PaperlayerView exact revision on arXiv

Source-linked AI summary

SerenAI: State-transition system inspired by text-based world AI models

Elvin Babayev, Artem Sinitsa, Arash Hajisharifi, Kabir Bakhshaei

arXiv:2609.06647v1cs.AIcs.CL

TL;DR

Professional workflows need more auditable alternatives to unconstrained language-model text when errors can create legal, operational, or financial failures. SerenAI addresses this gap with a text-based, world-model-inspired state-transition system trained through parameter-efficient fine-tuning and verifier-based reinforcement learning. On an internal benchmark, it improved structured transition prediction over an 8B open-weight baseline across JSON, schema, transition, reward, and termination measures, while remaining short of legal-grade reliability.

  • Problem

    Professional workflows lack guarantees that unconstrained generated text is factual, internally consistent, or compliant, making auditing difficult in high-stakes settings.

  • Method

    SerenAI predicts a four-field symbolic transition and combines parameter-efficient supervised fine-tuning with verifier-based reinforcement learning.

  • Results

    SerenAI improved structured transition prediction over an 8B open-weight baseline across JSON, schema, transition, reward, and termination measures.

  • Takeaways & Limitations

    The results support the narrower claim that verifier-compatible adaptation can improve structured transition prediction.

  • Takeaways & Limitations

    The current benchmark does not demonstrate reliable interpretation of court cases, case law, contractual language, or professional procedure.

Abstract

from arXiv · show

Although professional workflows leverage large language models widely, the interpretation for auditing unconstrained free-text generation is usually intractable if such generation demands legal, operational or financial workflow. We hereby demonstrate a text based system called SerenAI - inspired by world-models, it is a state transition system that outputs verifiable predictions rather than merely text: Provided with a description of the environment, state, and actions, the generated output contains 4 items: causal deltas that causally effect the given state, a next state that can logically follow from the given state and action, a validity reward, and a termination signal. For the released proto-model, we employ 2 steps of adaptation training, namely parameter efficient fine-tuning followed by verifier based RL over 50,000 exampled cause and effects in 12 environments spanning 10 reasoning domains. Compared to an initial internal evaluation of an 8B open-weight baseline, SerenAI increased JSON validity from 85.0% to 93.2%, schema validity from 55.0% to 84.0%, exact structured-output match from 0.0% to 41.5%, causal-delta exact match from 0.0% to 41.5%, resulting-state exact match from 0.0% to 42.0%, reward exact match from 1.0% to 80.5%, and termination exact match from 38.0% to 81.5%. These support the narrower claim that verifier-compatible adaptation can improve structured transition prediction. They do not yet establish legal-grade reliability. Accordingly, the paper also specifies a validation protocol for evidence-grounded legal workflows, calibration, human oversight, and sovereign on-premise deployment.

I. INTRODUCTION

SerenAI addresses the difficulty of auditing unconstrained language-model text in high-stakes workflows by representing professional reasoning as verifiable textual state transitions. Its scope is narrower than a general world model and does not yet establish legal-grade reliability.

  • I. INTRODUCTION: High-stakes workflow failures can arise when generated text contains unsupported regulatory claims, mis-referenced legislation, missing deadlines, or errant clauses.The paper frames grounding, constrained generation, verification, and human checks as responses to these risks.
  • I. INTRODUCTION: SerenAI targets structured-form reasoning for private and regulated environments, initially emphasizing legal document review, compliance, deadline monitoring, and legal discovery.Its technical documentation uses synthetic or collected environments and does not yet validate legal advice.
  • I. INTRODUCTION: SerenAI adapts the world-model state-transition principle to text by conditioning explicit symbolic predictions on workflow states and actions.The system is described as world-model-inspired rather than as a general physical or multimodal world model.
  • I. INTRODUCTION: The four-field transition representation specifies what changed, the next state, validity, and completion, while training combines parameter-efficient supervised fine-tuning with verifier-based reinforcement learning.The paper also presents preliminary internal benchmarks and a deployment and validation framework for regulated practice.
  • I. INTRODUCTION: Unlike smooth free-text generation, the transition model evaluates specific variables and requires a verifiable schema for entity identity, temporal ordering, preconditions, and termination conditions.The paper characterizes this as closer to symbolic workflow simulation than generic dialogue.

