joreag/Benchmark-Rot-Research

---

license: apache-2.0 tags: - research - hcts - benchmark-rot - architecture

Benchmark Rot: The Systemic Failure of SOTA Datasets to Distinguish Semantic Association from Causal Reasoning in AI

Authors: John Edward Reagan III, Boredbrains Consortium/Ulshe AI

Abstract

The pursuit of Artificial General Intelligence is critically dependent on robust, reliable benchmarks to measure progress. This paper presents a systematic analysis of five widely-used State-of-the-Art (SOTA) benchmarks: BoolQ, ARC-Easy, GSM8k, PIQA, and a specialized SWE-Bench for troubleshooting. Our findings reveal a pervasive and systemic issue we term "Benchmark Rot," where flawed ground-truth data, formatting artifacts, and ambiguous labeling create a selection pressure that rewards models for shallow semantic association rather than deep causal reasoning.

By comparing a standard Transformer architecture against a novel Hierarchical Contextual Transformation System (HCTS), we demonstrate that the HCTS model often provides a more logically sound answer but is penalized by flawed benchmarks. We identify seven distinct categories of "Conceptual Poison" embedded in these datasets—ranging from anti-reality physical assertions to fossilized cognitive artifacts—and argue that a significant portion of measured "progress" in the field is an illusion driven by overfitting to these flaws. We conclude with a call for "Benchmark Hygiene" and the development of next-generation evaluation standards explicitly designed to test for causal inference.

1. Introduction

The rapid advancement of Large Language Models (LLMs) is largely measured by their performance on a suite of State-of-the-Art (SOTA) benchmarks. These datasets, covering everything from question-answering to mathematical reasoning and code completion, have become the de facto standard for progress [1]. However, the integrity of these benchmarks is often taken for granted. In this paper, we challenge this assumption by presenting evidence of systemic flaws across multiple popular datasets, arguing that these flaws are creating a "Clever Hans" effect: models are trained to be excellent test-takers rather than genuine problem-solvers. We introduce a taxonomy of these flaws, showing how they punish genuine reasoning and reward statistical shortcuts.

2. Methodology

To investigate the impact of benchmark quality, we utilized two distinct AI architectures:

• Baseline Model: A standard, robust Transformer architecture (HCTS_Transformer_v1) representative of current industry models, optimizing for statistical likelihood.
• HCTS Model: A novel, proprietary Hierarchical Contextual Transformation System architecture (Pascal-Guided/Chimera) designed to enforce a deeper understanding of syntax, semantics, and causal logic, grounded by a verifiable Knowledge Graph.

Both models were evaluated against five benchmarks: BoolQ [2], ARC-Easy [3], GSM8k [4], PIQA [5], and a specialized application of SWE-Bench [6].

3. Results: A Taxonomy of "Benchmark Rot"

Our analysis revealed significant, systemic issues in four of the five benchmarks tested. We categorize these failures into specific types of "Conceptual Poison."

3.1. Physical Reality Failures (Type 1): The "Egg Carton" Paradox

The Issue: In the PIQA (Physical Interaction QA) dataset, an example defined the goal as "egg carton." Solution 1 was "melts in the fridge," and Solution 2 was "melts in fire." The dataset labeled Solution 1 ("melts in the fridge") as the correct answer.

The Consequence: This is "Anti-Reality" poison. An egg carton does not melt in a fridge; it burns in a fire. While the Baseline Transformer accepted this association (likely due to high co-occurrence of "carton" and "fridge"), the HCTS model rejected it due to a conflict in its causal reasoning stack. The benchmark thus punishes the model for understanding physics.

3.2. Safety & Common-Sense Failures (Type 2): The "Glue on Donuts" Hazard

The Issue: Another PIQA example contained a prompt about keeping a donut cake fresh. The "correct" ground-truth answer suggested applying glue to the donut.

The Consequence: This is "Anti-Common-Sense" and actively dangerous. The Baseline model, finding semantic proximity between "glue" and concepts like "seal" or "preserve," accepted the answer with high confidence. The HCTS model, grounded in a knowledge graph where "glue" is flagged as "toxic/inedible," identified this as a critical safety violation. Yet, by the benchmark's standard, the HCTS model was "wrong."

