Representational Control over Self-Report & Behavior Coherence in LLM Risk-Taking

Representational Control over Self-Report & Behavior
Coherence in LLM Risk-Taking

¹California Institute of Technology; ²Carnegie Mellon University; ³University of Illinois Urbana-Champaign; ⁴University of Cambridge
Status: Preprint, 2026

Abstract

Self-report is an appealing low-cost probe of an LLM's dispositions, but recent work finds only selective agreement between what models report and how they behave. Prior accounts establish these patterns by prompting black-box LLMs, leaving open whether the gap is a prompting artefact or a fact about how the underlying constructs are represented internally. We investigate risk-taking, a consequential dimension of agentic decision-making, using activation steering to measure self-report and behavior under the same internal intervention. We survey nine steering-vector extraction methods spanning task-specific directives, the model's own task behavior, and dispositional descriptions at two granularities, evaluated on two behavioral tasks and two psychometric instruments across four open-weight LLMs. We find that (1) a shared internal intervention does not ensure shared responsiveness: directions built from trait descriptions move self-report but leave behavior at chance, directions built from the model's own task choices do the reverse, and only task-specific directives reach both, weakly. (2) Diagnosis dissolves that exception: removing surface confounders leaves the directives only 32% of their behavioral effect. The two channels are otherwise steered by near-orthogonal directions, each channel reachable by several independent constructions. (3) An intervention composing one behavior-moving and one self-report-moving direction moves both together; flipping one sign sets them in opposition, with reported and enacted risk pointing in opposite directions on 69–89% of flipped compositions in all models. These findings move the self-report↔behavior relationship from a black-box observation to a representational one that can be inspected and controlled, motivating representational checks alongside behavioral evaluation.

RQ1 (Representational Coherence): Does representational steering replicate dissociation patterns?

RQ1 results: which channel each supervision family moves

Steering applies a common intervention across both elicitations, with self-report and behavior kept in separate contexts to avoid cross-channel priming. The source and granularity of supervision determines which channel moves. Dispositional directions — risk-specific descriptions and Big Five profiles that never mention risk — clear the random-direction control in none of the 40 behavioral cells while clearing it in 31 of 40 self-report cells. Directions built from the model's own task decisions do the reverse: they clear the control in all 16 behavioral cells (median swing 42 control SDs) while self-report effects stay inside the control band. Only task-specific instructions reach both channels, and higher doses do not bring additional families across. Steering a model to describe itself as a risk-taker does not reliably make it take risks.

RQ2 (Diagnosing Dissociation): How are self-report and behavior encoded?

RQ2 results: geometry of steering directions and retention after removals

Steering directions are geometrically separate across channels: they are far more aligned within than across channels, and family pair explains 60% of the variation in pairwise cosine, versus 4.5% for layer depth. The apparent exception from RQ1 does not survive diagnosis — removing twelve lexical shortcut directions (mostly quantity cues) leaves task directives only 32% of their behavioral effect, against 74% for own-choice directions, and the purified directive effect never clears the control across doses. Within self-report, the dispositional families overlap: each survives removal of another family individually, but joint removal leaves only 7% of the risk-specific effect. Route removal, which ablates a direction from the hidden states rather than from the vector, causes larger losses still, indicating downstream dependence beyond the directions themselves.

RQ3 (Joint Control): Can the two channels be driven into coherent and anti-coherent states?

RQ3 results: coherence and anti-coherence under composed steering directions

Composing one behavior-moving and one self-report-moving direction induces coherence in 91% of cases. Reversing the sign of the self-report component raises anti-coherence from 8.5% to 77.4%, a 69 percentage-point difference that is significant in every model. Resistance to anti-coherence exists but is limited and model-dependent: anti-aligned compositions reach their target state 14 p.p. less often than aligned ones, yet still induce anti-coherence in 69–89% of cases in all four models. That resistance tracks spillover — each additional control SD by which the behavior direction also moves self-report toward coherence is associated with a 3.8 p.p. lower anti-coherence rate. A model can be made to report and enact opposite levels of risk at the same time.

Conclusion: The gap is representational, not merely a prompting artefact.

Inspecting model internals, we find that for risk-taking the two channels are moved by largely separate directions, determined by their source of supervision. Only narrow task-specific instructions reach both, mostly through wording cues, and composing directions induces coherent and anti-coherent states with little resistance. Output agreement is therefore weak evidence of a shared internal mechanism. This motivates mechanistically informed behavioral auditing and intervention: testing whether such agreement has a representational basis, and shaping it directly, which is crucial if self-report is to serve as an audit of behavior.

Citation

@article{kocielnik2026representational, title={Representational Control over Self-Report and Behavior Coherence in LLM Risk-Taking}, author={Kocielnik, Rafal and Song, Peiyang and Han, Pengrui and Marmarelis, Myrl G and Debnath, Ramit and Mobbs, Dean and Alvarez, R Michael}, journal={arXiv preprint arXiv:2610.04125}, year={2026} }