B. Efficient Adaptation and Verifier-Guided Learning

SerenAI combines efficient adaptation with verifier-guided learning to produce structured transition outputs that can be checked field by field. The approach builds on parameter-efficient adaptation, verification, constrained-generation research, and evidence-grounded legal evaluation.

  • B. Efficient Adaptation and Verifier-Guided Learning: Parameter-efficient approaches such as QLoRA train low-rank adapters on a frozen quantized backbone, supporting memory-efficient adaptation for sovereign AI.The cited related work presents this as a way to reduce training and deployment costs while retaining control of weights and data.
  • B. Efficient Adaptation and Verifier-Guided Learning: SerenAI uses automated verification as scalable supervision and applies GRPO-style reinforcement learning to sampled structured outputs without claiming a novel policy-optimization algorithm.Its stated contribution is fusing verification with the task's structured output format.
  • B. Efficient Adaptation and Verifier-Guided Learning: Legal AI evaluation requires task-level benchmarks and expert consultation because generic reasoning benchmarks alone do not ensure safe legal performance.The related-work discussion points to LegalBench and CUAD as examples of task-specific legal evaluation.
  • B. Efficient Adaptation and Verifier-Guided Learning: The output tuple contains a causal delta, resulting state, reward or validity indicator, and termination signal.This representation separates transition content from task status.
  • B. Efficient Adaptation and Verifier-Guided Learning: The schema supports deterministic validation of syntax, required keys, data types, and exact field values.A canonical output names the fields causal_delta, resulting_state, reward, and done.

IV. SYSTEM ARCHITECTURE

The SerenAI architecture separates research-time adaptation and verification from deployment-time evidence handling and review controls. It is designed around data sovereignty, with audit feedback returning to evaluation rather than automatically training on customer data.

  • IV. SYSTEM ARCHITECTURE: The research pipeline uses curated state-action-transition data, supervised fine-tuning, deterministic verification, and reinforcement learning to improve outputs satisfying format and target-state conditions.The deployment side adds citations, confidence signals, logging, and human-review gates.
  • IV. SYSTEM ARCHITECTURE: Data sovereignty keeps sensitive prompts, files, intermediate stages, and responses on premises while local retrieval anchors outputs in customer-controlled repositories.Audit logs record each response and model version.
  • IV. SYSTEM ARCHITECTURE: The architecture's dashed return path represents audit feedback from deployment into evaluation rather than automatic customer-data training.Solid arrows denote data or inference flow.

V. DATASET AND TRAINING PROCEDURE

Training uses a 50,000-example cause-and-effect corpus spanning multiple environments and reasoning domains, followed by parameter-efficient supervised fine-tuning on structured target sequences. The objective is token-level prediction of the target transition format.

  • V. DATASET AND TRAINING PROCEDURE: 50,000 cause-and-effect examples span 12 environments and 10 reasoning domains, with 45,000 training, 2,500 validation, and 2,500 test records.Examples contain an initial context, an action or event, and a target four-field transition.
  • V. DATASET AND TRAINING PROCEDURE: The first training phase uses parameter-efficient supervised fine-tuning while keeping the quantized backbone frozen and learning task-specific low-rank adapters.The supervised objective minimizes token-level negative log-likelihood for structured target sequences.
  • V. DATASET AND TRAINING PROCEDURE: The training loss sums prediction errors across all examples and across every target token.The formulation interprets the model's error as a positive penalty based on next-token probabilities.
  • V. DATASET AND TRAINING PROCEDURE: For each token, the model is trained to assign probability to the actual next correct token given the prompt and preceding target tokens.The notation distinguishes the task input, prior answer prefix, and next target token.

Summary:

The system uses supervised training followed by verifier-based reinforcement learning to produce structured state-transition outputs. Training loss declines rapidly before stabilizing, while the second stage scores sampled outputs with an automatic verifier.

  • The training objective evaluates next-token probabilities against correct answers across the database.
  • Training loss declines rapidly at the beginning and then stabilizes as the model fits the structured target format.
  • The second stage uses an automatic verifier to score sampled outputs during GRPO optimization.
  • Probability-ratio clipping is applied with ϵ = 0.2.

VI. EVALUATION METHODOLOGY

The internal evaluation measures whether outputs are parseable, schema-compliant, and exactly match transition components. Exact matching is strict and may undervalue semantically equivalent text states.

