DeepNews: Workflow for Deep Financial Reporting

• DeepNews Framework is an agentic workflow designed to overcome the 'Statistical Smoothing Trap' by integrating evidence retrieval, schema-guided planning, and adversarial constraints.
• It uses a tri-stream retrieval strategy and a strict 10:1 evidence-to-output ratio to ensure deep factual grounding and maintain logical coherence in financial journalism.
• Empirical evaluations show DeepNews delivers data-rich, structurally nuanced narratives with improved acceptance rates compared to conventional LLM outputs.

Searching arXiv for the cited DeepNews-related papers and adjacent work to ground the article. DeepNews Framework is an agentic workflow for long-form generation in vertical domains, introduced for deep financial reporting as a system that models information acquisition, strategic planning, and constrained realization rather than relying on direct next-token continuation alone. It is explicitly motivated by the “impossible trinity” of low hallucination, deep logical coherence, and personalized expression, and by the claim that ordinary LLM generation falls into a “Statistical Smoothing Trap,” producing prose that is fluent yet over-smoothed, structurally average, and vulnerable to factual fabrication under sparse evidence (Jiang, 10 Dec 2025).

1. Conceptual foundation

DeepNews is defined against a specific failure model of contemporary LLMs. The framework argues that RLHF-trained autoregressive systems are incentivized to produce average, safe, and high-probability continuations, which yields grammatically smooth text but suppresses the burstiness, asymmetry, and hard factual grounding characteristic of expert financial journalism. In that formulation, the central problem is not only hallucination, but the joint collapse of factual precision, causal sharpness, and stylistic distinctiveness into a probabilistically smoothed median register (Jiang, 10 Dec 2025).

The paper formalizes the framework’s response as a constrained objective:

$$\mathcal{L}_{DeepNews}(x) = -\log P(x|C) + \lambda_1 \cdot \mathcal{H}_{Hallucination}(x, S) - \lambda_2 \cdot \mathcal{B}_{Burstiness}(x)$$

where $x$ is the generated text sequence, $C$ is the retrieved context, $S$ is the schema structure, $\mathcal{H}_{Hallucination}(x,S)$ is a hallucination penalty, and $\mathcal{B}_{Burstiness}(x)$ is a burstiness reward. The burstiness term is defined as

$$\mathcal{B}(x) = \sigma(\mathrm{Var}(\mathrm{len}(s_i)))$$

with $s_i$ denoting the sentences in passage $x$. The intent is not to maximize stylistic irregularity in isolation, but to force the generator away from local optima built around average sentence length and uniform discourse pacing.

This makes DeepNews distinct from a conventional prompting strategy. It treats long-form news writing as entropy reduction from a large, redundant evidence base into a low-entropy argument, and therefore frames retrieval, planning, and realization as separate computational stages rather than one prompt-response exchange.

2. Information foraging and retrieval regime

The first major module is a dual-granularity retrieval mechanism grounded in information foraging theory. Retrieval is not minimal-answer retrieval; it is explicit evidence saturation. DeepNews uses a tri-stream information foraging design consisting of an Ecological Stream, a Quantitative Stream, and a Narrative Stream. The Ecological Stream retrieves supply-chain context, upstream and downstream partners, and competitors. The Quantitative Stream retrieves financial reports, research reports, and macro indicators. The Narrative Stream retrieves event details, key statements, and conflicts. The framework’s claim is that these streams are orthogonal enough to support triangulation rather than single-source narrative collapse (Jiang, 10 Dec 2025).

Retrieved material is then reorganized into two granularities. “Atomic Facts” preserve micro-structure such as dates, prices, named entities, and local claims. “Context Blocks” preserve macro-structure such as market trends, causal background, and strategic setting. This separation is motivated by the claim that expert comprehension requires both local propositions and global gist, rather than one undifferentiated context dump.

A central operational rule is the 10:1 saturated information input ratio, described as the “cost of truth” or a “Cognitive Tax.” The framework argues that high-fidelity financial article generation requires input context length to output length of approximately $10:1$, and it operationally insists on retrieval packs exceeding $x$0 characters for a typical deep report. This threshold is justified through the paper’s “Knowledge Cliff” analysis rather than through a learned stopping criterion.

The same study varies retrieved context over $x$1, $x$2, $x$3, $x$4, and $x$5 characters. It reports a “Noise Zone” below $x$6 characters with Hallucination-Free Rate below $x$7, a “Collapse Point” at $x$8 characters with HFR about $x$9, a “Phase Transition Zone” at $C$0 characters where HFR jumps to $C$1, and a “Saturation Zone” at $C$2 characters where HFR rises to $C$3 with diminishing returns. HFR is defined verbally as the percentage of articles in which key data, factual statements, and causal attributions remain free from unverifiable or erroneous information. This establishes the framework’s Minimum Viable Context at roughly $C$4 characters (Jiang, 10 Dec 2025).

