Technical Summary: X-SYNTH — Enterprise Context Synthesis from Observed Human Attention

1. Problem Statement

In enterprise operations, the context required for AI agents to complete complex tasks is scattered across systems of record, static stores, and communication channels. Current prevailing approaches rely on Retrieval-Augmented Generation (RAG), which matches request content to stored data. While effective for narrow, content-driven queries (e.g., "Find invoice INV-34231"), this approach fails for complex agentic tasks requiring synthesis from dispersed behavioral evidence.

The core failure mode is the absence of a principled relevance signal. Standard retrieval systems treat data as static text, missing the "ordered sequence of actions" and "implicit reward signals" that encode how work actually happens. Consequently, agents suffer from low True Lead Rates (TLR) and high False Lead Rates (FLR) because they cannot distinguish between noise and causally relevant behavioral patterns specific to an organization, team, or individual. The paper argues that enterprise context synthesis is not a retrieval problem, but a relevance problem, where human attention serves as the ground truth.

2. Methodology: The X-SYNTH Framework

X-SYNTH is a four-stage agentic pipeline that grounds context synthesis in human attention—defined as the digitally observable interaction signatures of workers. The system models each individual's behavioral baseline as a Digital Twin Signature (DTS) and uses a learned router to select among seven distinct attention filters.

Core Components

A. Digital Twin Signature (DTS)
For each individual $u$, the system maintains a rolling behavioral profile $DTS(u, \tau)$ over a window $\tau$. It consists of five components:

1. $v_{dom}$ (Domain Attention): Concentration of attention across work domains.
2. $v_{rhythm}$ (Behavioral Rhythm): Statistics on dwell time, revisits, and transitions.
3. $v_{base}$ (Baseline): Extended lookback statistics establishing the individual's "normal."
4. $v_{resp}$ (Responsibility Profile): Inferred domain ownership derived from activity, not declared roles.
5. $v_{div}$ (Divergence): The difference between recent attention patterns and the long-term baseline.

B. Seven Attention Filters
A learned modality selector $f_\theta$ (a 3-layer MLP) maps a query $q$ and the user's DTS to a probability distribution over seven qualitatively distinct filters:

1. Proportional: High dwell time signals importance (expected attention).
2. Inverse: Absence of expected attention signals importance.
3. Differential: Deviation from the individual's own baseline signals importance.
4. Recurrent: Frequent return to an artifact signals importance.
5. Comparative: Rapid alternation between similar artifacts signals active evaluation.
6. Sequential: Deviations from expected workflow sequences.
7. Collective: Aggregates attention across a cohort to find consensus or outliers.

C. The Four-Stage Pipeline

1. Subject Scoping: Identifies the target set of individuals $U_q$ based on the query.
2. Modality Selection: For each $(q, u)$ pair, the system computes $m(q, u) = \text{softmax}(f_\theta([q; DTS(u, \tau)]))$ to select the appropriate filter(s).
3. Retrieval & Weighting: Artifacts are ranked by a combined weight $w(a_j, q, u) = I_{attn}(a_j, q, u) \cdot I_{content}(a_j, q)$. This ensures artifacts are both content-relevant and behaviorally significant.
4. Synthesis: An LLM synthesizes the response using the ranked evidence and modality annotations.

D. Feedback Loop
The system decomposes terminal feedback into stage-level failure probabilities. The modality selector is updated only when failure is attributed to Stage 2 (Modality Selection), preventing the training signal from being polluted by retrieval or synthesis errors.

3. Key Contributions

The paper outlines four primary contributions:

1. Human Attention as Behavioral Ground Truth: Operationalizing human attention as a learnable, discriminative relevance signal without external labels, encoded via the DTS.
2. Seven-Filter Attention Modality Framework: A structured set of qualitatively distinct filters selected dynamically per individual and query via a lightweight router.
3. Four-Stage Agentic Synthesis Pipeline: A workflow that integrates subject scoping, individualized modality selection, attention-weighted retrieval, and synthesis.
4. Credit-Attributed Feedback Loop: A mechanism to isolate and update the modality selector only when specific to modality misclassification, ensuring clean training signals.

4. Evaluation and Results

The framework was evaluated on a benchmark of 302 instances derived from real digital interaction traces of five knowledge workers at a Fortune 500 company (210 positive instances of business opportunities, 92 negative).

Baseline Performance:

• A frontier model (Claude Opus 4.6) operating directly on raw interaction sequences achieved a True Lead Rate (TLR) of 9.5% and a False Lead Rate (FLR) of 90.5%.
• A smaller model (GPT-4o-mini) achieved a TLR of 0%.

X-SYNTH Performance:

• Frontier Model + X-SYNTH: TLR increased to 61.9% (a 6.5× improvement), capturing 130 of 210 real leads. FLR dropped to 18.8%.
• Small Model + X-SYNTH: TLR increased from 0% to 57.1%, demonstrating that the framework can make smaller models functional for complex tasks.

Qualitative Findings:

• Vocabulary Gap: X-SYNTH successfully interpreted organization-specific abbreviations (e.g., "FZ" for "Formal Zone") by leveraging the DTS, whereas direct inference failed due to lack of organizational priors.
• Scale & Selectivity: Direct inference is infeasible at scale due to token volume (up to 2.5M tokens/day per seller). X-SYNTH's selective retrieval makes evaluation tractable.
• Temporal Integration: X-SYNTH successfully integrated signals across multiple days and individuals, whereas single-day windows missed critical context.

5. Significance and Claims

The paper claims that enterprise context synthesis is not a retrieval problem, but a relevance problem. The absence of a relevance signal grounded in human behavior is the primary bottleneck for enterprise AI agents, not the reasoning capability of the underlying LLM.

X-SYNTH demonstrates that by modeling individual behavioral baselines (DTS) and selecting attention filters dynamically, systems can:

• Recover causal relevance from dispersed behavioral evidence.
• Overcome the "enterprise vocabulary gap" by grounding inference in accumulated organizational priors.
• Scale to high-volume interaction streams through behaviorally-grounded compression rather than keyword matching.

The authors conclude that human attention is the most reliable ground truth for enterprise relevance, and that X-SYNTH provides a principled mechanism to operationalize this signal, significantly improving the utility of AI agents in complex enterprise workflows.