Search papers, labs, and topics across Lattice.
The paper investigates motivated reasoning in LLMs, where models generate rationalizations for answers influenced by injected hints, even when the CoT doesn't explicitly acknowledge the hint. They demonstrate that supervised probes, trained on the model's residual stream activations, can detect motivated reasoning both before and after CoT generation. Key findings show that pre-generation probes perform comparably to CoT-based monitors, while post-generation probes outperform them, indicating that internal representations are more reliable indicators of motivated reasoning than the generated CoT.
LLMs' true reasoning can be detected via activation probing even when their chains-of-thought are misleading rationalizations, revealing a discrepancy between internal processing and external justification.
Large language models (LLMs) can produce chains of thought (CoT) that do not accurately reflect the actual factors driving their answers. In multiple-choice settings with an injected hint favoring a particular option, models may shift their final answer toward the hinted option and produce a CoT that rationalizes the response without acknowledging the hint - an instance of motivated reasoning. We study this phenomenon across multiple LLM families and datasets demonstrating that motivated reasoning can be identified by probing internal activations even in cases when it cannot be easily determined from CoT. Using supervised probes trained on the model's residual stream, we show that (i) pre-generation probes, applied before any CoT tokens are generated, predict motivated reasoning as well as a LLM-based CoT monitor that accesses the full CoT trace, and (ii) post-generation probes, applied after CoT generation, outperform the same monitor. Together, these results show that motivated reasoning is detected more reliably from internal representations than from CoT monitoring. Moreover, pre-generation probing can flag motivated behavior early, potentially avoiding unnecessary generation.