Interactive Natural Language Processing

• iNLP addresses limitations of generative language models, including misalignment, hallucinations, weak grounding, imprecise mathematics, and limited complex reasoning.Existing models such as BART, T5 (Raffel et al., 2020), GPT-3 (Brown et al., 2020), and PaLM (Chowdhery et al., 2022) support many NLP tasks but retain these limitations.Source: §1 · p. 3
• iNLP defines language models as agents that observe, act, and receive feedback in a loop with humans, knowledge bases, models, tools, and environments.Observation includes inputs, action includes generation or requests for external objects, and feedback includes messages such as human scores.Source: §1 · p. 3
• Interactions provide targeted benefits: humans improve alignment with needs and values, knowledge bases reduce hallucinations, tools and models improve reasoning and mathematical exactitude, and environments enhance grounded reasoning and embodied tasks.These interactions extend beyond the conventional human-in-the-loop focus to external resources that address distinct language-model limitations.Source: §1 · pp. 3–4
• The survey fills gaps in prior interaction research by offering a unified and systematic review spanning knowledge-base interaction and methodologies alongside models/tools, environments, and human interaction.Existing works commonly emphasize tool use, reasoning, decision-making, reinforcement learning, embodiment, or evaluation while omitting parts of the broader iNLP spectrum.Source: §1 · p. 4
• The paper organizes iNLP by interactive objects, interaction interfaces, and interaction methods, then surveys evaluation, applications, ethical and safety issues, and future directions.The survey proceeds through interactive objects (§2), interfaces (§3), methods (§4), evaluation (§5), applications (§6), ethics and safety (§7), and future directions and challenges (§8).Source: §1 · pp. 4–5

2 Interactive Objects

• The interactive objects in iNLP are humans, knowledge bases, models/tools, and environments that actively participate in model training or inference.Source: §2 · p. 7

2.1 Human-in-the-loop

• Human-in-the-loop NLP improves Human-LM alignment through conversational prompts, parameter-updating feedback, configurable systems, and scalable user simulation.Prompt interaction supports continuous dialogue but can be labor-intensive because it does not adapt model parameters; feedback directly optimizes parameters, while configuration offers simpler customization.Source: §2.1 · p. 7

2.2 KB-in-the-loop

• KB-in-the-loop augments training with factual representations and inference with dynamically retrieved knowledge for more accurate, contextually relevant responses.Corpus knowledge is controllable but limited and potentially outdated, whereas internet knowledge is broader and current but noisier and harder to control; retrieval quality therefore remains critical.Source: §2.2 · pp. 9–10

2.3 Model/Tool-in-the-loop

• Model/tool-in-the-loop decomposes complex tasks through thinking, acting, and collaborating, assigning subtasks to language models, specialized models, or external tools.External tools are especially useful for faithful or exact tasks and problems requiring information beyond the language model’s training corpus.Source: §2.3 · pp. 11–12
• Model/tool-in-the-loop systems decompose complex tasks across tools or specialized models, enabling executable reasoning, tool use, and tangible effects in external environments.Examples include ToolFormer, ART, ReAct, TaskMatrix.AI, HuggingGPT, OpenAGI, ChatGPT Plugins, and LM-Nav.Source: §2.3 · pp. 12–13
• Collaboration addresses tasks whose division of labor is difficult to specify explicitly by allowing specialized language-model agents to assume roles and coordinate implicitly.In MindCraft, agents with complementary recipe skills communicate to synthesize materials and may discover a division of labor automatically.Source: §2.3 · p. 13
• Closed-loop interaction uses feedback between agents, models, humans, or environments to improve adaptability, robustness, instruction completion, and embodied task execution.Planner-Actor-Reporter returns environmental state to an LLM planner, while Socratic Models and Inner Monologue incorporate multimodal or human feedback into action and planning.Source: §2.3 · p. 13
• Theory-of-Mind and communicative-agent approaches model beliefs, roles, language emergence, and social behavior to support multi-agent cooperation; ChatGPT completed 93% of reported Theory-of-Mind tasks.Communicative systems include referential games, role playing, machine-language emergence, teacher–student interaction, and simulated social behavior.Source: §2.3 · pp. 13–14