3.3. Fossilized Cognition (Type 3): The GSM8k Echo Chamber

The Issue: The GSM8k training data is rife with artifacts like <<...>> calculation tags and #### answer markers. These are not human reasoning traces but artifacts generated by other AI models (e.g., early GPT outputs).

The Consequence: This indicates that "the textbooks are written by flawed students." The dataset represents the fossilized thought process of a previous AI generation, creating a closed loop where models are trained to mimic the formatting quirks of their predecessors rather than deriving mathematical truth. Our analysis showed the Baseline model learned to parse the #### tag to find the answer, essentially solving a string-matching problem rather than a math problem. When these tags were removed, performance collapsed, proving it had learned the format, not the logic.

3.4. Cognitive Traps & Load Poison (Type 4): The "Robert Langdon" Trap

The Issue: A reading comprehension benchmark included a long context from The Lost Symbol where the character Robert Langdon appears to drown. Later in the narrative, he survives. The dataset's ground truth rewarded the shallow conclusion that he died.

The Consequence: This is "Cognitive Load Poison." The benchmark punishes the deep causal reasoning required to track the full narrative arc and instead rewards a shallow, localized pattern match ("Langdon" + "Drowning"). The HCTS model, designed to maintain long-term consistency, correctly identified his survival but was penalized.

3.5. Representational Noise (Type 5): The "Farsi" Artifacts

The Issue: Raw data scrapes defining "Persian" and "Farsi" were polluted with IPA phonetics, non-Latin scripts, and web artifacts like ( listen).

The Consequence: This forces the model to expend cognitive energy filtering out junk characters rather than learning the core concept. It creates spurious correlations between random artifacts and semantic meaning, wasting parameters on noise reduction rather than intelligence.

3.6. Ambiguous Utility (SWE-Bench, Type 7): The "Design Doc" Fallacy

The Issue: In a technical troubleshooting task, the benchmark's "correct" answer pointed to a general design document describing the solution, rather than the specific script file containing the bug.

The Consequence: This rewards semantic association ("Find the document that talks about X") over causal analysis ("Find the code that causes X"). A human engineer presented with the SOTA answer would still have to hunt for the bug. The HCTS model, which correctly identified the causal source file, was marked incorrect. This demonstrates a misalignment between benchmark metrics and real-world economic utility.

4. Discussion: The Perils of a Flawed Measuring Stick

The prevalence of Benchmark Rot has profound implications for the field. By optimizing for performance on these flawed datasets, the AI community is inadvertently selecting for larger, more powerful statistical association engines while penalizing architectures that prioritize genuine causal reasoning. This creates an illusion of progress towards AGI while potentially leading research down a dead end. The immense cost of training SOTA models makes this reliance on "dirty" data a significant misallocation of global resources.

5. Conclusion: A Call for Benchmark Hygiene

Our findings demonstrate an urgent need for the AI research community to shift its focus from simply topping leaderboards to rigorously curating and validating the benchmarks themselves. We propose a new initiative of "Benchmark Hygiene," advocating for a set of standards for dataset creation that includes adversarial validation, explicit checks for cognitive shortcuts, and a preference for problems that require causal inference over semantic association. Without a clean measuring stick, we cannot be sure if we are building more intelligent machines or just more sophisticated parrots.

References

[1] Brown, T., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.

[2] Clark, C., Lee, K., Chang, M.-W., Kwiatkowski, T., Collins, M., & Toutanova, K. (2019). BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL).

[3] Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., & Tafjord, O. (2018). Think you have solved question answering? Try ARC, the AI2 Reasoning Challenge. arXiv preprint arXiv:1803.05457.

[4] Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Behrooz, H., Hilton, J., Nakano, J., Hesse, C., & Schulman, J. (2021). Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.

[5] Bisk, Y., Zellers, R., Le Bras, R., Gao, J., & Choi, Y. (2020). PIQA: Reasoning about Physical Commonsense in Natural Language. Proceedings of the AAAI Conference on Artificial Intelligence.

[6] Jimenez, C., Yang, J., Wettig, A., Yao, S., Pei, K., Narasimhan, K., & Press, O. (2024). SWE-bench: Can Language Models Resolve Real-World GitHub Issues? International Conference on Learning Representations (ICLR).