  • The evaluation uses seven exact or validity-based metrics covering JSON, schema, complete outputs, transitions, states, rewards, and termination.
  • JSON validity measures whether an output parses as JSON, while schema validity checks for all four required fields and data types.
  • Exact metrics assess complete-output, causal-delta, resulting-state, reward, and termination matches.
  • Exact match is automatically verifiable but can undervalue semantically identical text states.

B. Preliminary Results

The preliminary results indicate improved machine-readability and field-level exact matching after adaptation. The benchmark supports task-learning evidence, but does not distinguish memorization, limited rules, transfer, or adversarial robustness, and verifier scoring introduces reward-misspecification risks.

  • The reported results indicate better machine-readability and larger gains in field-level exact matches.
  • Observed causal-delta and resulting-state improvements reached roughly 42%, supporting task-learning evidence on the internal benchmark.
  • Reward and termination were the largest absolute improvements, reaching 80.5% and 81.5% exact-field accuracies under majority-baseline analysis.
  • A verifier scoring only precise fields may encourage policy formatting or exploitable easy fields instead of difficult transition content.

B. Required Ablation Matrix

The paper compares a base model with the final SFT-and-RL system and proposes ablations and stress tests for a reproducible rerun. The current benchmark supports structured transition prediction, not reliable legal interpretation.

  • Figure 3 compares a base model with the final system after supervised fine-tuning and reinforcement learning.
  • The reported comparison shows gains especially in JSON validity, schema validity, and termination-flag accuracy.
  • The ablation matrix includes SFT-only, constrained decoding, and inference-time verifier conditions to isolate their effects.
  • Stronger evaluation should use held-out environment, composition, and counterfactual stress splits rather than relying only on random record splitting.
  • Robustness testing should vary serialization, context, paraphrase, contradictions, missing preconditions, and incomplete evidence while reporting correctness and coverage.
  • The current benchmark supports adapting language models to verifier-compatible structured transitions but does not demonstrate reliable legal or professional interpretation.

A. Legal Task Taxonomy

The proposed legal benchmark spans four task families, while validation must extend beyond task coverage to evidence quality, calibration, failure analysis, and expert-gated release decisions.

  • The first legal benchmark should cover contract state tracking, procedural timelines, compliance workflows, and evidence-grounded research.
  • High-risk legal validation requires qualified legal professionals and versioned primary sources because hallucinations and errors are unacceptable in these environments.
  • Validation should measure citation precision and recall, support entailment, retrieval coverage, confidence calibration, selective risk under abstention, and repeated-run consistency.
  • No paid pilot should advance from assistive review to production based on aggregate accuracy alone; gates require preregistered criteria, failure analysis, and independent expert sign-off.

IX. SOVEREIGN DEPLOYMENT AND GOVERNANCE

Sovereign deployment keeps sensitive workflow materials within an approved boundary, but governance must address model, retrieval, application, and confidentiality risks through explicit controls and audit evidence.

  • On-premise or customer-controlled deployment keeps documents, prompts, outputs, and audit logs within an approved boundary.
  • Governance should include data minimization, purpose limitation, local processing, documented roles, human oversight, technical documentation, monitoring, and incident response.
  • Sovereign deployment reduces third-party cloud dependence but does not remove model, retrieval, or application-layer risk.
  • The security boundary includes model weights and adapters, customer documents, retrieval indexes, prompts, tool permissions, generated outputs, and audit logs.
  • Retrieved material must remain evidence rather than executable instruction, with separate instruction and evidence channels to address indirect prompt injection.
  • Confidentiality can still fail through access controls, logs, backups, generated text, or model memorization despite on-premise inference limiting routine data transfer.

C. Evidence Freshness and Jurisdiction Drift

Legal evidence must remain tied to jurisdiction, timing, procedural context, and source authority, while the system’s structured interface supports validation without establishing legal-grade reliability.

  • Legal validity depends on jurisdiction, effective date, procedural posture, and source hierarchy, so authentic citations may still be inapplicable.
  • The retrieval layer should store source type, authority, publication date, effective interval, and version, with jurisdiction and temporal-validity fields in legal-research outputs.
  • High-severity actions such as deadline calculations, filing decisions, compliance approvals, and contract terminations require explicit human confirmation.
  • A reproducible next experiment should use a fixed model revision, public or escrowed test manifest, SFT-only and SFT+RL ablations, at least three seeds, stratified results, and an expert-reviewed legal benchmark.
  • The four-field output contract enables automatic verification and interfaces language-model adaptation with regulated workflow controls.
Loading 2609.06647v1…