Grounded Text Generation

• Grounded Text Generation is defined as producing text that is conditioned on external grounding signals, enhancing traceability and control.
• It employs diverse substrates such as textual sources, video frames, and structured data to support coherent and evidence-backed outputs.
• Key methodologies include retrieval-plus-generation, decode-time composition, and planning-first strategies to optimize factual grounding and performance.

Grounded Text Generation (GTG) refers to generating text that is explicitly conditioned on external grounding signals rather than relying only on a model’s internal priors. In the literature, those signals include external textual sources, dialog belief state and knowledge base results, retrieved documents, video frames, non-linguistic world state for embodied agents, source spans used as fine-grained attributions, structured graphs, and structured user profiles. Across these settings, the output is expected to be supported by the grounding input, and many formulations make provenance explicit through retrieved documents, highlighted spans, triplet indices, or selected frames (Prabhumoye et al., 2019, Zhang et al., 2021, Kim et al., 2023, Huang et al., 2023, Slobodkin et al., 2024, He et al., 24 Jan 2025, Dey et al., 13 Oct 2025).

1. Scope and historical development

An early explicit formulation of GTG appeared as Content Transfer for long-form text generation. In that setting, the task is to generate a next sentence in a document that both fits its context and is grounded in a content-rich external textual source such as a news story. The specific instantiation used curated Wikipedia context $C$, an external source document $S$, and a next sentence $y$, thereby framing GTG as control over content rather than only control over form such as style, persona, or politeness (Prabhumoye et al., 2019).

A distinct line of work defined GTG as a hybrid, stateful paradigm for task-oriented dialog systems. There, a large pre-trained Transformer is combined with explicit symbolic modules that encode dialog belief state and perform knowledge base inference, and responses are generated conditioned on dialog context, belief state, and deterministic knowledge base outputs. This formulation emphasizes interpretability, controllability, and reliability in task completion rather than only surface fluency (Gao et al., 2020).

A further expansion connected GTG to large-scale retrieval. RetGen placed grounded generation inside Information-aware Text Generation, where a prompt $x$ may leverage relevant documents from a large collection $Z$ to improve the next text $y$. Its central contribution was to jointly train a dense retriever and a grounded generator using LLM signal, thereby reducing reliance on oracle document supervision (Zhang et al., 2021).

Later work generalized GTG across modalities. SeViT treated a video as an external data store and combined a non-parametric frame retriever with a parametric generator for long untrimmed videos, while Grounded Decoding formulated token-level composition between an LLM and grounded models conditioned on non-linguistic state for embodied agents (Kim et al., 2023, Huang et al., 2023). In visually-grounded dialogue, GTG was used to generate definite descriptions that summarize coreferential information found in dialogue context before zero-shot grounding with a frozen vision-LLM (Willemsen et al., 2023).

Recent work has also sharpened the attributional and user-facing aspects of GTG. “Attribute First, then Generate” decomposes generation into content selection, sentence planning, and sequential sentence generation so that selected source spans become sentence-level attributions, while LAQuer introduces user-initiated, span-level attribution queries over generated text (Slobodkin et al., 2024, Hirsch et al., 1 Jun 2025). In parallel, GTG has been extended to graph-to-text with explicit reordering and attribution subtasks, to plan-based retrieval for long-form factual generation, to profile-grounded personalization, and to lightweight out-of-context detection in RAG deployments (He et al., 24 Jan 2025, Godbole et al., 2024, Dey et al., 13 Oct 2025, Poey et al., 2024).

2. Formal problem formulations

Despite the diversity of substrates, GTG formulations typically optimize a conditional distribution over text given grounding inputs. In Content Transfer, the model maximizes the likelihood of the next sentence conditioned on both local context and an external source:

$$\max_{\theta}\ \sum_{t=1}^{|y|} \log P_{\theta}\!\left(y_t \,\middle|\, y_{<t}, C, S\right).$$

This formulation makes coherence with $C$ and support from $S$ simultaneous requirements of the decoder (Prabhumoye et al., 2019).

In stateful task-oriented dialog, the grounded response is conditioned on dialog context $x$, belief state $S$0, and KB facts $S$1:

$S$2

The same framework also factorizes a dialog turn as $S$3, with the KB step deterministic given $S$4 (Gao et al., 2020).

In retrieval-grounded generation, RetGen models

$S$5

where the retriever assigns $S$6 and the generator models $S$7. This mixture over documents is approximated over the top-$S$8 retrieved candidates, and the retriever is updated from generator-derived signal rather than oracle retrieval labels (Zhang et al., 2021).

Grounded Decoding takes a different route by composing probabilities during decoding. Let $S$9 be a token sequence, $y$0 the language prefix, and $y$1 the non-linguistic world state. Then

$y$2

with multiple grounded models composed multiplicatively inside $y$3. This turns affordances, safety constraints, and preferences into token-level decoding factors without fine-tuning the LLM (Huang et al., 2023).

