Stability and Legibility Account
- The Stability-and-Legibility Account defines systems that achieve robust performance while clearly conveying intent through human-interpretable, measurable criteria.
- It bridges diverse domains such as robotics, human-centered AI, and prediction markets by using tailored metrics like posterior goal inference and transfer utility.
- Research highlights trade-offs between efficiency and explanatory clarity, advocating context-sensitive controls to balance performance with observer-friendly outputs.
Searching arXiv for the cited papers and closely related work to ground the article. arxiv_search(query="(Lúčny et al., 7 Aug 2025) OR (Wallkotter et al., 2022) OR (Faria et al., 2022) OR (Kirchner et al., 2024) OR (Seth et al., 2023) OR (Hemment et al., 2023) OR (Sreevatsa et al., 9 Feb 2026) OR (Crétides et al., 17 Mar 2026) OR (Roytburg et al., 20 Mar 2026) OR (Brandt, 30 Sep 2025) OR (Adegbenro, 13 Jun 2026)", max_results=10) arxiv_search(query="(Lúčny et al., 7 Aug 2025)", max_results=5) The Stability-and-Legibility Account, as it appears across recent research, links two desiderata that are often optimized separately: stable system behavior and legible system behavior. In robotics, legibility is the degree to which an observer can infer an intended goal from a movement snippet rather than merely observe efficient execution; in human-centered AI, legibility is associated with transparency, interpretability, and explainability; in adaptive dialogue systems, stability denotes long-term persona consistency; in trusted monitoring, legibility concerns whether a weaker overseer can understand and evaluate a stronger model’s actions; and in prediction markets, settlement legibility denotes the degree to which an uncertainty can be worded, sourced, and credibly resolved by third parties. The literature suggests that the account is better understood as a recurring design principle than as a single formal theory: systems should remain reliable, predictable, or tractable while exposing intentions, rationales, prompts, or resolution rules in forms that humans or weaker evaluators can actually use (Faria et al., 2022, Hemment et al., 2023, Brandt, 30 Sep 2025, Sreevatsa et al., 9 Feb 2026, Adegbenro, 13 Jun 2026).
1. Conceptual scope and domain-specific meanings
A central feature of the account is that neither “stability” nor “legibility” has a single invariant meaning across domains. In sequential decision tasks, legibility is defined as the degree to which an observer can infer the agent’s intended goal by observing a snippet of the agent’s actions, and it is explicitly distinguished from both predictability and optimality. Optimality concerns maximizing expected cumulative reward; predictability concerns making actions likely given the policy or action history; legibility concerns maximizing inference of the true goal as early and clearly as possible (Faria et al., 2022). In human-centered AI and XAI, by contrast, stability is described as reliability, predictability, and robustness, whereas legibility is described as transparency, interpretability, and explainability (Hemment et al., 2023).
| Domain | Stability emphasis | Legibility emphasis |
|---|---|---|
| Robot motion and sequential policies | tractable computation; robust behavior under uncertainty | intent inference from movement or action snippets |
| Experiential AI | reliability, predictability, robustness | transparency, interpretability, explainability |
| Adaptive chatbots | long-term persona stability | prompt legibility, instruction churn, register flip rate |
| Trusted monitoring | successful oversight without performance/safety tradeoff | rationale behind actions legible to the trusted monitor |
| Prediction markets | platform settlement infrastructure | wording, sourcing, and credible third-party resolution |
| Visual text perturbations | robustness to adversarial attack | whether a literate person can visually decipher altered text |
| Reasoning traces | stable task performance | traces that are checkable, teachable, and useful to weaker models |
The robotics literature gives the most explicit canonical distinction. One paper states: “A legible movement is characterized, not by its efficiency in reaching the goal, but by its distinctiveness” (Faria et al., 2022). Recent motion-generation work restates the trade-off in probabilistic terms: legibility maximizes , whereas efficiency or predictability maximizes (Crétides et al., 17 Mar 2026). This distinction underlies much of the later cross-domain borrowing of the account.
