Even Correct Answers Score Zero: TRACE System Exposes Scoring Challenges in AI Scientific

Recently, the GroundLM 2026 TRACE system released a review report, with the LitTraceQA subtask drawing significant attention. The core requirement is that models must not only provide answers to scientific questions but also simultaneously output standardized paper IDs, page numbers, and citations.

Back to news

Tech Insights 好游模型Published Nov 6, 2026 151 views

Overview

Recently, the GroundLM 2026 TRACE system released a review report, with the LitTraceQA subtask drawing significant attention. The core requirement is that models must not only provide answers to scientific questions but also simultaneously output standardized paper IDs, page numbers, and citations. In other words, a complete response demands dual alignment of "answer + source."

However, evaluation results revealed a controversial phenomenon: even when the answer content is factually correct, models may receive zero points due to non-standard formatting, missing citations, or unverifiable sources. In short, "correct answers still score zero" directly summarizes the stringent standards in current AI scientific QA evaluation mechanisms. This TRACE review actually exposes a long-overlooked aspect of the field: answer correctness does not equate to usability or credibility.

Accompanying image

From the task designers' perspective, LitTraceQA's strict requirements are not meant to penalize models arbitrarily but stem from a pursuit of scientific rigor. In real research scenarios, a conclusion's value lies not just in "what" it states, but "where" it originates. Precise labeling of paper IDs and page numbers is crucial for determining if information is traceable and reproducible. Thus, TRACE aims to push AI models from "capable of answering" to "answering credibly" through this dual assessment of "answer + citation."

Illustration 1

Background

Understanding why TRACE emphasizes citation norms requires tracing the development of AI scientific QA. As large language models improve, more researchers are using AI to assist with literature retrieval, knowledge Q&A, and scientific reasoning. Early models often produced plausible answers but struggled to guarantee reliable sources behind them. This phenomenon of "confidently stating nonsense" has long hindered the practical deployment of AI in research.

GroundLM, focusing on "grounded" language models, aims to ensure models rely on verifiable knowledge sources rather than fabricating content. The 2026 TRACE shared task represents a systematic attempt in this direction. LitTraceQA refines the "traceable citation" concept into quantifiable, evaluable metrics: precise output of paper IDs and page numbers.

Macroscopically, this trend aligns with the broader AI industry's pursuit of "trustworthy AI." Whether driven by academic reproducibility requirements or industrial regulatory pressures for accuracy, the field is shifting from "fluent generation" to "credible generation." TRACE's strict citation assessment embodies this shift in scientific QA, marking an upgrade in evaluation standards from single "answer correctness" to dual dimensions of "answer + source."

Illustration 2

Impact

The TRACE review report has generated multi-layered impacts on the AI scientific QA field.

For the industry, the report highlights an undervalued issue: answer correctness does not automatically translate to usability. When models cannot stably output standardized paper IDs and page numbers, their scientific value diminishes significantly. This prompts evaluators, developers, and research institutions to redefine "high-quality scientific QA," driving evaluation standards toward greater rigor and traceability.

For users, the "correct answers score zero" phenomenon means adjusting expectations. When relying on AI for research assistance, users cannot judge reliability solely by answer fluency but must verify citation norms and source traceability. It also reminds users that AI research tools still need refinement in "credibility" and cannot yet replace rigorous academic verification processes.

For the market, this trend may reshape the competitive landscape of AI research products. Models and products that stably output standardized citations and achieve dual alignment of "answer + source" will gain a competitive edge. It also points to key development areas: balancing answer quality with citation accuracy and standardization will become a critical research focus.

In the long term, the "traceability" philosophy represented by TRACE could drive the entire AI research ecosystem toward greater transparency and rigor. When models learn to cite sources for every conclusion, AI's role in research becomes clearer—not a black box "spitting out answers," but an "evidence-based" auxiliary tool. This shift marks a crucial step toward truly trustworthy AI.