2.4 Environment-in-the-loop

• Environment-in-the-loop NLP grounds language by connecting models with real or virtual environments that provide observations, rewards, and state transitions for reasoning, planning, decision-making, and embodied tasks.Grounding is organized along concrete–abstract and low-level–high-level dimensions, with modality grounding and affordance grounding addressing the mismatch between environments and language models.Source: §2.4 · pp. 14–15
• Modality grounding adds visual or audio information, whereas affordance grounding aligns model outputs with scene-scale perception and executable actions; their synergy can produce knowledge-informed yet feasible action sequences.Methods include dual- and single-tower multimodal models, frozen vision-language model interaction, object-action proposals scored by environmental value functions, executable code, and visual navigation.Source: §2.4 · pp. 15–16

3 Interaction Interface

• The interaction-interface taxonomy comprises natural, formal, and machine languages, plus edits and shared memory.Shared memory differs from the other interfaces by enabling indirect communication through information written to a memory pool for later retrieval.Source: §3 · p. 17, §3.5 · p. 21

3.1 Natural Language

• Natural language is the most common interface because it is expressive, accessible without specialized training, interpretable during interaction, and aligned with language-model pretraining.Its flexibility is offset by ambiguity, incoherence, limited parsability, subjectivity, and sensitivity to prompt organization and wording, often requiring prompt engineering.Source: §3.1 · pp. 17–18

3.2 Formal Language

• Formal language improves precision, clarity, parsability, and logic-based interaction through rule-governed representations such as programs, queries, mathematical expressions, grammars, and knowledge triples.Its limitations include specialized accessibility requirements, dependence on formally trained models, high sensitivity to small errors, and reduced expressive flexibility.Source: §3.2 · pp. 18–19

3.3 Edits

• Edits provide an iterative interface that transforms source text through operations such as deletion, insertion, and substitution, either within one model or across cooperating models and modules.Editing-based models support low-resource performance and faster inference, but automatic prototype generation for general NLG remains under-explored, limiting broader adoption.Source: §3.3 · pp. 19–20

3.4 Machine Language

• Machine language enables non-human-readable communication through discrete tokens or continuous vectors, including quantized image and hidden-state sequences, dense visual representations, soft tokens, and metric signals.Discrete machine language uses quantized representations, whereas continuous machine language uses dense scalars or vectors; metric signals are classified according to whether their values are discrete or continuous.Source: §3.4 · pp. 20–21

3.5 Shared Memory

• Shared memory supports indirect interaction by storing messages in hard, human-readable logs or soft, queryable and human-unreadable memory slots.Memory enables access to past information and high-throughput communication, but noise, staleness, retrieval cost, unpredictability, and uncertainty require careful design.Source: §3.5 · pp. 21–22

4.1 Pre-trained Language Models

• The section first reviews pretrained language models as a foundation for understanding and processing interaction messages.Source: §4 · p. 22

4.2–4.6 Interaction Methods

• It then categorizes interaction methods into prompting without training, parameter-updating fine-tuning, active learning, reinforcement learning, and imitation learning.Source: §4 · p. 22

4.7 Interaction Message Fusion

• Finally, it reframes and formalizes these approaches through a unified interaction message fusion framework.Source: §4 · p. 22

• PLMs, especially LLMs, have become foundational to language intelligence, with Transformer, BERT, GPT-3, and ChatGPT identified as major milestones.PLMs are generally Transformer-based and categorized by architecture and pre-training strategy.Source: §4.1 · p. 22

Architectures

• Encoder-only, encoder-decoder, and decoder-only architectures respectively suit discriminative, sequence-to-sequence, and generative tasks.Encoder-only models produce input representations, encoder-decoder models autoregressively decode from encoded inputs, and decoder-only models generate from contextual input text.Source: §4.1 · p. 23

Pre-training Strategies

• Common pre-training objectives include CausalLM, PrefixLM, MLM, Seq2Seq, and sentence relationship capturing, while right-to-left and permutation language modeling are less common.These objectives respectively model preceding-token prediction, bidirectional-prefix prediction, masked-span prediction, encoded-input generation, and inter-sentence relationships.Source: §4.1 · p. 23

• The section surveys representative pre-trained language models, large language models, and multimodal foundation models in Tables 1, 2, and 3.The tables provide overviews of PLMs, LLMs, and MFMs, respectively.Source: Table 1 · p. 22, Table 2 · p. 23, §4.1 · p. 24

