PlanGEN: Multi-agent Planning & Generation in AI

Updated 17 July 2026

PlanGEN is a collection of planning-centric systems that use explicit intermediate objects, such as constraints, content plans, or spatial layouts, to guide complex reasoning and generation.
The multi-agent framework leverages constraint, verification, and selection agents to enable adaptive, instance-level algorithm selection and reward-based optimization.
PlanGEN also encompasses controlled data-to-text generation and unified vision-language models, achieving superior results on benchmarks compared to traditional methods.

PlanGEN refers to several planning-centric systems in contemporary arXiv literature rather than a single unified method. The label is used most directly for a model-agnostic, easily scalable multi-agent framework with constraint, verification, and selection agents for complex planning and reasoning (Parmar et al., 22 Feb 2025). Closely related uses include "Plan-then-Generate," a controlled data-to-text framework that separates content planning from surface realization (Su et al., 2021), and a unified layout planning and image generation model in auto-regressive vision-LLMs (He et al., 13 Mar 2025). Across these works, planning is treated as an explicit intermediate object—constraints, content plans, or spatial layouts—rather than as an implicit by-product of generation.

1. Scope and disambiguation

In the cited literature, the name appears in multiple technically distinct systems. Capitalization varies by paper, and the underlying task changes substantially.

Name in paper	Domain	Defining characterization
PlanGEN	Complex planning and reasoning	Multi-agent framework with constraint, verification, and selection agents (Parmar et al., 22 Feb 2025)
PlanGen	Data-to-text generation	Plan-then-generate pipeline with explicit content plans (Su et al., 2021)
PlanGen	Vision-language generation	Unified layout planning and image generation in one autoregressive transformer (He et al., 13 Mar 2025)

This multiplicity is important because "PlanGEN" is not a standardized acronym tied to a single benchmark family or implementation. In one line of work, it denotes iterative verification and adaptive inference-time algorithm selection for natural-language planning and reasoning (Parmar et al., 22 Feb 2025). In another, it denotes explicit content planning for ToTTo and WebNLG (Su et al., 2021). In a third, it denotes spatial layout planning as a precursor to image synthesis and manipulation (He et al., 13 Mar 2025). The shared theme is explicit plan formation, but the plan object itself differs: reasoning trajectories, ordered content tokens, or layout tokens.

2. Multi-agent PlanGEN for complex problem solving

The 2025 PlanGEN framework is presented as a model-agnostic and easily scalable agent framework with three key components: constraint, verification, and selection agents (Parmar et al., 22 Feb 2025). Its stated purpose is to improve the planning and reasoning capabilities of LLMs on complex problem-solving tasks, especially where prior methods either verify only at the task level without considering constraints or apply inference-time algorithms without adapting to instance-level complexity.

The Constraint Agent extracts instance-specific constraints from the problem statement. The examples given in the source summary include participant availabilities in calendar scheduling and formula correctness in mathematical problems. The Verification Agent evaluates candidate plans against those constraints, produces detailed feedback, and assigns a numerical reward. The Selection Agent decides which inference-time algorithm to apply next, using a modified Upper Confidence Bound policy enriched with an LLM prior, diversity bonus, and recovery bonus.

PlanGEN is defined in several operating modes. "PlanGEN (Best of N)" augments sampling-based selection with agent-guided reward selection; "PlanGEN (ToT)" replaces generic node evaluation in Tree-of-Thought with constraint-aware verification; "PlanGEN (REBASE)" uses agent-derived rewards to guide search and pruning; and "PlanGEN (Mixture of Algorithms)" dynamically selects among these algorithms for each instance (Parmar et al., 22 Feb 2025). The interaction loop is iterative: constraint extraction, candidate plan generation, verification, possible algorithm switching or refinement, and termination when a sufficiently high reward is reached or the computational budget is exhausted.

3. Verification, reward, and adaptive selection

A central feature of PlanGEN is its constraint-guided iterative verification loop. The Verification Agent scores plans on a range from -100 to 100, and the typical threshold for acceptance is $\geq 95$ (Parmar et al., 22 Feb 2025). The paper summary reports that the agent produces both natural-language feedback and a numerical reward, and that logistic regression shows a direct positive relationship between reward score and success probability. This makes the reward signal both an optimization primitive and an interpretability device.

The Selection Agent is described as using a modified UCB policy whose components include average historical reward per algorithm, an exploration bonus, an LLM-prior-score for expected suitability, a diversity bonus, and a recovery bonus (Parmar et al., 22 Feb 2025). The prior is weighted heavily at the start and decayed over iterations or trials. This is intended to balance exploitation, exploration, and adaptation to instance-level complexity rather than committing to a single fixed inference-time scaffold.

