The challenge begins with a fundamental mismatch. Large language models are built to process linear sequences of text, while tables are inherently two-dimensional structures. To make a table readable by a model, engineers must flatten it into text, usually in Markdown or JSON format. For a small table this works reasonably well. But for an industrial-scale table with hundreds of thousands of rows, the flattened version can exhaust the model’s context window entirely, leaving no room for the model to actually reason about the question. Even when the table fits, researchers have documented a phenomenon known as Loss in the Middle, in which models pay less attention to information buried in the middle of long inputs.

A team of researchers at the Institute of Artificial Intelligence (TeleAI) at China Telecom, led by Sishi Xiong and Ziyang He, set out to tackle these bottlenecks head-on. Their starting point was a simple observation: for any given question, the vast majority of cells in a large table are irrelevant noise. A question about average sales per publisher does not need the model to read every customer review column. Rather than forcing the model to digest the entire table, TableZoomer first builds a compact summary and then zooms in on only the parts that matter.

The framework’s first innovation is schema-based representation. Instead of feeding the model the full verbalized table, a component called the Table Describer parses the spreadsheet using the Pandas toolkit and generates a structured JSON schema. This schema captures each column’s data type, statistical properties such as maximum, minimum, mean and median values for numerical data, unique categories for categorical data, and a handful of randomly sampled cell values as examples. The result reduces the token complexity of the table representation from O(M times N), where M and N are the numbers of rows and columns, down to O(N), a reduction of orders of magnitude. Crucially, the schema is generated only once per table and can be reused across many questions.

Next comes the zooming mechanism, which the authors describe as query-aware table focusing. A Query Planner, guided by chain-of-thought prompting, decomposes the user’s question into atomic sub-queries and classifies them as either column-only queries, which can be answered from aggregate statistics, or row-column queries, which require identifying both relevant rows and columns. The Table Refiner then performs three operations: column selection to discard irrelevant features, entity linking to match names mentioned in the question with their surface forms in the table using a Longest Common Subsequence algorithm with an overlap threshold of 0.6, and table zooming to produce a minimal, query-relevant sub-schema. On the benchmark datasets, this column selection compressed the original column counts to just 5.2, 19.8 and 22.6 percent of their original sizes on DataBench, TableBench and WikiTableQA respectively.

Once the table has been zoomed, the framework turns to execution. Rather than asking the model to reason about numbers in natural language, a strategy prone to numerical hallucination, TableZoomer employs the Program-of-Thoughts approach, prompting the language model to generate executable Python code. The code runs in an isolated environment, and if execution fails, an error feedback mechanism returns the code and error traces to the model for self-correction. The entire workflow is orchestrated through the ReAct paradigm, an iterative cycle of thought, action and observation, with reasoning rounds capped at five. An Answer Formatter finally packages the result into whatever strict output format the task demands, from pure numbers to complete sentences.

The performance numbers are striking. When implemented with the Qwen3-8B-Instruct model, TableZoomer achieved accuracy improvements of 19.34 percent on the large-scale DataBench dataset and 25 percent on the Fact Checking task of TableBench compared with conventional Program-of-Thoughts methods, plus an 11.84 percent gain on TableBench’s Numerical Reasoning task. Perhaps most remarkably, the framework narrows the gap between models of different sizes: the TableZoomer-enhanced Qwen3-8B outperformed the much larger baseline Qwen3-32B model. For companies weighing the cost of deploying large models, this suggests that a well-designed agent framework can let smaller, cheaper models punch far above their weight.

Scalability proved to be another standout strength. The researchers sorted DataBench tables by cell count into large subsets averaging 1,519,065 cells, medium subsets averaging 18,131 cells and small subsets averaging 2,882 cells. Textual chain-of-thought reasoning degraded severely as table size grew, while code-based Program-of-Thoughts held up better. TableZoomer, however, showed minimal performance degradation across all three size categories, an attribute the authors attribute to the Table Describer and Table Refiner components excelling at locating essential information within massive tables. The framework also proved remarkably robust to how tables are serialized: while a baseline model’s accuracy fluctuated by up to 9.18 percentage points across Pandas-String, Markdown, JSON and structured formats, TableZoomer varied by only 0.57 points, staying between 86.59 and 87.16 percent.

