering in AI systems. Contributions/Results: (1) It establishes context engineering as an independent research domain by delineating its conceptual boundaries and core theoretical tenets; (2) it identifies the dynamic evolution of context modeling as AI autonomy increases; and (3) it provides transferable conceptual tools and standardized methodologies for context design, enabling AI systems to transition from reactive behavior to contextual awareness. This work advances the theoretical foundations and practical design paradigms for context-aware intelligent systems.

Existing in-context learning approaches lack systematic solutions to challenges such as credit assignment, overfitting, catastrophic forgetting, local optima, and high variance. This work proposes Reflective Contextual Learning (RCL), a novel framework that, for the first time, systematically introduces classical optimization primitives into contextual space. RCL enables agents to optimize their learning through interaction with the environment, reflection on actions and failures, and iterative refinement of context representations. The framework unifies prior methods while incorporating mechanisms including directional update signals, contextual perturbation, batch processing, failure replay, trajectory grouping, curriculum strategies, and variants of optimizer states. Experimental results demonstrate that RCL significantly outperforms strong baselines on AppWorld, BrowseComp+, and RewardBench2, validating both the efficacy and cross-task transferability of its constituent optimization primitives.

This work addresses the issue of cross-batch mode collapse in large language models, wherein repeated independent prompting leads to a progressive decline in output diversity. To systematically mitigate this problem without model fine-tuning or architectural modifications, the authors propose the Dynamic Context Evolution (DCE) framework. DCE integrates three key components: model self-evaluation to filter high-probability outputs, persistent memory deduplication via semantic embeddings (all-MiniLM-L6-v2) and HDBSCAN clustering, and adaptive prompt evolution. Experiments demonstrate that DCE achieves 0% mode collapse—compared to a 5.6% baseline—while consistently generating 17–18 distinct semantic clusters per seed across multiple domains and models. The method operates at a cost of approximately $0.50 per thousand generations, offering robustness, scalability, and sustained structural coherence alongside diversity.