2. Formalizations and measurement strategies
The account is operationalized through several distinct but structurally related measurement schemes. In sequential robotics, legibility is formalized as a posterior over goals conditioned on partial behavior, , and PoL-MDP defines a per-action legibility reward
thereby rewarding actions that are distinctive for the true goal relative to alternatives (Faria et al., 2022). Framework-level benchmarking in robot motion uses a different strategy: a framework-independent trajectory dataset is combined with a reusable human baseline, and frameworks are compared by correlation between framework scores and empirical human legibility judgments. This method enables N-way comparison, fine-grained analysis over progression levels, and decoupling of scoring functions from trajectory optimization (Wallkotter et al., 2022).
Outside robotics, the same account is instantiated through domain-specific observables. In adaptive chatbots, prompt legibility is quantified by prompt churn, register flip rate, and cache hit rate; in reasoning-language-model evaluation, legibility is decomposed into efficiency-based metrics such as trace length, redundancy, and backtracking, and transfer-based metrics such as first-order transfer utility, second-order transfer utility, and regression rate (Brandt, 30 Sep 2025, Roytburg et al., 20 Mar 2026). Transfer utility is especially notable because it defines legibility not as intrinsic textual clarity but as pedagogical usefulness to a weaker model. In prediction markets, settlement legibility is encoded by an ordinal codebook with primary score
where the scored dimensions are template repeatability and settlement determinacy, with closure precision used as a robustness check (Adegbenro, 13 Jun 2026).
These formalizations share a common structure. Each replaces a purely internal or efficiency-centered criterion with an observer-relative one: posterior goal inference, human-baseline alignment, verifier checkability, prompt maintainability, or third-party resolvability. A plausible implication is that the account is best characterized by its measurement stance: behavior is not fully evaluated until its meaning remains recoverable to the relevant observer.
3. Robot motion, multimodality, and sequential decision-making
The most developed empirical literature under the account concerns robot motion. In a pointing experiment with the NICO humanoid robot, 28 adults observed arm movements truncated at 60% or 80% of the full trajectory, with cue conditions varying among gaze alone, pointing alone, and gaze-plus-pointing with congruent or incongruent gaze. Prediction bias was significantly lower at 80% trajectory than at 60% trajectory, confirming that more motion information yields higher legibility; the multimodal superiority and ocular primacy hypotheses were both supported, with gaze-plus-pointing improving target prediction and gaze-only producing the fastest responses (Lúčny et al., 7 Aug 2025). The same study reports that gaze helped anchor predictions along the horizontal axis and that combined cues reduced error, especially lateral bias. This suggests that legibility in humanoid motion is not merely kinematic but multimodal and cue-hierarchical.
A broader benchmarking study compared 10 legibility frameworks from 6 papers using framework-independent trajectories. Overall correlations with human legibility scores were moderate, with Spearman ; Bodden Legibility, especially its velocity metric variant, best matched human judgments in one scenario; and viewing angle and trajectory progression materially affected human evaluation (Wallkotter et al., 2022). One consequence is methodological rather than purely substantive: legibility depends not only on the trajectory generator but on observer geometry and on which fraction of the trajectory is visible.
For stochastic sequential tasks, PoL-MDP extends legibility beyond single motion segments. It is designed to handle uncertainty while remaining computationally tractable, and in simulated experiments it always found a solution with 0% fail rate, whereas L-MDP exhibited 40–60% fail rate in larger settings. In a user study with 150 Prolific participants observing a mobile robot in a maze, PoLMDP-generated policies yielded 85% correct goal identification versus 70% for optimal-policy demonstrations, and the PoLMDP group took, on average, 3 seconds less to decide (Faria et al., 2022). Here the account becomes explicit: policies are valuable not solely because they reach the goal, but because they systematically support reliable observer inference.