Ablation studies on DataBench using Qwen3-8B confirmed which components matter most. Simply replacing the full-table Markdown text with the schema representation improved accuracy by 6.51 percentage points. Adding the table zooming mechanism then pushed accuracy from 74.33 to 86.40 percent, marking it as the framework’s core engine. Removing entity linking cost 1.73 points, and the ReAct reflection loop delivered a modest but consistent additional gain. Case studies illustrated the practical differences: where a baseline program-of-thoughts method misread a hierarchical Parent ID column and returned the number 150 instead of the category name Attractions, TableZoomer’s schema-level semantic annotations enabled an extra lookup that produced the correct answer. In another case, the framework’s denoising stage recovered a string-encoded dictionary field that stumped the baseline, correctly extracting an overall rating.

Efficiency analysis showed TableZoomer’s core workflow requires a minimum of five language model invocations, spanning the Table Describer, Query Planner, Code Generator, Answer Formatter and ReAct reflection, with the schema generated once and reused across multi-turn question answering. Competing frameworks often demand more: Chain-of-Table scales its inference count with the number of planning and execution operations, PoTable typically needs at least eight calls, and TableMaster at least six. On WikiTableQA, TableZoomer using LLaMA3.1-70B trailed only TableMaster, by 1.43 percent, while outperforming the second-ranked PoTable by 10.96 percent. The authors are candid about limitations: the predefined fixed workflow remains a bottleneck, and they plan to explore a central brain module that autonomously selects optimal action paths, as well as extending the framework toward vision-language models that could perceive table layouts directly. The source code is publicly available on GitHub, inviting the community to build on a framework that may well redefine how machines read the world’s spreadsheets.

Subject of Research: Large language model agent framework for table question answering over large-scale tabular data

Article Title: TableZoomer: a collaborative agent framework for large-scale table question answering

Article References: Xiong, S., He, Z., He, Z., Zhao, Y., Pan, C., Zhang, J., Song, S., & Li, Y. (2025). TableZoomer: a collaborative agent framework for large-scale table question answering. Vicinagearth, 2(1), Article 11. https://doi.org/10.1007/s44336-025-00016-x

Image Credits: AI Generated

DOI: 10.1007/s44336-025-00016-x

Keywords: table question answering, large language models, TableZoomer, agent framework, Program-of-Thoughts, ReAct, schema representation, entity linking, DataBench, TableBench, numerical reasoning, code generation

The core idea behind MR-UIE is a shift from treating large language models as passive extractors to training them as active reasoners. Rather than simply asking a model to map input text onto a fixed output structure, the framework teaches it to consider multiple analytical perspectives, weigh different reasoning strategies, and select the path most likely to yield a correct, faithful extraction. The researchers describe this as enabling models to understand not only what to extract but also how to reason, a distinction they argue is critical when texts are riddled with ambiguity, implicit causal links, and context-dependent information that simple pattern matching cannot resolve.

Technically, the framework rests on four pillars. The first is a unified schema representation, drawing on function-call semantics, that converts task-specific schemas for entities, relations, and events into a common hierarchical JSON structure with class identifiers, property-defining arguments, and semantic descriptors. This abstraction allows a single model to handle named entity recognition, relation extraction, and event extraction within one consistent output grammar, and it lets schemas for unseen tasks be projected into the same space at inference time, supporting zero-shot generalization without task-specific adapters.

The second pillar is the construction of a multi-perspective reasoning dataset, modeled on a human cognitive cycle of brainstorming, generating multiple solutions, and converging on the best one. For each input text and schema, a large language model acting as a strategy generator produces fine-grained thinking strategies along three dimensions: cognitive perspective, professional role, and heuristic rules. These raw strategies are then clustered into keyword-defined paradigms, ranked by TF-IDF embeddings and cosine similarity, and sampled to balance distinctiveness against representativeness. DeepSeek-R1 is then prompted to produce step-by-step chain-of-thought rationales together with structured predictions for each retained strategy, and rejection sampling discards any strategy whose prediction disagrees with the ground truth. Samples that fail to accumulate a minimum number of correct strategies are withheld from supervised training and routed instead to the reinforcement learning stage, ensuring the model learns from cases where reasoning demonstrably helps.

Training proceeds in three stages. The model, built on Qwen3-8B, first undergoes supervised fine-tuning on a curated corpus of roughly one million instances drawn from the bilingual IEPile collection, teaching it to interpret instructions and emit outputs that strictly conform to the unified schema. In the second stage, multi-perspective reasoning fine-tuning injects strategy prefixes into inputs and introduces special thinking tokens that segment reasoning steps, with a composite objective that optimizes both the chain-of-thought rationales and the structured predictions. To prevent overfitting to any single strategy, the team employs a regularization trick called strategy hiding, in which ten percent of training examples instruct the model to skip reasoning entirely. The final stage applies reinforcement learning using a rule-based Group Relative Policy Optimization procedure, in which the model generates multiple candidate reasoning chains per question and learns from a multi-grained reward function that scores both results and process.

