Workflow is All You Need: Escaping the "Statistical Smoothing Trap" via High-Entropy Information Foraging and Adversarial Pacing¶
🕒 Published (v1): 2025-12-10 22:13 UTC · Source: Arxiv · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
DeepNews is an agentic workflow for long-form financial news generation that encodes expert journalist cognition into three coupled modules: saturated multi-stream retrieval, schema-guided hierarchical planning, and adversarial constraint prompting. The system empirically demonstrates a "Knowledge Cliff" requiring a 10:1 input-to-output context ratio to suppress hallucinations. In a real-world blind submission test, a prior-generation model (DeepSeek-V3-0324) running DeepNews achieved 25% editorial acceptance vs. 0% for zero-shot GPT-5.
Problem¶
Standard RLHF-trained LLMs fall into a "Statistical Smoothing Trap": their probability distributions bias toward average, safe outputs that lack the logical depth, stylistic burstiness, and factual precision required for expert-level vertical-domain writing (e.g., financial journalism). Existing RAG pipelines treat retrieved content as uniform slices, missing the micro/macro information duality that human experts naturally exploit. There is no quantified threshold for how much context is needed to physically prevent hallucination in long-form generation.
Method¶
DeepNews implements a DAG-based Map-Reduce pipeline with three modules:
-
Tri-Stream Information Foraging: Three orthogonal parallel search streams (Ecological/structural context, Quantitative/hard data, Narrative/event arc) enforce a 10:1 input-to-output ratio (≥30k characters). Retrieved content is restructured into Atomic Facts (micro-structure) and Context Blocks (macro-structure), implementing Kintsch's Construction-Integration model.
-
Hierarchical Strategic Planning: A three-layer planning stack: (a) Macro-Schema Injection selects from DNFO-v5, a curated library of 5 narrative categories × 19 sub-scenarios derived from journalist SOPs; (b) Narrative Orchestrator controls pacing by enforcing alternation of section types (e.g., "Deep Analysis" must follow a "Vignette"); (c) Atomic Block System decomposes each section into single-function paragraph types (Data Anchor, Narrative Cut-in, Deep Insight).
-
Scoped Execution & Adversarial Prompting: Each parallel Writer Agent receives only the context slice mapped to its outline section (mitigating "Lost in the Middle"). Adversarial Constraint Prompts inject "cognitive noise" via Rhythm Break (forced sentence-length variance), Logic Fog (suppression of explicit connectors), and Lexical Hedge (professional/colloquial juxtaposition). The generation objective explicitly maximizes a Burstiness reward term
B(x) = σ(Var(len(sᵢ)))alongside a Hallucination Penalty.
Key Contributions¶
- Information Compression Rate (ICR) and Knowledge Cliff: Empirically establishes that HFR (Hallucination-Free Rate) follows a sigmoid curve with a phase transition at ~30k characters (10:1 ICR); below 15k characters HFR collapses below 40%.
- Dual-granularity retrieval: Operationalizes Kintsch's micro/macro text representation distinction as an engineering primitive within a RAG pipeline.
- DNFO-v5 schema library: Formalizes journalist narrative schemas (e.g., "Game Theory Model," "Regulatory Black Swan") into slot-filling templates, converting open-ended generation into constrained slot completion.
- Adversarial Constraint Prompting formalized: Proposes a generation objective that explicitly penalizes hallucination and rewards syntactic burstiness as a penalty term, not merely a qualitative prompt guideline.
- Ecological validity benchmark: A 40-day real-world editorial blind test at a named Chinese tech-finance outlet, yielding concrete acceptance rates as the evaluation metric.
Results¶
- Knowledge Cliff (Experiment 1, N=100): HFR by context window: <10k → <20%; 15k → 40%; 30k → 85%; 40k → 90%.
- Blind editorial test (Experiment 2): DeepNews (DeepSeek-V3-0324) — 3/12 accepted (25%); GPT-5 zero-shot — 0/10 accepted (0%).
- ROI (Experiment 3): DeepNews effective cost per accepted article ¥2.8 (¥0.7/run × 25% acceptance); GPT-5 effective cost = ∞ (¥0.5/run, 0% acceptance).
- Ablation (Experiment 4, N=80): Full model vs. human expert vs. ablated variants on three metrics:
- Structural Entropy: Full=1.298, Human=1.328, w/o Schema=1.101, w/o Tactics=1.214
- Burstiness: Full=0.656, Human=0.537, w/o Schema=0.589, w/o Tactics=0.321
- Subjectivity Score: Full=17.2, Human=13.7, w/o Schema=8.5, w/o Tactics=5.6
Limitations¶
- Single-domain evaluation: all experiments concern Chinese financial journalism; generalizability to other vertical domains is asserted but untested.
- Small-scale blind test: 22 total submissions (12 Blue, 10 Red) over 40 days at one outlet — insufficient statistical power for significance claims.
- DNFO-v5 is proprietary and human-curated by a single author; schema quality depends on author expertise and is not systematically validated or transferable.
- No comparison against other agentic baselines (e.g., AutoGPT, LangGraph pipelines, ReAct agents); GPT-5 zero-shot is the only baseline.
- Burstiness metric (sentence-length variance) is a surface proxy for stylistic quality; it does not directly measure informational depth or editorial value.
- Cost analysis uses a single media outlet's acceptance rate, which may reflect editorial idiosyncrasies rather than general content quality.
- The "adversarial" framing of prompting tactics is conceptual, not adversarial in the technical ML sense.
Relevance to Harnesses / Meta-Harnesses¶
DeepNews is a concrete exemplar of a domain-specific agentic harness: it wraps a base LLM with a fixed DAG topology of specialized agents (searchers, cleaners, planners, writers, assembler) and encodes domain SOPs as first-class workflow primitives. The empirical finding that workflow architecture dominates base model capability (older model + harness > SOTA zero-shot) is a direct empirical validation of the harness design philosophy. The DNFO-v5 schema library functions as a meta-harness concern — a configurable, domain-swappable knowledge layer that steers subordinate agents — analogous to how meta-harnesses parameterize sub-workflows. The scoped context injection strategy (partitioning retrieval by outline slot) is a generalizable pattern for any multi-agent harness managing long contexts across parallel worker agents.