Owlgorithm: AI in SRL & Hallucination Mitigation

• Owlgorithm is a dual-concept framework offering an AI-powered self-regulated learning platform for competitive programming and a causal attention method to mitigate hallucinations in vision-language models.
• Its educational segment automates metacognitive reflection by generating tailored prompts aligned with cognitive taxonomies, boosting debugging and strategic refinement.
• Its vision-language component employs layer-wise attention adjustments to enhance visual grounding and reduce object hallucinations, achieving state-of-the-art improvements on benchmarks.

Owlgorithm refers to two distinct but thematically unified concepts within the recent research literature: (1) an educational GenAI platform for self-regulated learning (SRL) in competitive programming that leverages LLM-generated reflection prompts, and (2) a causally-grounded attention intervention methodology for object hallucination mitigation in Large Vision-LLMs (LVLMs). Both share a metacognitive orientation—supporting either reflective learning or reliability in multimodal AI—yet are technically and contextually independent. This article delineates both usage domains, underlying theoretical constructs, operational frameworks, empirical results, and anticipated future developments.

1. Self-Regulated Learning Support in Competitive Programming

Owlgorithm, as introduced by Claros-Salinas et al., is a web-based, LLM-driven reflection tool designed to scaffold the “reflection” phase of Zimmerman’s SRL cycle within competitive programming contexts (Nieto-Cardenas et al., 13 Nov 2025). Traditional competitive programming (CP) workflows emphasize forethought (e.g., task selection, planning) and performance (iterative problem solving, code submission), but rarely institutionalize systematic reflection on code correctness, failure causes, or prospective strategy optimization.

Owlgorithm fills this cognitive gap by automatically generating context-sensitive, metacognitive prompts using GPT-4o. These prompts query students about their reasoning, design choices, diagnostic methods, and plans for revision, thus operationalizing SRL reflection. Theoretical alignment is achieved through the dual reference frameworks of Zimmerman’s three-phase SRL model and the Revised Bloom’s Taxonomy, ensuring that generated questions range from knowledge recall to higher-order evaluative and creative reasoning.

The platform automates prompt creation, scoring, and feedback delivery, minimizing instructor burden while promoting learners’ internalization of debugging heuristics, complexity analysis, and algorithmic design principles.

2. System Architecture and Operational Workflow

Owlgorithm employs a multi-agent LLM pipeline to separate pedagogical question generation from secondary tasks such as prompt refinement, rubric construction, and feedback formulation (Nieto-Cardenas et al., 13 Nov 2025). The architecture consists of the following primary components:

• Generator (LLM #1): Drafts initial reflective or diagnostic prompts based on the latest student code and contest verdict (pass/fail).
• Reviewer (LLM #2): Filters, refines, and aligns prompts with Bloom’s taxonomy, constructing exemplar answers and a 0–3 rubric for each question.
• Formatter: Packages curated materials as JSON for UI deployment.
• Feedback Agent: Scores student responses against the rubric and supplies concise, actionable hints.
• Summary Agent: Compiles per-question results into a structured summary table.

Workflow bifurcation is based on student-reported code outcome:

• All Cases Passed Flow: Generator produces ~20 higher-order prompts (generalization, complexity, algorithm choice), Reviewer selects the 10 most salient.
• Partial/Failed Cases Flow: Generator synthesizes a correct reference solution, drafts ~10 targeted debugging prompts, Reviewer prioritizes the 5 most diagnosis-relevant.

The entire reflection cycle—upload, prompt presentation, response entry, feedback, and summary—is designed for completion in under five minutes.

3. Reflective Prompt Typology and Exemplars

Prompt engineering is tightly coupled to both the outcome of the student’s submission and established cognitive taxonomies. For correct code, the system emphasizes transferable understanding and in-depth analysis:

• “Justify your choice of a priority queue over a simple sort. Under what input distributions might your solution degrade?”
• “Analyze the worst-case time complexity. Could a more efficient data structure reduce it further?”

For incorrect or partially correct code, prompts target granular fault localization and revision planning:

• “Identify which test cases your code fails and pinpoint the logic gap causing the Wrong Answer verdict.”
• “Formulate a revision plan: which sections will you instrument with debug prints, and why?”

This adaptivity seeks to promote not just remediation but durable metacognitive skill acquisition.