The reward design is notable for its granularity. The result reward decomposes correctness into category-level and argument-level components, reweighted to reflect the fact that category prediction is easier than argument extraction, and combined as a weighted harmonic mean to encourage joint optimization. The process reward evaluates faithfulness, checking whether the reasoning strictly follows the schema, effectively uses the source text, and reflects a sound strategy. A convex combination of the two guides the policy, allowing the model to internalize effective reasoning patterns rather than memorize specific templates. During reinforcement learning, difficult samples are deliberately emphasized to raise the ceiling of model capability while leaving reasoning formats unconstrained, encouraging autonomous exploration of optimal strategies.

The experimental results are striking. Across eleven supervised benchmarks spanning newswire, finance, biomedicine, and general science, MR-UIE set new state-of-the-art Micro-F1 scores on five datasets, and on the remaining six never trailed the best published result by more than 3.2 F1 points. The wins span the full spectrum of extraction tasks: the framework topped the mit-restaurant dataset with its eight colloquial entity types and ambiguous boundaries, led on the terminology-dense biomedical AnatEM corpus despite limited training data, achieved the best overall scores on the skewed eighteen-category KBP37 relation benchmark, and ranked first on argument extraction for DUEE-fin, a financial event dataset with thirteen event types, ninety-one argument roles, and severe role overlap.

Zero-shot performance tells perhaps the more important story. Evaluated on datasets from domains and schemas never seen during training, spanning literature, music, law, and social media in both English and Chinese, MR-UIE achieved top-two results on most tasks and beat GPT-4 in five experimental settings, including a new state-of-the-art for zero-shot named entity recognition on CrossNER. The researchers attribute this generalization to the unified schema, which lets supervised capabilities transfer without structural adaptation, and to adaptive chain-of-thought reasoning that dynamically decomposes problems into finer-grained sub-steps, helping the model distinguish easily confused elements on schema-heavy datasets.

Ablation studies confirm that each component earns its place. Removing the unified schema cost between 0.42 and 0.72 F1 points across subtasks, while dropping multi-perspective reasoning training caused a maximum decline of 2.48 F1 points, demonstrating that explicit multi-angle fine-tuning is essential rather than decorative. Reinforcement learning alignment proved to internalize the diverse strategies rather than memorize them: when compared directly against static alternatives, the reinforcement-learned dynamic strategy beat random template selection, which actually degraded performance below a no-strategy baseline, and outperformed similarity-based template matching, whose effectiveness varied by task. Intriguingly, the training dynamics showed that as reinforcement learning progressed, generated reasoning chains shortened and converged to roughly 300 tokens while reward scores steadily climbed, suggesting that concise, efficient thinking benefits extraction performance more than lengthy elaboration.

The authors are candid about limitations, noting that the framework still struggles to dynamically integrate diverse reasoning paths and to handle highly noisy text, and they point toward noise-aware reasoning modules, contrastive path selection, refined reward design, and extension to multimodal settings as future directions. Even so, the work makes a compelling case that the road to robust structured understanding runs through reasoning itself. By rewarding models for faithful, efficient thinking and giving them multiple analytical lenses to think through, MR-UIE demonstrates that teaching language models how to reason may matter as much as teaching them what to know, a conclusion likely to shape how the next generation of extraction systems is built.

Subject of Research: Multi-perspective reasoning with reinforcement learning for universal information extraction

Article Title: MR-UIE: multi-perspective reasoning with reinforcement learning for universal information extraction

Article References: Li, Z., Wang, S., Fang, R., Bao, M., Wu, Z., Song, S., Li, Y., & He, Z. (2025). MR-UIE: multi-perspective reasoning with reinforcement learning for universal information extraction. Vicinagearth, 2(1), Article 17. https://doi.org/10.1007/s44336-025-00019-8

Image Credits: AI Generated

DOI: 10.1007/s44336-025-00019-8

Keywords: information extraction, multi-perspective reasoning, reinforcement learning, large language models, chain-of-thought, universal information extraction, named entity recognition, relation extraction, event extraction, GRPO, zero-shot generalization, schema representation