4.2 Prompting

• Prompting calls a model through prompts without parameter updating, typically using task descriptions and few-shot examples in context; the discussion focuses on large-scale generative language models.Prompting can be difficult for small language models, which may require fine-tuning with prompts.Source: §4.2 · p. 24

4.2.1 Standard Prompting

• Standard prompting organizes task descriptions and demonstrations through instance selection, processing, and combination, but remains vulnerable to context sensitivity and limited robustness.These operations include retrieving or learning relevant examples, expanding, filtering, editing, formatting, ordering, and structuring demonstrations.Source: §4.2.1 · p. 25
• Standard-prompting methods also communicate retrieved knowledge or task instructions in knowledge-base- and environment-in-the-loop settings, making context organization important for interaction performance.Examples include feeding retrieved knowledge through in-context learning and generating instructions that support cross-environment transfer.Source: §4.2.1 · p. 26

4.2.2 Elicitive Prompting

• Elicitive prompting adds step-by-step guidance to improve reasoning and planning, using techniques such as chain-of-thought, least-to-most decomposition, zero-shot cues, and executable program-based reasoning.Program-of-Thought, PAL, and ViperGPT can verify intermediate results by executing programs, unlike vanilla text-based chain-of-thought.Source: §4.2.2 · p. 26
• Elicitive prompts can also unlock tool use and support reasoning and planning during interactions with other objects, especially in model/tool-in-the-loop settings.Specialized prompts can elicit web search, calculator, and physical-simulation actions.Source: §4.2.2 · p. 27

4.2.3 Prompt Chaining

• Prompt chaining cascades multiple language-model runs with interdependent prompts, decomposing tasks into multi-hop reasoning or pipelined subtasks; the survey classifies schemes by topology and discusses customization and automatization.The four topologies are sequential, branched, decomposed, and interactive.Source: §4.2.3 · p. 27
• Sequential, branched, decomposed, and interactive chains respectively support linear reasoning, aggregation over alternative paths, recursive task decomposition, and alternating communication among specialized node groups.These structures cover examples including question answering, tree search, hierarchical subproblems, multimodal interaction, and thought verification.Source: §4.2.3 · pp. 27–28
• Prompt-chaining systems may be user-configurable through visual interfaces or constructed automatically according to context, improving configurability while enabling dynamic tool selection and invocation.Customization supports transparent pipelines for applications such as writing and advertisement generation, whereas ReAct, MM-ReAct, and ToolFormer automate thought-versus-tool decisions.Source: §4.2.3 · p. 29

4.3 Fine-Tuning

• Fine-tuning uses accumulated interaction messages to adapt language models for instruction following and grounding while targeting catastrophic forgetting, task generalization, and efficient adaptation.The section organizes this work into supervised instruction tuning, continual learning, parameter-efficient fine-tuning, and semi-supervised fine-tuning.Source: §4.3 · p. 29

4.3.1 Supervised Instruction Tuning

• Supervised instruction tuning combines task instructions with multitask data to improve query response and enable zero-shot generalization to unseen tasks.Instruction formats, task mixtures, model architectures, data scale, and chain-of-thought integration are explored; instruction tuning and RLHF improve alignment with human intent, truthfulness, and toxicity, while FLAN-T5 and FLAN-PaLM show robust performance across translation, reasoning, and question answering (Ouyang et al., 2022; Glaese et al., 2022; Chung et al., 2022; Longpre et al., 2023).Source: §4.3.1 · p. 30

4.3.2 Continual Learning

• Continual learning continuously incorporates new interaction knowledge while preserving prior learning, because naïve sequential fine-tuning can cause catastrophic forgetting and degrade earlier-task performance.Regularization, rehearsal, and modularization mitigate forgetting, but the problem remains especially challenging for large language models, with continual-learning evaluation and sample efficiency still unresolved.Source: §4.3.2 · pp. 30–32

4.3.3 Parameter-Efficient Fine-Tuning

• Parameter-efficient fine-tuning updates only a small parameter subset or added modules, reducing GPU memory use and overfitting compared with full-model tuning.Partial fine-tuning selects existing parameters, whereas adapter-based methods freeze the backbone and add trainable modules such as Prefix Tuning, Prompt Tuning, LoRA, and Compacter; these methods support modularity, composition, shareability, and interactive applications across models, knowledge bases, and environments.Source: §4.3.3 · pp. 32–33