Recent diffusion-policy work introduces an adaptive variant of the same idea. Style-Conditioned Diffusion Policy freezes a base diffusion policy and trains a lightweight scene encoder and conditioning predictor so that, at inference time, an ambiguity detection module activates legibility conditioning only in ambiguous scenes and reverts to efficient paths otherwise. In navigation, SCDP achieved in ambiguous cases, close to the Legibility Diffuser baseline at $0.62$, but in non-ambiguous cases achieved versus $0.32$ for the always-legible baseline, precisely by not enforcing exaggerated trajectories when they were unnecessary (Crétides et al., 17 Mar 2026). This turns the account into a context-sensitive control rule rather than a global preference for expressive motion.
4. LLMs, adaptive dialogue, and trusted oversight
In language-model research, the account shifts from physical motion to the legibility of outputs, prompts, and rationales. One line of work studies chain-of-thought solutions for grade-school math and shows that optimizing only for answer correctness can make reasoning less legible. Prover-Verifier Games address this by alternating between small verifiers, “helpful” provers that produce correct solutions the verifier accepts, and “sneaky” provers that produce incorrect solutions that fool the verifier. Over training, the helpful prover’s accuracy and the verifier’s robustness to adversarial attacks increase, and human accuracy under time constraints increases when checking helpful-prover solutions while decreasing when checking sneaky-prover solutions (Kirchner et al., 2024). The relevant legibility criterion is explicitly checkability by weaker agents.
A related but broader evaluation of 90k reasoning traces from 12 Reasoning LLMs introduces transfer utility as the usefulness of a strong model’s reasoning trace for guiding a weaker, non-reasoning model to the correct answer. The study finds that the reasoning traces of the highest-performing models rank among the lowest for legibility, uncovers tensions between efficiency-based measurements such as trace length and transfer utility, and concludes that reward models used to train RLMs do not intrinsically reward legibility (Roytburg et al., 20 Mar 2026). This directly challenges the common assumption that better reasoning models naturally emit better explanations.
Adaptive chatbots instantiate the account at the prompt-policy level. Using an 8-dimensional style vector and a closed-loop base+delta prompting architecture, one study evaluates explicit adaptation policies—Uncapped, Cap, EMA, Dead-Band, and Hybrids—and maps a clear Pareto frontier between synchrony and long-term persona stability. A Hybrid (EMA+Cap) raises stability from 0.542 to 0.878 while reducing synchrony by only 17%, and the same bounded policies reduce prompt volatility: register flip rate drops from 0.254 to 0.092 under the Hybrid policy (Brandt, 30 Sep 2025). Here legibility means that dynamic instructions are easier to reason about, debug, and maintain.
Trusted monitoring extends the account into AI control. In APPS coding evaluations where an adversarial model attempts to write backdoored code without detection, legibility protocols require the untrusted model to document code with comments rather than remove comments. The study reports that commenting protocols improve safety without sacrificing task performance relative to comment-removal baselines, that commenting disproportionately benefits honest code because honest code typically has a natural explanation that resolves monitor suspicion, and that gains from commenting increase with monitor strength (Sreevatsa et al., 9 Feb 2026). The paper’s formalization is operational: legibility protocols are protocols that encourage the untrusted model to make the rationale behind its actions legible to the trusted monitor.
5. Human-centered AI, adversarial text, and institutional legibility
The account has also expanded beyond intent communication and oversight to human-centered creative systems, adversarial NLP, and market design. In Experiential AI, legibility is not restricted to model internals or saliency maps. The field is described as making AI tangible and explicit through embodied, interactive experiences, with the New Real Observatory Platform and the SLIDER tool enabling artists to define user-specified dimensions, visualize latent spaces, and probe associations within image and word models. The reported outcome is increased artistic agency, greater understanding of how data and inputs affect outputs, and improved predictability and reliability through interactive experimentation (Hemment et al., 2023). This redefines legibility as contextual, embodied, and public-facing.
