Models & Agents — Episode 5

Models & Agents

Date: March 02, 2026

FireRedTeam releases FireRed-OCR-2B, a 2B-parameter model tackling structural hallucinations in document parsing for tables and LaTeX.

What You Need to Know: The standout development today is FireRed-OCR-2B, a new open-source model from FireRedTeam that uses GRPO optimization to eliminate common errors in large vision-language models when handling complex document structures like tables and LaTeX. Meanwhile, a wave of arXiv papers introduces innovative multi-agent systems, from payment workflows and urban planning to suicide ideation detection, showcasing how LLMs are being integrated into agentic frameworks for real-world tasks. Pay attention this week to how these agent hierarchies could streamline your workflows in domains like healthcare and simulation, and test them against benchmarks for practical gains.

Top Story

FireRedTeam has released FireRed-OCR-2B, a 2B-parameter model designed to solve structural hallucinations in document parsing, particularly for tables and LaTeX, using Gradient-Response Prompt Optimization (GRPO). This model treats document parsing as a unified task, avoiding the multi-stage pitfalls of traditional LVLMs that lead to disordered outputs or invented elements. Compared to prior approaches, it offers better accuracy on structured data without needing separate layout detection and text extraction steps, making it a step up from models like those in the Florence or PaliGemma families for software developers dealing with code or scientific docs. Practically, this enables more reliable OCR for automating code reviews or extracting formulas from papers, so developers in data-heavy fields should care if they've struggled with hallucinated outputs. Keep an eye on community fine-tunes for domain-specific adaptations, and try integrating it into your pipelines via Hugging Face. What to watch: Potential expansions to multimodal agents that combine this with tools like LangChain for end-to-end document intelligence.

Source: marktechpost.com

Model Updates

FireRed-OCR-2B: MarkTechPost

FireRed-OCR-2B is a new 2B-parameter flagship model from FireRedTeam that uses GRPO to unify document digitization, fixing structural hallucinations in LVLMs for tables and LaTeX without multi-stage processing. It outperforms traditional methods on benchmarks by preserving order and syntax, comparing favorably to smaller models like MiniCPM-V but with specialized focus on developer tools. This matters for your work if you're building apps that parse code or scientific docs, as it reduces errors in automated extraction at low inference cost.

Source: marktechpost.com

Enhancing CLIP Robustness: cs.MA updates on arXiv.org

This paper introduces COLA, a training-free framework using optimal transport to improve CLIP's adversarial robustness via cross-modality alignment, boosting zero-shot classification by 6.7% on ImageNet variants under PGD attacks. It filters non-semantic noise and aligns image-text features better than fine-tuned baselines, addressing gaps in models like CLIP or Flamingo. For practitioners, this means more reliable VLMs in security-sensitive apps, though it requires augmented views for full effect.

Source: arxiv.org

Toward General Semantic Chunking: cs.CL updates on arXiv.org

A new discriminative model based on Qwen3-0.6B handles ultra-long documents for topic segmentation, supporting 13k-token inputs and compressing representations via vector fusion, outperforming Jina's Qwen2-0.5B models with faster inference. It beats generative LLMs by two orders of magnitude in speed while improving F1 on WIKI-727K, making it a practical upgrade over tools like Llama for long-context tasks. This is key for your RAG pipelines if you deal with massive texts, but it's optimized for English Wikipedia-style data.

Source: arxiv.org

Agent & Tool Developments

Hierarchical Multi-Agent System for Payments: cs.MA updates on arXiv.org

This introduces HMASP, a hierarchical LLM-based system with agents like CPA and Supervisor for end-to-end payment workflows, using MCP for secure data access and modular execution. It's a first for agentic payments, improving over static LLMs by handling multi-level coordination, and works with open or proprietary models like those from OpenAI or Claude. Try it today via the open-source code for prototyping financial agents, though it requires LLM API access for full functionality.

