Dynamic Atomic Fact Extraction

• Dynamic atomic fact extraction is a process that decomposes complex claims into minimal, self-contained atomic facts to facilitate independent verification.
• It employs adaptive reinforcement learning and feedback-driven techniques to refine fact granularity based on task-specific constraints.
• This approach enhances claim verification, knowledge graph updates, and natural language inference through improved factual accuracy and system stability.

Dynamic atomic fact extraction refers to the adaptive and context-sensitive process of decomposing complex textual claims or narratives into minimal, independently-verifiable factual units ("atomic facts") in a manner optimized for downstream reasoning, verification, or retrieval tasks. This paradigm merges advanced decomposition policies, explicit atomicity quantification, and feedback-driven optimization (commonly via reinforcement learning) to dynamically align fact granularity with the requirements of subsequent models or evaluative frameworks. It contrasts with static, hand-crafted decomposition strategies by iteratively refining and controlling the decomposition process according to empirical system feedback, task constraints, or changing input distributions.

1. Formalization of Atomicity and Atomic Facts

Atomic facts are minimal, self-contained propositions, each corresponding to an independently checkable factual claim. Formally, given a textual span $c$, its atomicity $\alpha(c)$ is quantified as the base-2 logarithm of the number of independent atomic facts it conveys: $\alpha(c) = \log_2 N(c)$ where $N(c)$ denotes the count of atomic facts in $c$ (Lu et al., 19 Mar 2025). In practice, atomic facts are typically represented as simple triples $(s, r, o)$ or short self-contained sentences; minimality demands that no further decomposition yields independently verifiable sub-facts, and decontextualization ensures that each atomic fact can be interpreted correctly without reference to surrounding context (Gunjal et al., 28 Jun 2024, Yan et al., 22 Oct 2024, Lu et al., 19 Mar 2025).

Distinctions between atomic and "molecular" facts have been proposed to capture the tension between minimality and decontextuality—molecular facts are extended forms of atomic facts that resolve ambiguity (e.g., by disambiguating pronouns or generic entities, but with the least possible additional information to maximize verifiability) (Gunjal et al., 28 Jun 2024).

2. Model Architectures and Optimization Objectives

The construction of dynamic decomposers is commonly framed as a bilevel optimization problem, seeking a decomposition policy $\pi_d$ that, for each input claim $C_i$, chooses a decomposition $\{c_j\}$ such that an external verifier $\mathcal{V}$ achieves maximal accuracy: $$\max_{\pi_{d}}\; \mathbb{E}_{i} \left[ \mathbf{1} \left(Y_i = \bigwedge_{c \in \{c_j\}} \mathcal{V}(c \mid \pi_v) \right) \right]$$ Here, $Y_i$ is the gold label, and the logical AND reflects strict compositional verification semantics (Lu et al., 19 Mar 2025). Approximating the globally optimal policy is strongly NP-hard, motivating the use of reinforcement learning (RL), with policy networks trained using feedback such as verification confidence. The policy network observes the current set of subclaims and their embeddings, and dynamically chooses whether (and where) to further decompose, aiming to maximize downstream verifier performance rather than simply minimizing claim length or maximizing number of splits (Lu et al., 19 Mar 2025).

Encoder-only architectures (such as JEDI) sidestep generative decomposition at inference by learning extractive span rationales and optimally combining global and span-level inferences, demonstrating that atomic fact extraction and inference can be efficiently accomplished in a single pass given appropriate supervision (Popovič et al., 23 Sep 2025).

3. Dynamic Extraction Algorithms and Feedback Mechanisms

A typical dynamic atomic fact extraction loop integrates incremental decomposition, evidence retrieval, verification, and adaptive control. For instance, AFEV implements a multi-stage loop:

1. Iterative Decomposition: At each step $t$, extract atomic fact $$F_t = \text{Extractor}(C, F_{1:t-1}, y_{1:t-1}, r_{1:t-1})$$, where $F_{1:t-1}$ are prior facts, $y_{1:t-1}$ their labels, and $r_{1:t-1}$ rationales.
2. Coverage Checking: Evaluate if the union of extracted facts covers the claim sufficiently. If not, extract the next atomic fact.
3. Evidence Retrieval & Reranking: For each $F_t$, perform bi-encoder retrieval (cosine similarity), followed by cross-encoder reranking.
4. Adaptive Demonstration Selection: Dynamically select in-context examples to adapt the reasoning for each atomic fact.
5. Reasoning & Aggregation: Produce a verdict and rationale per atomic fact, and aggregate all verdicts for the final claim classification (Zheng et al., 9 Jun 2025).

Reward or feedback signals for decomposition policies include end-task verification accuracy and proxy signals such as verification confidence (output probability margin of the verifier), providing continuous, label-free feedback for RL training (Lu et al., 19 Mar 2025). In ATOM, the feedback loop is extended to stability and exhaustivity metrics in the context of temporal knowledge graph induction (Lairgi et al., 26 Oct 2025).