4.3.4 Semi-Supervised Fine-Tuning

• Semi-supervised fine-tuning uses labeled and unlabeled interaction data when messages lack sufficient supervision, selecting confident generated labels or distributions for self-training or teacher-based student training.Self-training bootstraps filtered captions, reasoning paths, instructions, and tool interactions, while teacher models generate chain-of-thought data for fine-tuning smaller students.Source: §4.3.4 · pp. 33–34

4.4 Active Learning

• Active learning selects an optimal subset of unlabeled samples, obtains labels from an oracle, and repeats training to maximize performance gains with minimal data expansion.The unlabeled pool may come from corpora, the internet, or knowledge graphs, while the oracle may be a human or automated system.Source: §4.4 · p. 35

Query Strategy

• Query strategies balance informativeness, which maximizes information gained, with representativeness, which reduces sampling bias and excessive outlier weighting.Common approaches include uncertainty, disagreement, performance, density, and batch-diversity strategies.Source: §4.4 · p. 35
• PLM- and LLM-based query strategies can enrich representations, but adapting them may be prohibitively costly or harmful when downstream-task fine-tuning is poor.Schröder et al. (2022) report that LLM adaptation costs can outweigh expected savings, while Margatina et al. (2022) emphasize effective task-specific PLM adaptation.Source: §4.4 · pp. 35–36

Oracle

• Although humans remain the primary oracles, PLMs can generate pseudo-labels, new training data, or relabeled instructions, while knowledge bases, tools, and environments offer further labeling possibilities.Examples include ontology-based labeling, web search and code execution through ReAct and PoT (Chen et al., 2022d), and environment-grounded labeling in MineDojo (Fan et al., 2022) and NLMap-SayCan (Chen et al., 2022a).Source: §4.4 · p. 36

4.5 Reinforcement Learning

• RL for iNLP treats generated text or tool use as actions, surrounding text and non-text inputs as observations, and human preference or affordance grounding as primary rewards.The survey focuses on language-model interaction with environments and humans, drawing on RLEF, RLHF, and RLAIF.Source: §4.5 · p. 36

4.5.1 Feedback Loop

• Online RL updates language models from synchronous rewards, whereas offline RL asynchronously fine-tunes them from static feedback; online RL generally suits environment interaction, while offline RL better suits human interaction.The survey therefore mainly emphasizes online RL for LM–environment interaction and offline RL for LM–human interaction.Source: §4.5.1 · p. 37, Figure 23 · p. 37
• Online RL uses dense step-level scalar or boolean rewards for immediate learning, sparse episode-level rewards for long-horizon tasks, and buffered feedback for periodically redeployed services.These reward schedules accommodate continuous movement, multi-action tasks, and production-scale training costs.Source: §4.5.1 · p. 37
• Offline RL applies algorithms including PPO, A2C, ILQL, TRPO, and NLPO to align language models with human preferences, but alignment costs can degrade other-task performance.Pretraining-time offline RL, early stopping, and improved optimization target better alignment–quality tradeoffs, computational efficiency, robustness, and training stability.Source: §4.5.1 · p. 38

4.5.2 Reward Modeling

• Environment feedback supports affordance grounding through absolute, binary, handcrafted, comparative, and progress-based rewards, while human-feedback models provide task-sensitive scores or structured guidance for open-ended generation and reasoning.RLAIF extends model-based feedback by enabling self-adjustment from self-feedback and self-prompts, indirectly reflecting human-annotated data.Source: §4.5.2 · pp. 38–39

• Imitation learning trains an agent’s policy or reward function by mimicking expert demonstrations, supporting applications across autonomous systems, robotics, human–agent interaction, and NLP tasks.Unlike reinforcement learning, it relies on behavior demonstrations rather than manually designed rewards or learned reward models.Source: §4.6 · p. 39

Offline Imitation Learning

• Offline imitation learning uses stored state–action demonstrations to train models through supervised behavior cloning, including token- or sequence-level text generation.The expert policy is encoded in offline data, after which the model can perform the task independently.Source: §4.6 · pp. 39–40