PlanGEN’s integration with existing inference-time algorithms is explicit. For Best of N, the Verification Agent replaces simple post-hoc ranking or self-evaluation. For Tree-of-Thought, verified constraints are used to value partial paths and prune poor branches. For REBASE, the Verification Agent supplies the reward function that prioritizes subtrees with better verified rewards (Parmar et al., 22 Feb 2025). This suggests that PlanGEN should be understood less as a standalone planner and more as a supervisory layer over inference-time search and sampling procedures.

4. Benchmarks and reported performance

PlanGEN is evaluated on NATURAL PLAN, OlympiadBench, GPQA, and DocFinQA, using Exact Match, Micro-average Accuracy, Accuracy, and Accuracy/F1-score, respectively (Parmar et al., 22 Feb 2025). The reported models are primarily Gemini-1.5-Pro, with case studies on Gemini-2.0-Flash and GPT-4o to support model-agnosticism.

Benchmark	Reported PlanGEN result	Context
NATURAL PLAN	Calendar EM 60.70%	Highest EM; about +8% over baseline (Parmar et al., 22 Feb 2025)
OlympiadBench	Math accuracy 55.94%	Mixture achieves highest math accuracy; about +5% (Parmar et al., 22 Feb 2025)
GPQA	Accuracy 59.6%	Mixture shows strong lead; +13%/+12% vs. baselines (Parmar et al., 22 Feb 2025)
DocFinQA	Accuracy 31.16%, F1 29.45%	Improvement of about 7% (Parmar et al., 22 Feb 2025)

The source summary further states that multi-agent PlanGEN frameworks consistently outperformed both zero-shot and vanilla multi-agent baselines. For simpler problems, PlanGEN (ToT) or PlanGEN (Best of N) were sufficient, whereas for more complex problems—explicitly including OlympiadBench, GPQA, and complex scheduling—the Mixture/Adaptive approach outperformed fixed algorithms (Parmar et al., 22 Feb 2025). The same summary reports favorable scalability-efficiency tradeoffs: state-of-the-art performance is achieved with only modestly increased computation over baselines.

A recurrent misconception is that PlanGEN’s gains arise solely from sampling more trajectories. The reported ablations instead emphasize two separate effects: constraint-guided verification improves inference-time algorithms, and adaptive selection provides an additional boost on complex planning and reasoning problems (Parmar et al., 22 Feb 2025).

5. PlanGen in controlled data-to-text generation

In "Plan-then-Generate: Controlled Data-to-Text Generation via Planning," PlanGen is a two-stage, pipeline-based framework for controllability and diversity in neural data-to-text generation (Su et al., 2021). The planning phase predicts a Content Plan ( $C$ ), defined as an ordered list of tokens from the input data specifying what information to include and in which sequence it should appear. The generation phase then realizes text from the concatenated representation $[T:C]$ , where $T$ is the structured input.

The planner uses a pre-trained BERT-base content encoder, a selection module for slot keys or predicates, and an ordering predictor implemented as a linear-chain Conditional Random Field. The generator is BART-base, trained first with maximum likelihood estimation and then with a structure-aware reinforcement learning objective that combines BLEU between generated and reference text with BLEU between predicted and reference content plans (Su et al., 2021). Users may also directly specify content plans, which gives full structural control over both intra-sentence and inter-sentence organization.

Evaluation is performed on ToTTo and WebNLG. On ToTTo, the summary reports that PlanGen with 140M parameters matches or outperforms state-of-the-art systems such as T5-3B on BLEU, PARENT, and BLEURT, with an example on the Non-Overlap split of BLEU 41.5 vs. 41.4 and PARENT 54.6 vs. 54.2. On WebNLG, it reportedly outperforms previous systems on both BLEU and METEOR, with overall BLEU 60.51 compared with 59.70 for T5+Prefix, and achieves the best planning metrics with planning accuracy 0.72 vs. 0.62 and BLEU-2 84.97 vs. 80.10 (Su et al., 2021).

The paper also emphasizes controllability and diversity. Human evaluation is summarized as showing that PlanGen follows user-supplied or predicted content plans very accurately, with an accuracy ~1.74/2. When content plans are shuffled, Self-BLEU reportedly drops from 100 to ~26, indicating structural diversity that decoding changes alone did not achieve in end-to-end baselines (Su et al., 2021). In this usage, PlanGen denotes explicit symbolic or semi-symbolic planning inside neural generation, not multi-agent verification.