In adversarial NLP, legibility becomes a property of perturbed text rather than explanations or trajectories. “LEGIT” is a human-annotated dataset of visually perturbed words, constructed through pairwise judgments of which perturbation is more legible. The resulting models achieve up to 0.91 F1 in predicting whether an input is legible and 0.86 accuracy in predicting which of two perturbations is more legible, and legible perturbations drawn from LEGIT are reported to be more effective at lowering NLP model performance than best-known attack strategies (Seth et al., 2023). The implication is that robustness claims are incomplete if they are benchmarked only against heuristically chosen perturbations rather than the broader space of human-legible ones.
Prediction-market research extends the vocabulary of legibility to institutional formation. Using an audited dataset of 6,047 Africa-topic and Latin America-topic contracts on Polymarket and Kalshi, one study defines settlement legibility as the degree to which an uncertainty can be worded, sourced, and credibly resolved by third parties. The coding protocol reaches ordinal reliabilities of 0.92 and 0.96 on the primary dimensions and blind human benchmarks of 0.97 and 0.92. Legibility orders the listed inventory steeply, with sports and elections near the top of the scale and conflict at the bottom; in a formation logit, the legibility coefficient is 0.433 with 0, while among listed contracts the association between legibility and trading value is negative, with coefficient 1 and 2 (Adegbenro, 13 Jun 2026). In this setting, the account no longer concerns explanation of model behavior but the platform-side preconditions under which uncertainty becomes tradable at all.
6. Trade-offs, Pareto frontiers, and unresolved issues
A recurring empirical pattern is that stability and legibility are jointly desirable but rarely costless. In robot motion, longer or more distinctive trajectories can be more legible even when they are less efficient, and RL-based frameworks sometimes negatively correlate with human data because they penalize trajectory length even though longer trajectories can provide more time and information for disambiguation (Wallkotter et al., 2022). In adaptive dialogue, bounded style-adaptation policies are Pareto-efficient because they substantially improve stability at only modest synchrony cost (Brandt, 30 Sep 2025). In reasoning-language-model evaluation, no single model dominates across correctness, efficiency-based legibility, and transfer utility, and the work explicitly identifies a legibility Pareto frontier (Roytburg et al., 20 Mar 2026). In SCDP, the proposed resolution is conditional activation: prioritize expressive motion only for ambiguous goals and revert to efficient paths otherwise (Crétides et al., 17 Mar 2026).
The literature also rejects several simplified interpretations. First, high task performance does not imply high legibility: optimizing chain-of-thought purely for correctness can reduce legibility, and the highest-performing reasoning models may rank low on transfer utility (Kirchner et al., 2024, Roytburg et al., 20 Mar 2026). Second, legibility is not equivalent to minimalism: verbose traces may be more pedagogically useful, while short traces may omit the bridging steps that weaker models or humans need (Roytburg et al., 20 Mar 2026). Third, legibility is not merely a property of the artifact itself; it is observer-dependent. Viewing angle affects motion legibility, prompt legibility matters for maintainability, and settlement legibility depends on the availability of source hierarchies and credible adjudication (Wallkotter et al., 2022, Brandt, 30 Sep 2025, Adegbenro, 13 Jun 2026). Fourth, public inventories should not be read naively: prediction-market inventories measure what platforms can settle as much as what traders believe (Adegbenro, 13 Jun 2026).
The cumulative record therefore supports a general interpretation of the Stability-and-Legibility Account as a multi-objective research program. Stability may refer to robustness, persona coherence, tractability, or oversight reliability; legibility may refer to intent expressivity, transparency, checkability, transfer utility, prompt maintainability, human readability, or settlement determinacy. What unifies these uses is the claim that systems should not be evaluated only by internal reward, final-answer correctness, or economic uptake. They must also be evaluated by whether their behavior remains recoverable, monitorable, teachable, or resolvable for the agents who interact with them. The literature suggests that future progress will depend less on treating legibility as an automatic byproduct of performance and more on making it an explicit design target, metric family, and deployment constraint (Hemment et al., 2023, Sreevatsa et al., 9 Feb 2026, Roytburg et al., 20 Mar 2026).