Online Imitation Learning

• Online imitation learning updates models through expert feedback during interaction, including feedback from simulated experts when human consultation is infeasible.Adjusting the expert simulator can reinforce or discourage behaviors to better align models with human expectations.Source: §4.6 · p. 40

• Imitation learning suffers from exposure bias because models rarely encounter states caused by their own actions, producing distribution shifts and accumulated errors.Alternating expert and model policies enables corrective feedback, but limited online access, demonstration quality, and data requirements remain unresolved challenges.Source: §4.6 · p. 40

Summary

• The three dimensions are not mutually exclusive, so methods can combine categories, with choices shaped by the task, data, resources, and performance–computational-efficiency trade-off.The framework is intended to help researchers and practitioners understand existing methods and design new approaches.Source: §4.7 · pp. 40, 42

Data dimension

• Along the data dimension, interaction messages are categorized as metrics, constraints, or freeform text.Metrics include rewards or ranking scores, constraints include keywords, templates, skeletons, and constrained vocabularies, and freeform messages are unstructured text.Source: §4.7 · pp. 40–41

• Interaction message fusion methods are organized across three dimensions: data, model, and training.Source: Figure 25 · p. 41

Model dimension

• Along the model dimension, methods are model-invariant, model-adjusted, or model-expanded depending on whether they preserve, modify, or add architectural components.Examples include prompting and tuning without architectural changes, VaLM and K-BERT with modified fusion mechanisms, and K-Adapter, Retro, RelationLM, and KELM with additional modules.Source: §4.7 · p. 41

Training dimension

• Along the training dimension, methods are classified as frozen, partial tuning, or full tuning according to how many model parameters are updated.Prompting and some constrained decoding methods can be frozen; K-Adapter tunes only adapters; ToolFormer, Kepler, and KnowBERT fully train model parameters.Source: §4.7 · pp. 41–42

5 Evaluation

• iNLP evaluation addresses interaction quality that reference-based NLG metrics often miss, motivating a four-way taxonomy by interactive object.Metrics such as BLEU reward lexical similarity to references, which may overlook user preferences; the survey therefore organizes interaction evaluation into human-, knowledge-base-, model/tool-, and environment-in-the-loop types.Source: §5 · pp. 42–43

5.1 Evaluating Human-in-the-loop Interaction

• Human-in-the-loop evaluation uses general alignment metrics and task-specific measures, but comprehensive benchmarks remain difficult because interaction processes and user diversity complicate data collection.General evaluation covers helpfulness, honesty, and harmlessness, while task-specific metrics include dialogue success, turn size, and dialogue-state accuracy; interactive users can also evaluate models differently from third-party annotators.Source: §5.1 · pp. 43–44

5.2 Evaluating KB-in-the-loop Interaction

• KB-in-the-loop evaluation measures retrieval effectiveness and knowledge-enhanced generation through retrieval recall, hallucination detection, factual attribution, and comparisons with human knowledge use.Examples include REALM’s “top-5 recall,” Knowledge F1, fact-checking benchmarks such as BEGIN and DialFact, AIS attribution, and human comparisons in WebGPT and GopherCite.Source: §5.2 · p. 44

5.3 Evaluating Model/Tool-in-the-loop Interaction

• Model/tool-in-the-loop evaluation covers chain-of-thought reasoning, tool-use triggering, accuracy and proficiency, and collaborative behavior through outcome- and process-oriented analyses.Evaluations examine reasoning errors and trajectory structure, API invocation and tool-selection success, task success, theory of mind, and communication efficiency.Source: §5.3 · pp. 44–45

5.4 Evaluating Environment-in-the-loop Interaction

• Environment-in-the-loop evaluation requires embodied task platforms and metrics capturing mission success, planning and execution success, action efficiency, and instruction-generation speed.Platforms include EvalAI, SAPIEN, Alexa Arena, ALFRED, ALFWorld, and MindCraft; metrics distinguish effectiveness from efficiency in embodied tasks.Source: §5.4 · pp. 45–46

6.1 Controllable Text Generation

• iNLP improves controllable text generation through human preferences, knowledge-aware training, classifier or self-feedback, and environment-based action scoring.These interactions support user-oriented, robust, fine-grained, complex, and reality-oriented control over generated text.Source: §6.1 · pp. 46–47

