arXiv subjects

Jason Starace

Publications and source records attributed to Jason Starace.

4 recordsLinked to original sources

arXiv2026

Scaffold Effects on GAIA: A Controlled Comparison

Published agent capability scores conflate what a model can do with what its scaffold lets it do, and the magnitude of this elicitation gap is not well characterized under controlled conditions. This study executes a pre-registered controlled comparison of three scaffolds (ReAct, a Planner-Actor-Rater multi-agent design, and planner-then-executor) across five models from three providers (Claude Opus 4.7, Sonnet 4.6, Haiku 4.5; Gemini 3.1 Pro Preview; GPT-5.5) on GAIA validation Levels 1 and 2, holding tasks and conditions fixed, with three attempts per question. Scaffold choice alone moves measured accuracy by as much as 28 percentage points within a single model (Opus, Level 2, robust slice), confirming the pre-registered hypothesis that scaffold variation produces gaps of at least 10 points. The pre-registered prediction that more capable models would be less scaffold-sensitive is rejected in direction: scaffold effects vary significantly by model in every dataset slice, but the most capable Anthropic model gains the most from structured scaffolds at the harder level, and tier-scaling holds only at Level 1 under the robust slice. The multi-agent advantage over ReAct at Level 2 appears within the Anthropic family but not for the cross-provider models, making model family rather than capability tier the conditioning variable, and the predicted planner-executor advantage on file-reading tasks is falsified. Structured scaffolds make fewer tool calls yet recover more often from mid-trajectory errors at the harder level, and a single cell (Gemini with planner-then-executor) is the cheapest at both levels and the most accurate at Level 2. These results indicate that single-scaffold capability numbers are scaffold-conditional estimates and that the elicitation gap is not guaranteed to shrink as models improve.

cs.AI↗

arXiv2026

Ethical Implications of Training Deceptive AI

Deceptive behavior in AI systems is no longer theoretical: large language models strategically mislead without producing false statements, maintain deceptive strategies through safety training, and coordinate deception in multi-agent settings. While the European Union's AI Act prohibits deployment of deceptive AI systems, it explicitly exempts research and development, creating a necessary but unstructured space in which no established framework governs how deception research should be conducted or how risk should scale with capability. This paper proposes a Deception Research Levels (DRL) framework, a classification system for deceptive algorithm research modeled on the Biosafety Level system used in biological research. The DRL framework classifies research by risk profile rather than researcher intent, assessing deceptive mechanisms across five dimensions grounded in the AI4People ethical framework: Pillar Implication, Severity, Reversibility, Scale, and Vulnerability. Classification follows a ``highest dimension wins'' approach, assigning one of four risk levels with cumulative safeguards ranging from standard documentation at DRL-1 to regulatory notification and third-party security audits at DRL-4. A dual-development mandate at DRL-3 and above requires that detection and mitigation methods be developed alongside any deceptive capability. We apply the framework to eight case studies spanning all four levels and demonstrate that ecological validity of the deceptive mechanism emerges as a consistent, non-independent indicator of classification level. The DRL framework is intended to fill the governance gap between regulated deployment and unstructured research, supporting both beneficial applications and defensive research under conditions where safeguards are proportional to the potential for harm.

cs.CY↗

arXiv2026

Intentional Deception as Controllable Capability in LLM Agents

As LLM-based agents increasingly operate in multi-agent systems, understanding adversarial manipulation becomes critical for defensive design. We present a systematic study of intentional deception as an engineered capability, using LLM-to-LLM interactions within a text-based RPG where parameterized behavioral profiles (9 alignments x 4 motivations, yielding 36 profiles with explicit ethical ground truth) serve as our experimental testbed. Unlike accidental deception from misalignment, we investigate a two-stage system that infers target agent characteristics and generates deceptive responses steering targets toward actions counter to their beliefs and motivations. We find that deceptive intervention produces differential effects concentrated in specific behavioral profiles rather than distributed uniformly, and that 88.5% of successful deceptions employ misdirection (true statements with strategic framing) rather than fabrication, indicating fact-checking defenses would miss the large majority of adversarial responses. Motivation, inferable at 98%+ accuracy, serves as the primary attack vector, while belief systems remain harder to identify (49% inference ceiling) or exploit. These findings identify which agent profiles require additional safeguards and suggest that current fact-verification approaches are insufficient against strategically framed deception.

cs.AI↗

arXiv2025

Behavioral Inference at Scale: The Fundamental Asymmetry Between Motivations and Belief Systems

How much information about an agent's underlying values can be recovered from its observable behavior? This question matters for any approach that infers agent properties from action sequences, yet remains empirically open at scale. We address it through controlled experiments: LLM-based agents (Llama 3.1-8B) assigned one of 36 behavioral profiles (9 belief systems x 4 motivations) generate over 1.5 million behavioral sequences in grid-world environments, providing ground truth unavailable in human behavioral studies. After filtering, classifiers train and evaluate on a shared canonical dataset of 10,338 episodes and 1,200,834 sequences. A fundamental asymmetry emerges in both magnitude and structure. Motivations achieve 98-100% accuracy and recover 97% of available mutual information across all architectures. Belief systems plateau at 24% for LSTMs regardless of capacity, and even transformers reach only 34.0%, recovering 16.3% of available information, a 6.1x asymmetry in extraction efficiency. Per-alignment accuracy ranges from 23.2% (Lawful Neutral) to 59.4% (Chaotic Evil). Confusion analysis maps the failure structure: a neutral zone of behavioral ambiguity centers on True Neutral, absorbing misclassified samples from adjacent Neutral and Good alignments whose prosocial or balance-keeping behavior lacks distinctive signal. Combined inference yields 12.2x improvement over random baseline for full 36-class profile classification, with the bottleneck located entirely in belief system inference. Signal enhancement and explanatory queries yield only marginal LSTM gains (+3.8%), confirming the recurrent ceiling is architectural rather than data-limited. Whether the transformer's 34.0% ceiling reflects a similar architectural-class limit or a more fundamental bound remains open. These results characterize what behavioral observation can and cannot reveal about LLM agent values.

cs.MA↗