6. PlanGen in auto-regressive vision-LLMs

In "PlanGen: Towards Unified Layout Planning and Image Generation in Auto-Regressive Vision LLMs," PlanGen denotes a unified model for spatial layout planning and layout-to-image generation (He et al., 13 Mar 2025). The paper states that, unlike previous diffusion-based models that treat layout planning and layout-to-image as two separate models, this PlanGen jointly models both tasks in one autoregressive transformer using only next-token prediction.

The model ingests text prompt tokens $\mathbf{t}$ , layout tokens $\mathbf{l}$ , and image tokens $\mathbf{g}$ . Layout conditions are represented as ordered context tokens rather than compressed by embed-and-pool operations. Prompting uses tokens such as <grounding>, <image_start>, <ref>...</ref>, and <box>...</box>, with bounding boxes standardized to [0,1000] scale (He et al., 13 Mar 2025). The multitask objective combines layout planning, layout-to-image generation, and image layout understanding: $\mathcal{L} = \alpha \mathcal{L}_{LP} + \beta \mathcal{L}_{LIG} + \gamma \mathcal{L}_{ILU},$ with empirical balance coefficients $\alpha = \beta = \gamma = 1$ .

This PlanGen also supports layout-guided image manipulation. The summary highlights teacher-forcing content manipulation, in which edited regions are sampled while non-edited regions copy original image tokens, and negative layout guidance, which uses negative layout tokens and a guidance scale $s$ , described as typically 5, to suppress undesired content during deletion or editing (He et al., 13 Mar 2025).

Reported results span multiple layout-related tasks. For layout planning, images generated after layout planning are reported to obtain FID 50.7. For layout-to-image generation, PlanGen reportedly achieves the best FID 13.9 and outperforms diffusion-based methods such as InstanceDiff, CreatiLayout, GLIGEN, HiCo, MIGC, and Ranni on region-wise and global image quality metrics. For image layout understanding, it is summarized as nearing the performance of specialized detection models such as Grounding-DINO and outperforming Qwen-VL-Chat and CogVLM-grounding, with an example of AP50 44.8 vs. 9.2 for Qwen-VL-Chat. For image manipulation, it reports object removal success 77.5% vs. 66.5% for SD-Inpainting (He et al., 13 Mar 2025). In this variant, planning is a spatial scaffold inside a multimodal generative transformer.

7. Later comparisons, extensions, and contested readings

Later papers use "PlanGEN" both as a baseline family and as a foil for new execution or generalized planning systems. In "RunAgent: Interpreting Natural-Language Plans with Constraint-Guided Execution," RunAgent is evaluated against baseline LLMs and state-of-the-art PlanGEN methods on Natural-plan and SciBench (Srivastava et al., 1 May 2026). On Natural-plan Calendar Scheduling, the summary reports EM Acc. 68.9 for PlanGEN (Best-of-N, Gemini 2.0 Flash) and 81.1 for RunAgent; an ablation without RunAgent’s constraint checking gives 75.4. On Natural-plan Trip Planning, the same summary reports 41.63 for PlanGEN (Best-of-N, Gemini-1.5) and 14.73 for RunAgent, while explicitly noting that the table seems to have an inversion (Srivastava et al., 1 May 2026). That caveat is significant: it indicates that benchmark narratives around "PlanGEN" require close reading of task definitions and reporting conventions.

A separate development appears in generalized PDDL planning. "GenePlan: Evolving Better Generalized PDDL Plans using LLMs" describes prior PlanGEN/Silver et al (2024) methods as CoT-based extraction of domainwide Python policies or generalized planners that are satisficing rather than optimized for plan length or cost (Murray et al., 10 Mar 2026). GenePlan positions itself as an optimization-enhanced successor, reporting average SAT score 0.91, compared with 0.93 for state-of-the-art planners and 0.64 for CoT prompting, while solving new instances in 0.49 seconds per task at $1.82 per domain using GPT-4o (Murray et al., 10 Mar 2026). This suggests that, in yet another subcommunity, "PlanGEN" is associated with generalized planner induction rather than natural-language reasoning or multimodal generation.

Taken together, these later references show that PlanGEN has become a family name for methods that externalize plan structure and use it to control downstream execution, generation, or search. The precise meaning, however, remains paper-specific: multi-agent constraint-guided reasoning (Parmar et al., 22 Feb 2025), explicit content planning for data-to-text (Su et al., 2021), autoregressive layout planning for image synthesis (He et al., 13 Mar 2025), or generalized planner extraction as discussed by successor systems (Murray et al., 10 Mar 2026).