DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation

TL;DR

DeepPresenter is a dual-agent framework (Researcher + Presenter) for end-to-end slide-deck generation that abandons predefined workflows and template-filling in favor of free-form HTML generation grounded in environmental observation. The key mechanism is environment-grounded reflection: rather than self-reflecting over reasoning traces or HTML source, the agent renders intermediate artifacts to pixels via an inspect tool and reasons over the rendered output, catching post-render defects (overflow, low contrast, broken layouts) that introspection misses. A 9B model fine-tuned on 802 filtered trajectories (synthesized with an extrinsic critic to break self-verification bias) approaches GPT-5 performance on a 128-task evaluation while costing substantially less per run.

Key claims

Self-reflection over internal signals (HTML source, reasoning traces) fails to detect rendering defects because agent observation and user perception operate in different state spaces; grounding reflection in rendered pixels via an inspect tool closes the gap [§2.3].
A two-agent decomposition (Researcher autonomously plans tool-augmented exploration; Presenter generates HTML slides from scratch with content-aware styling) outperforms single-agent and template-based baselines on Content, Style, and Diversity [§2.2, Tab. 2].
With Gemini-3-Pro as backbone, DeepPresenter scores 4.44 average vs. Gamma (commercial) 4.36 and best open-source baseline 3.92, and reaches Vendi-Score diversity 0.79 vs. 0.17–0.52 for template-based competitors [Tab. 2].
Extrinsic verification (independent critic conditioned on the rendered observation, separate from the generating trajectory) yields a 67% larger fine-tuning gain than SFT alone (+0.20 vs +0.12 Avg.) by mitigating self-verification bias during trajectory synthesis [Tab. 4, §5.1].
DeepPresenter-9B (fine-tuned GLM-4.6V-Flash) reaches 4.19 average — outperforming all open-source baselines and approaching GPT-5 (4.22) — with only 802 trajectories surviving the three-stage filter (rule-based constraint check, GLM-4.6 consistency check, GLM-4.6V quality check) [§3.3, Tab. 2].
Ablating environment-grounded reflection drops the 9B from 4.19 to 3.82; ablating dual-agent decomposition drops it to 3.23; ablating trajectory filtering drops it to 4.03 [Tab. 3].
Extrinsic verification detects 308 vs. 212 layout defects and 101 vs. 43 render defects over the same 300 trajectories compared to self-verification, indicating systematic self-rationalization in the latter regime [§5.1, Fig. 4].
Free-form generation is brittle: 43.0% of pre-filter trajectory failures are quality errors and 32.3% are environment failures (context overflow, infrastructure disruptions); constraint and consistency errors are 13.5% and 11.2% [§5.2, Fig. 5].
Backbone capability matters disproportionately for DeepPresenter (3.82 → 4.44 across GLM-4.6V-Flash → Gemini-3-Pro), whereas template-based baselines are flat across backbones, indicating fixed pipelines bottleneck stronger models [§5.3].

Method

DeepPresenter splits the task into two sequentially-executed agents sharing the same backbone. The Researcher invokes tools to retrieve and synthesize materials and persists a structured markdown manuscript with assets to a shared file system; depth and strategy adapt to the user intent (technical surveys vs. accessible illustrations). The Presenter consumes the manuscript, develops a global design plan (color theme, typography), then generates each slide as a standalone HTML file. Crucially, both agents have access to an inspect tool: for the Presenter, inspect rasterizes HTML to image pixels; for the Researcher, it returns structured diagnostics over the file system. Reflection (think) is then conditioned on the rendered/diagnostic observation, forming an observe–reflect–revise loop where the agent’s observation space matches the user’s perceptual state.

For the distilled DeepPresenter-9B, training trajectories are generated by running the framework with Gemini-3-Pro as both backbone and critic on 1,024 tasks (drawn from PersonaHub, arXiv, FinePDFs-Edu, augmented with verifiable constraints over slide count, language, aspect ratio). The trajectory synthesis adds an extrinsic verification step: after each inspect, an independent critic — run in a separate context from the agent’s own trajectory state — judges the artifact, identifies defects, and emits a reasoning trace specifying revisions; this trace is then appended to the agent context as a think call. A three-stage filter (rule-based constraint check, GLM-4.6 consistency check on reflection–action alignment, GLM-4.6V critical-defect check) reduces the 1,024 rollouts to 802 SFT samples. Fine-tuning runs MS-SWIFT on GLM-4.6V-Flash with batch 32, LR 1e-5, 5 epochs, ~80 GPU-hours on 8 A800.

Results

On 128 held-out tasks scored across Constraint (rule-based), Content + Style (GPT-5 MLLM judge from PPTAgent), and Vendi-Score diversity (DINOv2 features), DeepPresenter with Gemini-3-Pro reaches 4.44 average (Constraint 4.70 / Content 4.25 / Style 4.37, Diversity 0.79) vs. Gamma 4.36 (Diversity 0.52) and the best open-source baseline KCTV + Claude-Sonnet-4.5 at 3.92 (Diversity 0.20) [Tab. 2]. The Diversity gap is structural: templated baselines (PPTAgent, KCTV) sit at 0.17–0.35 regardless of backbone. DeepPresenter-9B reaches 4.19 (vs. GPT-5 backbone of DeepPresenter at 4.22) for substantially lower per-run cost; it also exceeds its GLM-4.6V-Flash base by +0.37 average. Ablations: w/o inspect (no grounded reflection) costs −0.12 on Gemini-3-Pro and −0.37 on the 9B; w/o dual-agent costs −0.40 and −0.96 respectively; w/o filtering costs −0.16 on the 9B [Tab. 3]. Extrinsic-verification ablation on a 300-trajectory subset: +0.20 over base vs. +0.12 with SFT-on-self-verified trajectories [Tab. 4].

Why it’s interesting

DeepPresenter sits at the intersection of two threads on the wiki. First, it extends Tool-Use Agents with a visual observation channel — the inspect tool is essentially the same multi-turn (instruction, observation, action, feedback) loop catalogued for VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents, but where VisGym uses image observations to test current VLMs and finds visual grounding is the bottleneck, DeepPresenter operationalizes the same channel for generation and reports that rendered-pixel feedback is what catches the dominant failure modes (layout, overflow, low contrast) that text-only introspection misses. Second, the extrinsic-verification recipe is the cleanest filed datapoint on a thread implicit in Synthetic Training Data: the verification gate dominates trajectory-synthesis returns, and a critic that is decoupled from the generating trajectory state outperforms self-verification even when the agent already has access to environmental observations. The 9B’s +0.20 vs. +0.12 result is a sharper version of the same effect that Action100M’s Self-Refine and Video-Thinker’s hindsight curation report from adjacent recipes (Action100M: A Large-scale Video Action Dataset, Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning). The contrast with MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling is also instructive: MiroThinker scales interaction depth (more tool calls) while DeepPresenter scales observation modality (text → rendered pixels) with relatively shallow rollouts.