Video-grounded GTG in SeViT explicitly separates retrieval and generation. A video $y$4 is treated as an external data store; a retriever $y$5 returns a subset $y$6 of query-relevant frames; and a generator $y$7 produces an output sequence. The retriever computes cosine similarity between query and frame embeddings and applies top-$y$8 retrieval, while the generator uses standard auto-regressive generation:

$y$9

Late fusion is then implemented either by marginalization or Fusion-in-Decoder (Kim et al., 2023).

Other GTG settings instantiate the same conditional pattern with different grounding objects. In visually-grounded dialogue, a causal LLM generates a description $x$0 from dialogue context and a marked mention, and a frozen VLM selects the referent via $x$1 (Willemsen et al., 2023). In graph-to-text, the objective is $x$2 under explicit constraints to use only information from the provided triplets and include all of the information in the triplets, with PlanGTG additionally supervising a triplet permutation and attribution markers (He et al., 24 Jan 2025). In profile-grounded personalization, a model generates a personalized description $x$3 conditioned on a structured profile $x$4 and a generic item description $x$5 (Dey et al., 13 Oct 2025).

3. Grounding substrates and task families

A large portion of GTG is text-grounded. Content Transfer uses free-form source documents to support context-appropriate next-sentence generation in Wikipedia editing, while RetGen uses a large Wikipedia index to improve prose and dialogue generation without oracle document supervision (Prabhumoye et al., 2019, Zhang et al., 2021). More recent attributed-generation work targets multi-document summarization and long-form question answering, where the grounding input is either user-provided or retrieved source text and the output must be attributable at the sentence or span level (Slobodkin et al., 2024, Hirsch et al., 1 Jun 2025, Godbole et al., 2024).

A second substrate is symbolic state and structured knowledge. In task-oriented dialog, the grounding object is a belief state represented as domain-slot-value tuples plus deterministic KB results, and generation is typically delexicalized and later lexicalized with retrieved facts (Gao et al., 2020). In graph-to-text, the grounding object is a graph linearized as RDF-like triples $x$6, and groundedness is tied to factual coverage, lack of hallucination, and explicit attribution from output spans to triplet indices (He et al., 24 Jan 2025).

Video-grounded GTG treats frames or segments as the evidence substrate. Representative tasks include video captioning, video-based visual question answering, iVQA, and Next-QA. In this setting, “grounded” means the generated text can be traced back to visual evidence, and SeViT operationalizes that idea by selecting a small set of query-relevant frames from long videos rather than attending over all frames (Kim et al., 2023).

Embodied GTG grounds text in non-linguistic observations and downstream executability. Grounded Decoding combines LLM token probabilities with grounded model probabilities derived from affordance functions, detector scores, safety rules, and preference constraints, producing plans or skill sequences that are both semantically likely and realizable in the current environment state (Huang et al., 2023). This differs from document-grounded GTG because the grounding signal is not merely evidential; it is also feasibility- and reward-bearing.

Visually-grounded dialogue introduces a discourse-heavy substrate. Here the grounding challenge is not only mapping text to images but first resolving elliptical, pronominal, and distributed references across multiple turns. The proposed solution is to generate a definite description that summarizes the referential chain and only then use a pretrained VLM for zero-shot retrieval (Willemsen et al., 2023).

Profile-grounded GTG uses user structure as the grounding object. GRAVITY constructs profiles from demographics, product interests, intrinsic values and beliefs, and personality traits, and uses profile-grounded synthetic preference pairs to tune personalized generation. This extends GTG beyond factual support and toward alignment with user-specific values, interests, beliefs, and personality (Dey et al., 13 Oct 2025).

4. Modeling patterns and system architectures

One recurring pattern is retrieval-plus-generation with explicit separation between non-parametric memory and parametric decoding. RetGen jointly trains a dense retriever and a grounded generator, scores retrieved documents via inner product, and combines multiple documents with a Mixture-of-Experts ensemble at the output distribution. SeViT uses the same high-level decomposition in video: a non-parametric retriever selects query-relevant frames from an external store, and a parametric generator aggregates them with late fusion through MAR or FiD. In both cases, grounding is concentrated into a small, high-utility subset rather than the full evidence pool (Zhang et al., 2021, Kim et al., 2023).

A second pattern is decode-time composition rather than end-to-end fusion. Grounded Decoding multiplies LLM next-token probabilities by grounded probabilities computed from affordance models, value functions, detectors, safety constraints, and preference functions. This formulation resembles probabilistic filtering and makes groundedness a token-level property of the decoding rule rather than only a property of encoder conditioning (Huang et al., 2023).