4. Empirical Assessment, Limitations, and Recommendations

Deployment in three Purdue University CP courses (Spring 2025), covering second- and third-year undergraduates, demonstrated that many students and TAs found the LLM-generated prompts useful for reflection and debugging (Nieto-Cardenas et al., 13 Nov 2025). However, quantitative assessment results highlighted specific limitations:

• Prompt Quality (TA survey; N=7): Correctness M=62.1%, debugging usefulness M=3.4/5.
• Feedback Quality: Correctness M=49.9%, usefulness ≈2.8/5.
• Student Usefulness: High predicted adoption for novices (CP1; 87.5%), minimal for advanced (CP2/CP3).
• Satisfaction: Only half of prompts rated “high-quality;” user satisfaction below 50%.

Common challenges included LLM hallucinations, non-specific or redundant prompts, scalability issues with context windows, and user experience frustrations due to static, non-conversational interactions.

Recommended improvements incorporate retrieval-augmented generation, larger-context LLMs, dynamic conversational prompting, and tighter LMS integration. Generalization to other iterative computing courses and non-CS domains is posited, contingent on alignment of outcome-branching and rubric-driven evaluation.

5. Causal Intervention for Hallucination Mitigation in LVLMs

In a separate domain, Owlgorithm refers to a causally-motivated attention intervention framework for mitigating object hallucination in LVLMs (Yu et al., 12 Nov 2025). Here, “Owl” (Bi-mOdal attention reWeighting for Layer-wise hallucination mitigation) models multimodal generation and hallucination via a structural causal graph. Nodes include visual and textual inputs ($X_v, X_T$), priors ($P_v, P_T$), mediator attention weights ($A_v, A_T$), and output ($Y_T$), with manipulable mediators.

A key metric, the Visual-to-Textual Attention Contribution Ratio (VTACR), quantifies modality dominance during layer-wise decoding:

$\mathrm{VTACR}^l(e) = \frac{V^l(e)}{T^l(e)}$

where $V^l(e)$ and $T^l(e)$ are average visual/textual attention strengths for token $e$ at layer $l$. Empirical analysis finds low VTACR tokens highly susceptible to hallucination.

Interventions are implemented as token- and layer-wise adjustments to attention weights, governed by empirically determined thresholds $V_0^l$, culminating in dual-path contrastive decoding. Specifically, two decodings are computed per token: a visual-favored path and a text-favored path, with final token probabilities contrastively fused:

$$\ell^\text{final}_t = (1+\lambda)\,\ell^+_t - \lambda\,\ell^-_t$$

$$P_{\mathrm{DCD}(y_t)} = \mathrm{Softmax}(\ell^\text{final}_t)$$

This orchestration systematically amplifies visually grounded outputs while suppressing hallucinated content.

6. Performance Benchmarks and Comparative Results

Evaluated on POPE, CHAIR, and multiple VQA benchmarks (Yu et al., 12 Nov 2025), the Owlgorithm intervention achieved state-of-the-art hallucination mitigation with minimal compromise in VQA accuracy and often improved answer detailedness. Representative improvements include:

Model	CHAIR Cs ↓ (Orig → Owl)	CHAIR Ci ↓ (Orig → Owl)	POPE QA ↑ (Orig → Owl)
LLaVA-1.5	31.8 → 26.2 (−17.6%)	10.3 → 8.1 (−21.4%)	84.6/84.4/83.1 → 90.2/88.1/90.5
MiniGPT-4	24.8 → 21.2 (−14.5%)	9.3 → 6.2 (−36.7%)	69.2/68.8/67.4 → 82.2/78.4/79.0
Shikra	37.6 → 29.3 (−22.1%)	12.9 → 9.7 (−24.8%)	81.5/78.1/79.2 → 85.2/82.3/83.4

When evaluated by GPT-4V, gains in output correctness (e.g., LLaVA-1.5 from 5.58 to 6.70) and detailedness were observed across models.

7. Significance and Future Directions

Owlgorithm, in both its metacognitive educational and hallucination-mitigation forms, exemplifies the operationalization of advanced AI in structured reflection and reliability-critical applications. In education, future work focuses on precision improvement, interactivity, deployment scalability, and cross-domain adaptation (Nieto-Cardenas et al., 13 Nov 2025). In vision-language processing, continued refinement of causally-motivated interventions and integration with retrieval augmentation or larger-context models are prominent directions (Yu et al., 12 Nov 2025).

A plausible implication is that the shared design pattern—dynamic, context-aware response generation and intervention using symbolic and statistical models—may find broad utility across domains requiring trustworthy, explainable, and adaptive AI support.