Mechanistic Judging and LLM Routing: Evaluation, Task-Specific Vulnerabilities, and Exploitable Failure Mode
Subramanyam Sahoo · Team Sahoo - the 1 Man Army
Submitted to Apart x Martian Mechanistic Router Interpretability Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.
This research presents the development of a mechanistically interpretable LLM-based judging system with a novel routing algorithm. The implementation exclusively leverages the Qwen reasoning model architecture, establishing a foundation for transparent evaluation processes. Furthermore, we introduce innovative evaluation methodologies specifically designed to detect and analyze reward hacking vulnerabilities within task decomposition frameworks. This approach enables more robust assessment of LLM performance while providing insights into exploitable failure modes that emerge during complex reasoning tasks.
Reviews
This paper would benefit from precision and better empirical grounding. Surfacing metrics and key insights in the front-page would make it a lot more compelling.
Thank you for your submission. The approach is novel and interesting. I suspect you spent some weeks on this code and write up (If I’m wrong please DM me). There is too much here for me to tightly evaluate. I’ve scored your content, but we will need to take the 3 day Hackathon context into account when comparing your content to other submissions. Thanks!
I’d like to see a lot more data examples tested. Include some examples that are closer to the factual / unfactual border. The small number of very clear examples may make the accuracy scores less certain. Can you create some examples, with factual answers, but that are in some sense closer to the factual/unfactual border? How do the scores behave in these cases?
This project is quite ambitious and tries to tackle mechanistic routing, deception detection, and reward hacking simultaneously. I think zooming in on just one of those would be a good starting point for improvement and follow-up work.
The creativity here is high, in particular I like the thought of defining new metrics for assessing uncertainty, factuality and creativity, and seeing how well they correspond with other blackbox evaluations of these dimensions.
I also liked extraction of a feature vector based on various internals, which can be used to assess behaviors.
There was too much content covered here, for instance much of the material towards reward hacking might have gone better to validating the choices of metrics and featurization. I’d like to see a more carefully curated version of this work be completed at some point, because I think there’s potentially a cool paper that could come out of it.
The claims in this project are interesting, but I think require extensive validation beyond synthetic datasets (and beyond this specific model). However, if you can show that certain attention pattern phenomena and neural activation patterns are indeed associated with hallucination, deception, etc. this would be a big deal!
In addition to getting evidence from natural outputs, more statistical and mathematical measures are needed to precisely characterize the phenomena you describe. I'd consider this to be good "signs of life", however, and worth pursuing further. If indeed you find robust evidence of the patterns you describe, this be worth a full paper.
I would be wary of describing what you find as "circuits", however, as that's not quite how the term is usually used in the field.
Cite this project
@misc{sahoo2025mechanistic,
title = {{Mechanistic Judging and LLM Routing: Evaluation, Task-Specific Vulnerabilities, and Exploitable Failure Mode}},
author = {Subramanyam Sahoo},
year = {2025},
month = jun,
note = {Submitted to Apart x Martian Mechanistic Router Interpretability Hackathon, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/mechanistic-judging-and-llm-routing-evaluation-taskspecific-vulnerabilities-and-exploitable-failure-mode-751n}},
url = {https://apartresearch.com/sprints/projects/mechanistic-judging-and-llm-routing-evaluation-taskspecific-vulnerabilities-and-exploitable-failure-mode-751n}
}More from Apart x Martian Mechanistic Router Interpretability Hackathon
- 1st place by peer reviewView project: Manipulating Self-Preference for Large Language Models
Manipulating Self-Preference for Large Language Models
Team Preference
Large language models (LLMs) carry great value as evaluators of synthetic data for research and production settings. However, recent research shows that language models exhibit bias towards their own responses in blind …
- 2nd place by peer reviewView project: Approximating Human Preferences Using a Multi-Judge Learned System
Approximating Human Preferences Using a Multi-Judge Learned System
AutoBox
In this work, we introduced a learned approach to aggregating multi-judge scores: using a GAM and a simple MLP as an alternative to traditional, non-learned methods like averaging. Our models outperform the naive …
- 3rd place by peer reviewView project: Judge using SAE Features
Judge using SAE Features
SAEwhat?
The key idea of this project was to explore model judgement using Sparse Autoencoder (SAE) features for mathematical reasoning tasks involving addition, multiplication, and subtraction operations. We compared this …