3. Schema-guided planning and Atomic Blocks

After retrieval, DeepNews transitions to schema-guided strategic planning. The framework uses a domain expert knowledge base called the DeepNews Financial Ontology, or DNFO-v5. DNFO-v5 is described as the fifth iteration of a distilled expert framework containing 5 root narrative classes, 19 sub-scenarios or leaf nodes, and more than 5,000 combination pathways. The five root classes are S1-BLIND, S2-VGAME, S3-SINGLE, S4-HGAME, and S5-INDUS (Jiang, 10 Dec 2025).

These schemas are not generic templates. They are domain-specific narrative programs encoding the editorial logic of seasoned financial journalists. In the “Vertical Market Game” schema, for example, the structure includes company pressure, counterpart response, transmission path, variable injection, game focus, and endgame scenarios. This indicates that schema selection functions as logical slot filling over an already saturated evidence base.

The planner operates in three layers. The first is Macro-Schema Injection, which detects the news theme, retrieves the matching DNFO-v5 schema, and fills scenario-specific slots. The second is a Narrative Orchestrator, which selects among six lede blueprints such as “Contradiction/Paradox” and “Dramatic Opening,” and imposes pacing variation across adjacent sections. The third is the Atomic Block System, which decomposes each section into standardized functional units (Jiang, 10 Dec 2025).

Atomic Blocks are defined as meso-structural units with a single communicative purpose. The appendix lists Data Anchor Block, Narrative Cut-in Block, Deep Insight Block, and Conflict Block. A Data Anchor Block provides high-density quantitative information and serves empirical credibility. A Narrative Cut-in Block introduces a concrete scene, quotation, or close-up and serves reader grounding. A Deep Insight Block provides causal attribution, trend prediction, or structural interpretation and serves analytical depth. A Conflict Block stages competing interests or contradictory positions and serves narrative tension. The planning logic therefore explicitly separates macro-structure from micro-realization: schemas define overall argument topology, while Atomic Blocks define local rhetorical function.

4. Scoped execution and adversarial prompting

DeepNews is described as a Map-Reduce-style architecture with a Directed Acyclic Graph topology. Search Nodes retrieve evidence in parallel, Data Clean Nodes restructure it, a Planner generates the hierarchical outline, Writer Agents execute section-level sub-tasks, and an Assembler stitches the final article. A crucial implementation principle is “Scoped Context Injection”: each Writer Agent receives only the evidence relevant to its section, rather than the entire retrieval pack. This is designed to reduce attention dilution and to avoid the “Lost in the Middle” problem (Jiang, 10 Dec 2025).

The execution module is coupled with adversarial constraint prompting. This is the framework’s explicit mechanism for breaking probabilistic smoothness during generation. Three named tactics are emphasized. “Rhythm Break” enforces alternation between long and short sentences. The appendix specifies that if

$C$5

with a high threshold such as 40 tokens, then

$C$6

with a low threshold such as 5 tokens; otherwise stochastic variation is applied with parameter $C$7. “Logic Fog” suppresses explicit discourse connectives such as “therefore” and “however,” so that causality is implied by juxtaposition rather than mechanically signposted. “Lexical Hedge” mixes professional terminology with colloquial or slang expressions to avoid sterile corporate-register output (Jiang, 10 Dec 2025).

The prompt stack used during execution has four layers: system role definition, context injection, adversarial constraint, and output requirement. In the appendix, this includes role conditioning such as “financial editor,” schema IDs, atomic fact lists, tactic IDs, forbidden connector lists, and the target Atomic Block type for the current section. This indicates that adversarial pacing is not a post-editing heuristic but a first-class drafting constraint.

A plausible implication is that DeepNews operationalizes style not as global “voice” conditioning but as local constraint satisfaction under section-specific evidence and function. That is, style emerges from pacing, omission of explicit connectives, lexical register mixing, and structural alternation rather than from a single persona prompt.

5. Empirical evaluation and reported performance

The framework is evaluated through four main empirical components: a Knowledge Cliff experiment, an ecological validity blind test, an ROI-style cost analysis, and an ablation study (Jiang, 10 Dec 2025).

