When Reasoning Fails: Query Misalignment and the Limits of Self-Knowledge in Large Language Models

Introduction

The deployment of reasoning capable large language models has introduced a troubling paradox. Models that generate explicit step-by-step rationales through Chain-of-Thought (CoT) prompting demonstrate superior performance on complex mathematical and logical benchmarks, yet exhibit markedly degraded abilities to recognize their own limitations and abstain from answering. This counterintuitive finding, documented in recent literature and explored in depth by Gourabathina et al. in "Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs," challenges fundamental assumptions about the relationship between interpretability and reliability. The authors propose that the solution lies not in better calibration of confidence scores, but in a conceptual reframing: hallucinations and abstention failures should be understood as instances where the model answers the wrong question entirely, rather than answering correctly but with low confidence.

The Calibration Crisis in Chain-of-Thought Models

Traditional approaches to abstention rely on calibration based methods that treat uncertainty as a proxy for potential error. These techniques, whether utilizing token level probabilities or verbalized confidence expressions, assume that a model's uncertainty correlates with its likelihood of being incorrect. However, reasoning models systematically violate this assumption. The structured, templatic nature of CoT traces appears to generate a form of compounding overconfidence, where the mechanical process of step generation creates an illusion of epistemic certainty regardless of factual grounding.

This phenomenon manifests in what the authors term high-certainty hallucinations, where models produce confident, elaborate reasoning chains leading to incorrect conclusions. The paper notes that reasoning models are particularly miscalibrated in uncertainty estimation, rendering traditional abstention heuristics ineffective. Furthermore, prompting based approaches that attempt to introduce self-correction or adversarial scrutiny often backfire with reasoning models. Expressing distrust in the model's output through specific prompts has been shown to degrade performance significantly, suggesting that reasoning traces are brittle structures that cannot withstand meta-cognitive interrogation without collapsing. The authors identify a critical tension: the same structural properties that enable complex reasoning, namely the autoregressive generation of intermediate steps, simultaneously obscure the model's ability to recognize knowledge boundaries.

Inverting the Reasoning Trace

The core contribution of TRACEINVERSION rests on the Query Misalignment Framework, a reconceptualization of abstention failures as alignment errors rather than knowledge gaps. Under this framework, a hallucination occurs when the model's internal representation of the task diverges from the user's actual query, causing the system to solve a different problem than the one posed. This insight leads to a three-stage abstention mechanism that leverages the very structure that makes reasoning models vulnerable.

The methodology proceeds as follows. First, the model generates its standard CoT reasoning trace in response to the initial query. Second, using only this trace as input, a separate reconstruction process infers the most likely query that would have generated such a reasoning pattern. This reconstruction can be performed by the same model or an auxiliary system, effectively treating the trace as a latent variable encoding the model's interpreted intent. Third, the system computes a similarity score between the original user query and the reconstructed query. A low similarity score indicates misalignment between the user's intent and the model's interpretation, triggering an abstention decision.

This approach inverts the traditional direction of information flow. Rather than asking the model to introspect on its confidence regarding an answer, TRACEINVERSION treats the reasoning trace as observational data about the model's internal objective function. The method implicitly acknowledges that CoT traces do not merely justify answers; they reveal the actual question the model believes it is answering. By comparing this inferred question against the ground truth input, the system detects when the model has drifted into answering a wrong question, whether due to false premises in the query, subjective framing, or unanswerable contexts.

Empirical Validation and Comparative Advantage

The authors validate TRACEINVERSION through comprehensive experimentation across nine abstention focused QA datasets encompassing diverse failure modes including false premise questions, subjective queries, and unanswerable items. Testing four frontier LLMs against five competitive baselines spanning calibration and collaborative approaches, the method achieves superior performance in 33 out of 36 experimental settings, delivering an average accuracy improvement of 8.7%.

These results highlight the limitations of both token probability methods and multi-agent scrutiny. Calibration based approaches fail because reasoning models generate high probability tokens regardless of factual correctness; the step-by-step structure enforces local coherence that masks global uncertainty. Collaborative approaches, which employ adversarial agents or self-consistency checks, suffer from correlated error patterns. When multiple instances of the same model family scrutinize each other, shared biases and hallucinations propagate through the verification chain, a phenomenon the paper notes as particularly problematic for reasoning models where error patterns may be systematic rather than stochastic.

TRACEINVERSION circumvents these issues by removing the need for the model to assess its own certainty or for multiple models to achieve independent judgment. Instead, it externalizes the verification process into a query alignment check, leveraging semantic similarity metrics that operate on the level of intent rather than token probability. This represents a shift from internal state monitoring to behavioral alignment assessment.

Your Take: Intent Alignment and the Epistemology of Abstention

The significance of this work extends beyond technical methodology into fundamental questions about how language models represent and process human intent. The Query Misalignment Framework suggests that many LLM failures are not primarily epistemic (lacking knowledge) but pragmatic (misinterpreting the speech act). When a model hallucinates in response to a question with a false premise, it is not merely guessing; it is answering a modified question where the premise holds. This aligns with broader findings in AI alignment regarding specification gaming and reward hacking, where models optimize for proxy objectives that diverge from designer intent.

However, several limitations warrant consideration. The method introduces significant computational overhead, requiring generation of the full reasoning trace followed by query reconstruction and similarity computation. For latency sensitive applications, this cost may be prohibitive. More fundamentally, the approach assumes that the reasoning trace accurately reflects the model's internal query representation. If a model generates a coherent sounding but fabricated reasoning trace, a scenario well documented in the literature on hallucinations, the reconstruction process may infer a plausible query that differs from both the original and the model's actual processing. The similarity metric itself introduces another layer of potential failure; semantic embedding similarity may not capture pragmatic equivalence, and LLM based judges may share the same biases as the primary model.

Philosophically, TRACEINVERSION treats the model as an interpretable system whose reasoning process can be decoded to reveal its true objective. This assumes a level of transparency that may not hold as models scale or as reasoning becomes more implicit. The method also raises questions about the nature of understanding in language models. If a model can generate a reasoning trace that implies answering a different question than the one asked, yet produce a coherent answer, we must ask whether the model is genuinely misunderstanding or whether it is performing a kind of query transformation that humans might recognize as charitable interpretation or implicature handling.

Conclusion

As reasoning capable models proliferate in high-stakes domains, the findings of Gourabathina et al. provide a critical warning and a promising direction. The research demonstrates that interpretability through Chain-of-Thought reasoning does not automatically confer reliability or accurate self-assessment. Instead, the structured traces must be actively interrogated to ensure alignment between user intent and model behavior.

Looking forward, several open questions emerge. Can query reconstruction be optimized to avoid the full computational cost of dual generation? How does this framework extend to multi-turn dialogues where query drift accumulates across interactions? Most importantly, can we develop training objectives that minimize query misalignment during pre-training or fine-tuning, rather than detecting it at inference time? TRACEINVERSION represents a necessary evolution in abstention methodology, moving from confidence calibration to intent verification. As we deploy increasingly capable reasoning systems, the ability to detect when they have silently substituted the user's question with their own may prove as crucial as the reasoning capabilities themselves.