A third pattern is planning-first generation. In graph-to-text, PlanGTG instructs the model first to output the “correct triplet order,” then to generate text with sequence numbers that attribute facts to triplets. In plan-based retrieval for long-form factual generation, the model first produces an outline, then facet-specific questions, then evidence retrieval, and only then writes the final output using only the retrieved evidence. This suggests that planning is used both as a coherence device and as a mechanism for improving evidence coverage (He et al., 24 Jan 2025, Godbole et al., 2024).

A fourth pattern is select-then-generate for local attribution. “Attribute First, then Generate” breaks GTG into content selection, sentence planning, and sequential sentence generation. Selected source spans become the conditioning context for one sentence and also the sentence’s attribution, so “select” becomes “attribute.” LAQuer extends this family by adding a post-generation, user-directed layer: a user selects output spans, the system decontextualizes their meaning, and then returns minimal supporting source spans, either with LLM prompting or with hidden-state alignment (Slobodkin et al., 2024, Hirsch et al., 1 Jun 2025).

A fifth pattern is modular interface design between discourse processing and grounding back ends. In visually-grounded dialogue, the causal LLM is not used to directly ground against images; instead, it generates definite descriptions that summarize coreferential information, and a frozen VLM performs the actual visual retrieval. This preserves an interpretable intermediate object—the generated description—while exploiting zero-shot VLM alignment (Willemsen et al., 2023).

A sixth pattern is neuro-symbolic grounding. Task-oriented GTG uses symbolic belief tracking, deterministic KB inference, action masks, templates, and lexicalization, while still relying on a shared Transformer backbone for generation. RAGulator, although not a generator, occupies an adjacent systems position: it inserts a lightweight sentence-level discriminator after generation to detect whether candidate sentences are semantically out-of-context with respect to retrieved documents (Gao et al., 2020, Poey et al., 2024).

5. Attribution, evaluation, and empirical findings

GTG evaluation is notably heterogeneous because different task families expose different notions of support. In Content Transfer, automatic evaluation uses ROUGE-L F1, BLEU, and METEOR, and human evaluation follows DUC 2007 quality guidelines plus pairwise judgments for “close to reference” and “coherent to context.” The same work explicitly notes that grounding-specific automatic attribution or entailment metrics are not reported, which underscores an early gap between text quality evaluation and support verification (Prabhumoye et al., 2019).

Retrieval-grounded generation has often combined text quality, diversity, and grounding proxies. RetGen reports NIST, BLEU, METEOR, diversity metrics such as Entropy and Dist-$x$7, and a groundedness proxy called Keyword Matching Ratio. It also includes pairwise human judgments for coherence, informativeness, and human-likeness, showing that retrieval and generation can improve together under joint training (Zhang et al., 2021).

In long-video GTG, the empirical case for retrieval is strongest on longer and causally complex inputs. SeViT reports new state of the art on four video-language datasets: iVQA (+4.8), Next-QA (+6.9), and Activitynet-QA (+4.8) in accuracy, and MSRVTT-Caption (+3.6) in CIDEr. The paper further reports that retrieval improves over uniform sampling especially on longer subsets, and that qualitative examples on 10-minute Activitynet videos show retrieval focusing on query-relevant segments while uniform sampling misses them (Kim et al., 2023).

For embodied planning, evaluation targets task success as much as linguistic plausibility. Grounded Decoding reports average success rates on simulated tabletop rearrangement, Minigrid 2D Maze, and real-world kitchen mobile manipulation, and compares greedy and beam decoding against ungrounded LLM planning and direct policy baselines. It also reports a token-efficiency comparison against SayCan, where Token Count relative to GD Greedy is 1× for GD Greedy, 4.3× for GD Beam, and 113.7× for SayCan, while maintaining comparable success. This frames grounding not only as a fidelity mechanism but also as a computational strategy in combinatorial action spaces (Huang et al., 2023).

In visually-grounded dialogue, evaluation is split between description generation and referent retrieval. The reported retrieval metrics are top-1 accuracy, MRR, and NDCG. On the AGOS benchmark, GPT-3 CRDG with BLIP-2 reaches .69 accuracy at $x$8 and .71 at $x$9, while ground-truth manual descriptions combined with larger context windows reach approximately .83 with full context. The same study reports that larger windows tend to help retrieval even when text-generation metrics slightly decrease (Willemsen et al., 2023).