6.2 Writing Assistant

• Writing assistants range from content support and iterative polishing to enrichment and human-AI co-creation, with increasing involvement in the generation process.Support tools preserve writer agency, while co-creation can improve coherence and preference but faces voice, creativity, and evaluation challenges.Source: §6.2 · pp. 47–48
• Content enrichment interactively supplements user input through autocomplete and keyword-to-sentence generation, enabling fine-grained preferences across writing and downstream NLP tasks.CueBot, for example, lets users select responses controlled by cues or keywords and trains with a keyword loss for lexically constrained outputs.Source: §6.2 · p. 48

6.3 Embodied AI

• Embodied AI uses textual observation, commands, sensors, and feedback for manipulation, navigation, exploration, and multi-role cooperation in physical or virtual environments.Agents can coordinate with humans and other agents for tasks including object rearrangement, tool use, driving assistance, negotiation, and scheduling.Source: §6.3 · pp. 48–49

6.4 Text Game

• Text games are intrinsic iNLP applications in which language models interpret commands, simulate states, update environments, and interact through text-only or text-aided interfaces.Platforms and tasks span adventure, social deduction, strategic, tabletop role-playing, life simulation, and social simulation games.Source: §6.4 · pp. 49–51
• Text-game research addresses partial observability, large state spaces, long-term credit assignment, semantic understanding, and extracting reliable game states and commands.PLM- and LLM-based agents reduce reliance on handcrafted templates, while language interfaces support communication, planning, policy generalization, and representation learning.Source: §6.4 · pp. 50–52

6.5 Other Applications

• Interactive objects can specialize language models for domains such as medicine and law by retrieving and reasoning over records, documents, knowledge graphs, and tools.This approach can improve data and computation efficiency relative to adapting models only through large domain-specific corpora.Source: §6.5 · p. 52
• Personalization adapts model behavior through user interactions, profiles, prompts, or external memory, improving relevance and trust while potentially magnifying toxicity.Deshpande et al. (2023) report that assigning a persona to ChatGPT can increase toxicity up to six times.Source: §6.5 · p. 52
• In-world context concerns a player’s interactions with other players’ characters and NPCs during role-playing.Source: §6.5 · p. 52
• On-brand context concerns a player’s interactions with other players and the game host outside role-playing.Source: §6.5 · p. 52
• Interactive evaluation combines statistical metrics, LLM judges, and model–human collaboration to assess generated content more accurately, efficiently, and automatically.LLMs have been studied as human-like evaluators across summarization, question answering, news generation, and translation, with some work targeting cheaper evaluation.Source: §6.5 · p. 53

• LLM applications raise societal risks involving education, bias and fairness, privacy, harmful content, and misinformation.Source: §7 · p. 53

Impact on Education

• Education systems face misuse of ChatGPT for assignments, while AI-content detectors remain ineffective in real-world settings and writing tools prioritize productivity over skill development.Tools including GPTZero 49, AI Classifier 50, and DetectGPT 51 commonly rely on perplexity and burstiness.Source: §7 · p. 53

Social Bias

• Language models trained on web corpora can amplify stereotypes, toxicity, racism, and discrimination in applications such as resume ranking and classification.Bias has been observed in embeddings and downstream outputs, including disability-related toxicity assessments (Hutchinson et al., 2020).Source: §7 · pp. 53–54
• Bias evaluation should address both broad societal patterns and sub-regional contextual effects, while mitigation is needed to prevent harms in decisions such as recruiting.HERB (Li et al., 2022i) evaluates bias in a sub-region to assess bias at the regional level.Source: §7 · pp. 53–54

Privacy Concern

• LLMs can expose sensitive personal information through training data and prompting, making privacy preservation increasingly urgent as human interaction with models expands.Sensitive information has been distilled from LLMs through prompting (Carlini et al., 2021; 2023), motivating policies for data collection, storage, and anonymization.Source: §7 · p. 54

Alignment

• Alignment research should detect knowledge blind spots, enable scalable knowledge editing, and replace vulnerable scalar rewards with diverse, fine-grained feedback.Factual alignment requires acknowledging missing knowledge, while value alignment must address reward hacking and improve training stability.Source: §8 · p. 54