4. Practical Instantiations: Human-in-the-Loop and Fully Automatic Systems

Approaches span the spectrum from fully automated RL-trained policies to semi-automatic and human-guided pipelines:

• RL-Based Policies: PPO-trained policies operate over learned state embeddings, choosing DECOMPOSE or STOP actions per claim or subclaim, invoking a frozen LLM as decomposer and an external verifier for reward calculation (Lu et al., 19 Mar 2025). Policy architectures are typically shallow MLPs, and effective state updates utilize GRUs and contextual BERT embeddings.
• Human-in-the-Loop Pipelines: Visual analytics frameworks introduce revision loops, visualization of semantic similarity and referential dependencies, coordinated views for annotation, and guided protocol for consensus atomic fact extraction and refinement, ultimately stabilizing annotation protocols and improving benchmarks for LLM factuality evaluation (Schmidt et al., 1 Sep 2025).
• Encoder-Only Models: JEDI demonstrates that robust atomic fact extraction and NLI reasoning can be supervised via synthetic rationale corpora, enabling single-pass, non-generative extraction that remains robust out-of-distribution (Popovič et al., 23 Sep 2025).
• Integration with Information Retrieval and Knowledge Graphs: ATOM demonstrates document-to-atomic-fact splitting, minimal context chunking, dynamic temporal KG construction (with dual-time modeling for observed/valid intervals), and LLM-independent merging for high-efficiency, parallelizable, and scalable dynamic fact KB construction (Lairgi et al., 26 Oct 2025, Li et al., 25 Mar 2025).

5. Empirical Results and Benchmarking

Dynamic atomic fact extraction consistently improves verification confidence and accuracy compared to static or prompt-engineered baselines. RL-trained dynamic decomposers achieve average improvements of +0.07 in confidence and +0.12 in verification accuracy across datasets and verifiers (Lu et al., 19 Mar 2025). Ablation studies show that expressivity (e.g., two-layer vs. one-layer policies), binary vs. ternary splits, and entropy-based exploration regularization significantly impact performance and generalizability.

The JEDI architecture outperforms strong extractive-only baselines, with marked improvements in out-of-distribution and adversarial NLI tasks (Popovič et al., 23 Sep 2025). ATOM achieves up to +31% factual exhaustivity, +17% temporal exhaustivity, and ~94% stability (centroid cosine), exceeding prior KG induction methods while reducing latency by up to 95% (Lairgi et al., 26 Oct 2025). Human-in-the-loop systems achieve convergence measured by inter-annotator Jaccard index beyond the 0.8 guideline, and embedding-based quantitative alignment (Schmidt et al., 1 Sep 2025).

AtomicTableLLM—trained on modular atomic skills—shows state-of-the-art table claim verification, halving error rates relative to chain-of-thought LLMs using only a fraction of data (Zhang et al., 8 Jun 2025).

6. Extensions, Generalization, and Open Challenges

Dynamic atomic fact extraction extends to diverse domains beyond pure textual verification, including dynamic temporal knowledge graph updates (Lairgi et al., 26 Oct 2025), attributed QA with clause-level editing and attribution reports (Yan et al., 22 Oct 2024), NLI with interpretable atomic inference (Stacey et al., 2023), and sequential agent planning, where extracted atomic facts serve as minimal abstractions guiding in-context LLM planning and lookahead evaluation (Holt et al., 10 Jun 2025).

Remaining open challenges include precise operationalization of atomicity for ambiguous or context-dependent facts, balancing decontextuality with minimality to optimize the tradeoff between human-interpretability and verification ease (Gunjal et al., 28 Jun 2024), handling cascading errors in automated cascades, resolving annotator disagreement, and scaling RL- or feedback-driven policies under high-latency conditions. Emerging work also demonstrates the need for robust, data-driven annotation protocols, integration of similarity-based fact alignment, and continuous, on-line optimization for evolving input sources (Schmidt et al., 1 Sep 2025, Ullrich et al., 7 Feb 2025).

7. Representative Instantiations and Comparison Table

The following table collates representative dynamic atomic fact extraction systems, their core methodology, and main task context, as substantiated by cited literature:

System / Paper	Extraction Policy	Feedback / Adaptation	Benchmark / Task Context
DyDecomp (Lu et al., 19 Mar 2025)	RL (PPO) over binary/stop actions with LLM decomposer	Verifier confidence; bilevel optimization	Claim verification (factuality)
JEDI (Popovič et al., 23 Sep 2025)	Encoder-only, extractive rationale heads	Multi-task learning with synthetic spans	NLI, fact-checking
AFEV (Zheng et al., 9 Jun 2025)	Iterative LLM-driven + feedback	Adaptive evidence reranking, demonstration selection	Complex claim verification
ATOM (Lairgi et al., 26 Oct 2025)	Few-shot LLM, chunk-wise, parallel extraction	Exhaustivity/stability metrics; explicit merging	Temporal KG induction
Visual Fact Annotation (Schmidt et al., 1 Sep 2025)	Guided human-in-the-loop, embedding alignment	Visual analytics, IAA via similarity/graphs	Fact annotation, LLM evaluation
FADER (Li et al., 25 Mar 2025)	LLM query-guided, multi-sample augmented	Query speculation, fact augmentation	Retrieval-QA, long context
AtomicTableLLM (Zhang et al., 8 Jun 2025)	Skill-chaining (modular, prompt-driven)	Plan-evidence-reason aggregation	Scientific table verification

Each approach varies in policy type, supervision signal, and deployment scenario, but shares the central principle of directly optimizing the extraction pipeline for dynamic, task-dependent, and verifiability-centric objectives.