Locally attributable GTG has introduced attribution-specific metrics that directly operationalize verification effort. “Attribute First, then Generate” uses AUTOAIS, average cited length per sentence (LEN), and the percentage of unattributed sentences (NO-ATT), alongside ROUGE-L and BERTScore. On long-form question answering with in-context learning, ATTR.FIRST-CoT reports AUTOAIS 89.3 with average citation length 48.2 tokens per sentence, versus ALCE’s citation length 2153.3, and human assessors require roughly 50% less verification time than with ALCE (Slobodkin et al., 2024). LAQuer pushes this further by evaluating user-selected sub-sentence facts. It reports substantial reductions in attributed text length, including about 54% fewer characters than sentence-level spans on SPARK and two orders of magnitude reductions relative to document-level attribution in Vanilla and ALCE settings, although decontextualized fact attribution remains difficult (Hirsch et al., 1 Jun 2025).

Graph-to-text and personalization bring yet other evaluation regimes. PlanGTG uses BLEU-4, METEOR, CHRF++, BartScore, and human counts of hallucinated and missed entities and relations, and reports strong zero-shot transfer gains after PlanGTG fine-tuning, such as WebNLG17 BLEU-4 17.32 → 28.76 for LLaMA2-7b-chat and 16.60 → 29.76 for Mistral-7b-chat (He et al., 24 Jan 2025). GRAVITY evaluates personalized book description generation with Top-1 WinRate, Preference Gain, Interestingness Score, and SentenceBERT similarity, and reports overall Preference Gain 82.5% in auto-evaluation and 86.73% in a user study (Dey et al., 13 Oct 2025). At the systems boundary, RAGulator evaluates sentence-level grounding detection with AUROC, AUPRC, and F1, and reports that DeBERTa-v3-large reaches AUROC 98.2, AUPRC 98.3, and F1 93.5 on the overall evaluation set while running at 19.08 examples per second (Poey et al., 2024).

6. Limitations, trade-offs, and research directions

A persistent limitation across GTG is dependence on the quality of the grounding mechanism itself. SeViT explicitly notes dependence on retriever quality and the difficulty of joint FiD retriever training; Grounded Decoding depends on the quality of affordance models, detectors, and partial-command scoring; RAGulator shows clear degradation under domain shift; and LAQuer reports that decontextualized facts remain hard to localize even when generation metadata is available (Kim et al., 2023, Huang et al., 2023, Poey et al., 2024, Hirsch et al., 1 Jun 2025).

Another recurring trade-off concerns attribution granularity. Coarse citations maximize coverage but increase verification burden, whereas local spans reduce reading cost but risk partial support. “Attribute First, then Generate” states this directly as a conciseness-versus-coverage trade-off and also notes a fluency-versus-locality trade-off under strict local conditioning. LAQuer makes a related point from the user side: sentence-level attributions often force users to read irrelevant content, but sub-sentence attributions may fail if the necessary context for decontextualization is not captured (Slobodkin et al., 2024, Hirsch et al., 1 Jun 2025).

Planning improves coverage, but it also adds complexity and latency. Plan-based retrieval requires an outline, question generation, a second retrieval stage, and sometimes an additional QA stage before generation, and the paper emphasizes that this adds cost even as it improves attribution and informativeness. In graph-to-text, performance declines as the number of triplets increases, and the paper identifies planning on complex graphs as a major bottleneck. This suggests that planning is beneficial, but not free: it changes the operating point of GTG systems rather than eliminating the underlying difficulty (Godbole et al., 2024, He et al., 24 Jan 2025).

Bias, transfer, and privacy are also substrate-dependent concerns. SeViT notes that a CLIP-based retriever may inherit biases from image-text pretraining; Grounded Decoding describes grounded models as domain-specific and leaves general-purpose grounding open; GRAVITY notes that synthetic profile inference introduces uncertainty and that cultural instruments may not capture all nuances of cultural identity, while also not specifying explicit privacy safeguards. In profile-grounded GTG especially, the grounding signal is not merely factual but inferential, which makes auditing and consent central implementation issues (Kim et al., 2023, Huang et al., 2023, Dey et al., 13 Oct 2025).

Several research directions recur across the literature. These include end-to-end differentiable retrieval, improved retrieval objectives and hard negative mining, segment- or event-level retrieval for video, richer multimodal memory structures, explicit causal reasoning modules, better attribution metrics, curriculum learning for graph complexity, and dynamic planning that adapts retrieval granularity to the task (Kim et al., 2023, He et al., 24 Jan 2025, Godbole et al., 2024). In discourse-heavy settings, entity tracking and larger context windows remain promising; in attributed generation, interactive interfaces that allow context expansion around concise citations are a natural extension (Willemsen et al., 2023, Slobodkin et al., 2024).

Taken together, the literature presents GTG not as a single model class but as a family of conditioning strategies for making generation supportable, attributable, and operationally useful. The common structure is the same: isolate an external grounding substrate, expose it to the generator either before or during decoding, and evaluate not only text quality but also support, provenance, and user verification cost. This suggests that the most durable advances in GTG are likely to come from tighter coupling between generation and evidence management, rather than from larger parametric models alone.