Source: arxiv.org

Rudder for Prefetching in GNN Training: cs.MA updates on arXiv.org

Rudder embeds LLM agents in AWS DistDGL for adaptive prefetching in distributed GNN training, using in-context learning to cut communication by 50% and boost performance up to 91% over baselines. It adapts to dynamic graphs better than static methods, integrating with frameworks like AutoGen for agentic control. Developers can test the GitHub repo on Perlmutter-like setups for graph ML tasks, but it's tailored to large-scale infra.

Source: arxiv.org

City Editing with Hierarchical Agents: cs.MA updates on arXiv.org

This presents a hierarchical agent framework for urban geospatial editing, using LLMs to handle dependency-aware modifications in GeoJSON, outperforming baselines in efficiency and validity on renewal tasks. It enables iterative planning with error mitigation, extending tools like LangGraph for spatial agents. Explore the code for urban simulation projects, noting it excels in multi-step edits but assumes clean input data.

Source: arxiv.org

Resp-Agent for Respiratory Diagnosis: cs.MA updates on arXiv.org

Resp-Agent is an agentic system with Thinker-A²CA for multimodal respiratory sound generation and diagnosis, using LLMs to bridge data gaps and reduce biases on the new Resp-229k dataset. It outperforms baselines in scarcity scenarios, complementing agent frameworks like CrewAI for health AI. Test it on GitHub for audio-text tasks, but it's compute-intensive for fine-tuning.

Source: arxiv.org

Practical & Community

Integrating LLMs in Social Simulation: cs.MA updates on arXiv.org

This position paper proposes Hybrid Constitutional Architectures for agent-based simulations, combining ABMs like GAMA/NetLogo with SLMs and LLMs to balance flexibility and transparency. It solves issues in benchmarks like Smallville by addressing biases and scalability, ideal for researchers in multi-agent systems. Get started by experimenting with the outlined integrations in NetLogo; limitations include undertraining risks in LLMs.

Source: arxiv.org

Mixed-Initiative Dialog for Collaboration: cs.MA updates on arXiv.org

MICoBot is a mixed-initiative system for human-robot dialog, using planners and affordance models to minimize human effort in tasks, outperforming LLM baselines in physical trials. It tackles collaborative manipulation with natural language, useful for robotics devs integrating Semantic Kernel. Try the UT Austin demo for task planning; it requires robot hardware for full tests.

Source: arxiv.org

CiteAudit Benchmark for Citation Verification: cs.CL updates on arXiv.org

CiteAudit is a 1.3M-sample benchmark with a multi-agent pipeline for detecting hallucinated citations in scientific writing, outperforming LLMs in accuracy via claim extraction and reasoning. It addresses LLM risks in research, complementing RAG tools for verification. Start with the dataset for fine-tuning evaluators; it's domain-agnostic but focused on academic texts.

Source: arxiv.org

Things to Try This Week

• Try FireRed-OCR-2B via Hugging Face for parsing LaTeX docs in your code pipeline — it's a low-cost way to cut hallucinations compared to standard LVLMs.
• Experiment with Rudder's LLM agents in DistDGL for GNN training if you're on AWS — it could slash your communication overhead in distributed setups.
• Integrate HMASP into a payment prototype using Claude or GPT APIs — see how hierarchical agents handle real workflows without custom coding.
• Test Resp-Agent on the Resp-229k dataset for health audio tasks — it's worth exploring if you're building diagnostic tools with multimodal LLMs.
• Compare COLA's robustness enhancements on CLIP for your image-text apps — it boosts adversarial defense without retraining, unlike fine-tuned alternatives.

On the Horizon

• Watch for expansions of FireRed-OCR-2B to broader multimodal tasks, potentially integrating with agent frameworks like LangChain.
• Expect more arXiv follow-ups on hierarchical agents, building on HMASP for domains beyond payments like logistics.
• Anticipate benchmarks evolving from CiteAudit, possibly incorporating MCP for standardized LLM evaluation in research tools.
• Look out for real-world deployments of systems like MICoBot in collaborative robotics, with betas in frameworks like Microsoft AutoGen.

Sources

marktechpost.com arxiv.org