Social Embodiment; Context Length; Long Text Generation

• Future systems should scale to complex social environments, support longer contexts and coherent long-form generation, and improve reasoning over extended interactions and action plans.Longer context windows support conversation history, large-document understanding, environmental observations, and long-form content, but increasing context length and generation quality remains challenging.Source: §8 · pp. 54–55

Plasticity

• Continual-learning methods should update iNLP models efficiently without excessive parameter changes or catastrophic loss of previously acquired knowledge.More human-like models and biological plasticity may guide adaptation to dynamic real-world situations.Source: §8 · pp. 54–55

Speed & Efficiency; Accessibility

• Efficient iNLP requires parameter-free inference acceleration that preserves performance while also accounting for robustness, bias, and alignment.Accessibility also requires more open models and deployment on edge devices, expanding offline interaction possibilities.Source: §8 · pp. 55–56

Analysis; Creativity

• Research should clarify language-model inner workings and develop creativity through explicit criteria, novel knowledge generation, useful outputs, and potentially new thinking paradigms.Interpretability may improve interaction behavior and facilitate knowledge updates, while creativity could enable richer writing assistance and debates between language-model agents.Source: §8 · p. 56

Evaluation

• iNLP evaluation should become more diverse by assessing interaction processes, establishing standardized benchmarks, and directly measuring model interactivity.Evaluation choices will substantially influence the design and optimization of interactive frameworks.Source: §8 · p. 56

• The survey proposes a unified definition and framework for iNLP and classifies its interactive objects, interfaces, and methods.This positions language models as interactive agents operating across diverse contexts.Source: §9 · p. 56
• It reviews iNLP evaluation methodologies, applications, ethical and safety issues, and future research directions.Source: §9 · p. 56
• Interactions with humans, knowledge bases, models, tools, and environments can improve alignment, personalize responses, enrich representations, avoid hallucinations, decompose complex tasks, and ground language in reality.Source: §9 · p. 56
• The survey provides a wide-angle view of iNLP’s current state and future potential as a reference for researchers entering this rapidly evolving field.Source: §9 · p. 57

Project Management and Literature Collection

• Ge Zhang and Zekun Wang co-led project management, including outlining, motivation, meetings, task allocation, member invitations, progress monitoring, and quality control.All members collected and presented papers, with Zekun Wang sharing most literature and Ge Zhang, Guangzheng Xiong, Kexin Yang, and Shaochun Hao sharing substantial literature.Source: A Contributions · p. 109

Foundational Sections

• Contributors wrote and edited the introduction, interactive objects, and interaction interfaces, including reorganizations of knowledge sources, tool use, multi-model collaboration, and interface types.Interactive-objects revisions divided Model/Tool-in-the-loop into Thinking, Acting, and Collaborating, while interface coverage included Edits, Natural Language, Formal Language, Machine Language, and Shared Memory.Source: A Contributions · p. 109

Interaction Methods

• The interaction-methods section combined contributions on pre-trained language models, prompting, instruction and fine-tuning, learning, message fusion, and collaborative editing and guidance.Its subsections covered standard, elicitive, and chained prompting; supervised, continual, parameter-efficient, and semi-supervised fine-tuning; active, reinforcement, and imitation learning; and interaction message fusion.Source: A Contributions · p. 109

Evaluation

• Kexin Yang wrote the evaluation section, with Zekun Wang adding material on knowledge acquisition, chain-of-thought capability, tool-use ability, and collaborative behavior analysis.Dayiheng Liu and Zekun Wang also provided suggestions and edited the section.Source: A Contributions · p. 109

Applications

• The applications section covered controllable text generation, writing assistance, embodied AI, text games, and other applications through contributions from multiple authors and editors.Ke Xu and Zekun Wang provided suggestions and edited the section.Source: A Contributions · p. 109

Ethics, Future Directions, and Supervision

• The paper’s ethics and safety and future-directions sections were authored collaboratively, while supervisors provided funding, discussions, and guidance throughout the project.Mong Yuan Sim authored ethics and safety; eleven contributors collaboratively wrote future directions; Shi Wang provided funding and discussions, and Jie Fu, Chenghua Lin, and Yike Guo guided the paper.Source: A Contributions · pp. 109–110