cs.AISep 6, 2026

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

Authors: Elvin Babayev, Artem Sinitsa, Arash Hajisharifi, Kabir Bakhshaei

Organizations: Collision Technologies S.R.L.S. Societa Benefit Trieste, Italy

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Fig. 1: SerenAI research and deployment architecture. Solid arrows denote data or inference flow; the dashed return path denotes audit feedback from deployment into evaluation rather than automatic customer-data training.
ItemReported value
Total records50,000
Training split45,000
Validation split2,500
Test split2,500 in split description; separate narrative says 5,000 evaluation samples
Environments12
Reasoning domains10; seven examples named in this study
TABLE I: Reported Corpus and Output Design
Fig. 2: Supervised-training loss function
Fig. 3: Reported preliminary benchmark results of SerenAI. Percentages are reproduced from the internal study of Collision Technologies team.
MetricBaseSerenAIΔ\Delta pp
JSON validity85.093.2+8.2
Schema validity55.084.0+29.0
Exact JSON match0.041.5+41.5
Causal delta exact0.041.5+41.5
Resulting state exact0.042.0+42.0
Reward exact1.080.5+79.5
TABLE II: Reported Preliminary Results
ConditionQuestion answered
Base, prompt onlyHow much structure and transition ability exists before adaptation?
Base + constrained decodingHow much validity gain is purely syntactic?
QLoRA SFTWhat is learned from curated demonstrations alone?
QLoRA SFT + verifier selectionDoes inference-time verification improve exact state prediction?
QLoRA SFT + RLVR/GRPOWhat is the incremental policy-learning gain over SFT?
SFT + RL + constrained decodingWhat is the strongest defense-in-depth configuration?
TABLE III: Minimum Ablation Matrix for a Reproducible Rerun
DimensionRequired evidence
State coherenceEntity, date, obligation, exception, and termination consistency
GroundingCitation precision/recall, passage support, source freshness
UncertaintyECE, Brier score, risk-coverage, abstention quality
RobustnessParaphrase, missing evidence, conflicting evidence, adversarial clauses
ReproducibilityModel hash, data manifest, seeds, prompts, verifier version
OperationsLatency, throughput, VRAM, failure recovery, audit-log completeness
TABLE IV: Recommended Legal Validation Dimensions
Fig. 4: Recommended release gates from internal benchmark to monitored legal pilot.
ThreatPrimary controlsRequired evidence
Indirect injectionInstruction/evidence separation, tool allow-list, content sanitizationRed-team corpus, tool-call logs, attack success rate
Cross-user leakageTenant isolation, RBAC, encryption, scoped indicesPenetration test, access review, isolation tests
Memory leakageNo default customer-data training, adapter isolation, output filteringExtraction audit, training-data manifest
Stale authorityVersioned sources, jurisdiction metadata, revocation processFreshness report, source lineage, temporal tests
Verifier bypassCanonical parser, adversarial verifier tests, independent semantic checksUnit-test coverage, failure cases, verifier hash
Automation biasMandatory review for high-severity actions, visible evidence and abstentionReviewer study, seeded-error detection, escalation rate
TABLE V: Operational Threats, Controls, and Audit Evidence
More in this topic

LLM World Models

Large language model world models predict how environments change so agents can plan, learn, and evaluate actions before execution.

See all 31 papers

Explore similar work

CardsList
Aug 17, 2026

Towards Computational Provenance: Carrying Causal-State Evidence in Generated Text

A language model's output does not by itself provide verifiable evidence about the internal computation that produced it. We study computational provenance: whether generated text can carry detectable evidence of which causally relevant internal state occurred. We test a bounded form of this idea in two controlled architectures: a modular feed-forward neural network and a transformer-based model. Both architectures are trained on the same arithmetic task with a mandatory pathway through two discrete intermediate states, allowing different internal paths to produce the same answer. We deliberately switch between these paths, authenticate the state actually used, and let that verified state determine a subtle statistical pattern in the generated text that can later be detected. The feed-forward and transformer systems each passed all 128 matched pairs in both their public and separately sealed protected end-to-end evaluations, with the detector recovering the signal associated with the authenticated internal state. The required causal computation also reproduced across five independently trained feed-forward models and three independently trained transformers. In a separate answer-only transformer experiment, our linear probes did not recover a naturally learned intermediate state. These results provide a controlled proof of concept that information about a verified, causally relevant internal state can be preserved in generated text even when the answer is unchanged.

Benjamin Belay

Independent Researcher.

Data Provenance in Language ModelsLanguage Model Auditing
AskPDF
Sep 21, 2026

World State Generator

Language agents solve complex tasks through plans and actions. A single step the world refuses puts the goal out of reach, and what the agent does next decides the task. Prompted planners fail at exactly this point, rewriting the refused step in new words, meeting the same refusal, and burning the attempt budget without moving. They fail because the plan was never tied to the world, so a refusal has nothing in the plan to attach to. A world is where a task runs, and it has its own rules, its own admissible actions, and its own constraints. We build synthetic worlds across 7 domains and extract training data from them. A program enforces each world's rules and grades its goal, and every world is admitted only if its goal is reachable from its initial state. Agents run inside and leave verified failures paired with repairs that carried the run to a state the world certified, a record of about 226K trajectories. On this record we train the World State Generator, a model that writes a plan as checkable states of the world and keeps that plan aligned with the world it runs in. That alignment is what a plan written in language lacks, since the world it runs in has physical limits, logical dependencies, and required orders the language never states, and the plan encounters these rules only when a state fails. WSG takes that failure as the rule the world has stated and rewrites the remaining states to obey it, so the plan bends to the world as the run goes on. Across 7 public benchmarks, WSG raises end-to-end success for two open models near 30B parameters over prompting and brings to the level of proprietary model.

Sungheon Jeong, Sanggeon Yun, Ryozo Masukawa +3

University of California, Irvine · Northeastern University

LLM Agent PlanningLLM PlanningWorld Model-Based Planning
AskPDF
Jul 10, 2026

Ceci n'est pas une pipe: AI systems as semantic abstractions

An AI system's output is not the fact or world state it appears to describe, but rather an engineered representation. We propose a semantic framework to describe AI systems, to be able to examine the correctness of such representations. To do so, we distinguish what is justified by accepted domain knowledge, what reference sources say, and what the system can currently use. This allows us to give precise definitions to common failures: extrapolation, refuted or unsupported assertion, sources versus knowledge mismatch, stale or refuted source, added hypotheses, unsupported use... We hope our framework gives a useful vocabulary for specifying and checking AI systems whose outputs, citations, tool calls, and world-changing actions must be justified by reliable claims and explicit authority rather than apparent fluency.

Jade Alglave, Patrick Cousot

Topic pending
AskPDF