The blind test is a 40-day single-blind evaluation conducted inside the submission system of a top Chinese technology media outlet. Human editors were unaware of AI involvement. The Red Team baseline used GPT-5 with a strong persona-based prompt. The Blue Team used DeepSeek-V3-0324 with the full DeepNews workflow: schema planning, retrieval saturation, and adversarial prompting. The Red Team submitted 10 pieces and had 0 accepted, for a 0% acceptance rate. The Blue Team submitted 12 pieces and had 3 accepted, for a 25% acceptance rate. Editor feedback reportedly described the baseline as lacking depth, sounding like a press release, and exhibiting overly smooth logic, while accepted DeepNews pieces were described as data-rich, tightly structured, and insight-bearing (Jiang, 10 Dec 2025).

The cost analysis introduces Effective Cost per Acceptance. In the main text, DeepNews is described as using roughly 200k tokens per article and costing about $C$8 per generation, while the GPT-5 baseline costs about $C$9 per generation. Because the baseline acceptance rate is 0%, its effective cost is treated as infinite. For DeepNews, the text gives

$S$0

yuan per accepted article. The appendix also reports a different currency table, with average per-article costs around $S$1 for the Blue Team, so the paper itself contains a unit inconsistency.

The ablation study evaluates Full Model, w/o Schema, w/o Tactics, and Human Expert reference on 20 financial topics, for $S$2 total samples. It reports the following values: Human Expert Structural Entropy 1.328, Burstiness 0.537, Subjectivity Score 13.7; Full Model Structural Entropy 1.298, Burstiness 0.656, Subjectivity Score 17.2; w/o Schema Structural Entropy 1.101, Burstiness 0.589, Subjectivity Score 8.5; w/o Tactics Structural Entropy 1.214, Burstiness 0.321, Subjectivity Score 5.6. This supports the specific claim that schema drives structural organization, while adversarial tactics drive burstiness and subjectivity (Jiang, 10 Dec 2025).

The evaluation also has explicit limitations. The paper does not report $S$3-values, confidence intervals, annotator agreement, or variance analyses for its main generation experiments. HFR is defined verbally rather than through a fully operational annotation protocol. Retrieval details remain underspecified. Context units alternate between characters and, in one teaser-style statement, tokens. These limitations mean the reported results are system-level and ecological rather than benchmark-standard in the strict experimental sense.

6. Position within broader news-AI research

DeepNews belongs to a broader family of multi-stage news systems, but it occupies a specific niche: long-form vertical-domain generation under high factuality constraints. Earlier work had already established several ingredients that DeepNews consolidates into one workflow. Weakly aligned multimodal news understanding appears in BreakingNews, which frames news as a domain where image-text relations are “connotative and ambiguous” rather than descriptively aligned (Ramisa et al., 2016). NewsEmbed shows that document-level news representations benefit from web-scale weak supervision, multilingual alignment, and freshness-aware retraining (Liu et al., 2021). Visually-Aware Context Modeling for news image captioning uses CLIP-based sentence retrieval, face-name grounding, and article-conditioned generation, illustrating the utility of retrieval-first conditioning and modular fusion in news generation tasks (Qu et al., 2023). Few-shot fake-news systems such as MetaFEND and the Detect–Investigate–Judge–Determine line of work show the value of event adaptation, retrieved exemplars, and multi-stage reasoning under scarce supervision (Wang et al., 2021, Liu et al., 2024). NewsLens extends the news pipeline idea into multi-agent adversarial bias navigation, where disagreement among agents is itself treated as an analytic signal (Bose, 17 May 2026). Pipeline reproducibility and orchestration concerns are made explicit in EdnaML, which models data collection, preparation, training, deployment, and monitoring as declarative sub-pipelines (Suprem et al., 2022).

A plausible interpretation is that DeepNews synthesizes these tendencies into a generation-first architecture: retrieval saturation rather than sparse evidence, schema-guided macro-planning rather than open-ended drafting, scoped execution rather than monolithic prompting, and adversarial pacing rather than fluency maximization. In that sense, it is less a generic “news model” than a strong statement about workflow design: high-quality vertical news generation is treated as a structured cognition problem rather than as pure sequence modeling.

Its main contributions are therefore architectural and epistemic. Architecturally, it proposes a DAG-based agentic workflow with tri-stream foraging, DNFO-v5 schema planning, Atomic Blocks, and adversarial drafting. Epistemically, it asserts that truthfulness in deep reporting requires redundancy, that logic should be scaffolded explicitly, and that stylistic human-likeness may require deliberate resistance to statistical smoothing. Whether those claims generalize beyond financial journalism remains an open question, but within the paper’s own formulation, DeepNews is presented as a domain-specific framework for converting high-entropy evidence into publishable, low-hallucination, structurally differentiated long-form news text (Jiang, 10 Dec 2025).