Human Reasons-Based Supervision Framework

• Human Reasons-Based Supervision Framework is an architecture that leverages explicit human rationales to guide AI decisions in safety-critical systems.
• It operationalizes values such as safety, comfort, and compliance through formal evaluations and recursive critique, ensuring responsible automated actions.
• The framework combines direct review methods and stakeholder-conditioned scoring to supervise behavior in domains like automated driving and scalable oversight.

Across the cited literature, a Human Reasons-Based Supervision Framework can be understood as a supervision architecture in which AI systems are evaluated, redirected, or authorized through human reasons rather than through end-outcome labels alone. In the direct automated-vehicle formulation, human reasons are the morally and practically relevant considerations that human stakeholders have regarding how the system should behave, including safety, comfort, efficiency, and compliance with traffic regulations; in adjacent scalable-oversight work, reasons appear as explicit rationales, critiques, values, intentions, expectations, or structured negative evidence that humans can inspect when direct solution or direct verification is too hard (Suryana et al., 31 Jul 2025, Wen et al., 7 Feb 2025, Bell et al., 26 Jan 2026).

1. Conceptual basis and scope

The framework is motivated by a common diagnosis: many important AI decisions cannot be supervised adequately by asking humans either to produce gold-standard answers or to rate final outputs directly. In automated driving, the literature argues that ethically important problems are often ordinary, recurring, context-sensitive situations rather than rare trolley-style dilemmas, and that system behavior should align with human reasons understood as values, intentions, and expectations that justify actions. This position is tied to the tracking condition of Meaningful Human Control, under which automated behavior should be responsive to the reasons of relevant humans rather than merely to fixed rules or opaque optimization criteria (Suryana et al., 18 Jul 2025).

A second strand of work argues that preferences alone are too weak and too ambiguous a supervision signal for alignment. “Beyond Preferences” states that principles are “underdetermined by preference annotations” and proposes Grounded Constitutional AI (GCAI), which uses human-written reasons for pairwise preferences to infer contextual principles and direct value statements to infer general principles. This establishes a central distinction for reasons-based supervision: some supervisory content comes from situated judgments about particular cases, while other content comes from broader statements about what AI should do independent of any one interaction (Bell et al., 26 Jan 2026).

A third strand provides the governance shell. The risk-based oversight framework links model influence and decision consequence to human oversight modes—Human-in-Command, Human-in-the-Loop, and Human-on-the-Loop—so that human review intensity is proportional to deployment risk rather than uniformly applied. A plausible implication is that a Human Reasons-Based Supervision Framework is not only a learning architecture but also an institutional arrangement specifying when reasons must be elicited, who may override the system, and how much autonomy is acceptable in a given domain (Kandikatla et al., 10 Oct 2025).

2. Formal representations of reasons as supervision objects

One formal route models reasons as explicit textual artifacts. “Scalable Oversight for Superhuman AI via Recursive Self-Critiquing” defines a response as

$R(Q) \rightarrow (T^0, A^0),$

where $Q$ is the question, $T^0$ is the full solution process or rationale, and $A^0$ is the final answer. A first-order critique is

$C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$

a second-order critique is

$C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$

and the general recursive form is

$$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$

In that formulation, the primary supervised object is not just $A$ but a tuple $(T, A)$ at each level, and humans judge not only end outcomes but the quality of these reason objects. This is an explicit process-based supervision scheme in which critiques and higher-order critiques are reasons about lower-level outputs and reasons about reasons (Wen et al., 7 Feb 2025).

A second formal route models reasons as stakeholder-conditioned evaluation functions over candidate actions. “Assessing the Alignment of Automated Vehicle Decisions with Human Reasons” defines a candidate trajectory set $\mathcal{T}=\{T_1,\ldots,T_k\}$, an agent set $Q$0, and for each agent $Q$1 a set of reasons $Q$2. Each reason is represented by a per-time-step evaluator

$Q$3

with trajectory-level score

$Q$4

These are aggregated into agent scores

$Q$5

and then into the overall alignment score

$Q$6

This makes reasons measurable supervision targets over temporally extended behavior rather than informal after-the-fact explanations (Suryana et al., 31 Jul 2025).

Taken together, these formalisms suggest two recurring supervision objects. One is a textual reason-bearing object $Q$7; the other is a structured action whose alignment with stakeholder reasons is scored by explicit functions. A plausible synthesis is that a Human Reasons-Based Supervision Framework treats supervision as operating on intermediate justificatory structure, whether natural-language rationale, critique chain, or stakeholder-reason score, rather than only on terminal correctness.

3. Recursive critique and interactive decomposition

The strongest scalable-oversight case for explicit reasons comes from recursive critique. The central hypotheses are that critique of critique can be easier than critique itself and that this ease relation may persist recursively. The paper operationalizes tractability through higher answer accuracy, lower or stable completion time, and higher self-reported confidence. In Human-Human experiments across CET-6 reading comprehension, GAOKAO Chinese, GAOKAO Math, KAOGONG reasoning, and Figure Reasoning, mean accuracy rises from response to critique to $Q$8; for example, GAOKAO Math improves from $Q$9, and CET-6 improves from $T^0$0. On CET-6 and GAOKAO Math, $T^0$1 improves further to $T^0$2 and $T^0$3, respectively. In Human-AI experiments on hard GAOKAO Math and TEM4 reading, humans are weaker than Qwen2.5 models at the response stage but often outperform prior AI stages when judging critique chains, which is the paper’s main evidence that humans can supervise superhuman systems better by evaluating model-generated reasons than by generating answers directly (Wen et al., 7 Feb 2025).

A related but distinct approach is Scalable Interactive Oversight. Here the hard problem is not only verification but specification: non-experts often cannot articulate precise intent up front or validate a complex final artifact. The framework therefore decomposes a user query $T^0$4 into a tree-structured plan $T^0$5, traverses the tree depth-first, elicits low-burden local feedback $T^0$6 at each node, updates a cumulative context $T^0$7, and finally generates a Product Requirement Document from the final tree and context. Feedback is closed-form—primarily selection-based or ranking-based—with explicit DontCare and DontKnow options, rather than open-ended rationale collection. The paper treats these signals as structured local decisions rather than full human-authored reasons, but the cumulative preference state functions as a trace of why the final artifact takes its eventual form. In web-development experiments, the method improves alignment from $T^0$8 to $T^0$9 over vanilla interaction in the Gemini case, which the paper describes as a $A^0$0 relative improvement, and it further shows reinforcement learning over interaction quality using User Reward, Progressive Reward, and Outcome Reward (Zhou et al., 4 Feb 2026).

These two lines occupy different points on the same design spectrum. Recursive critique supervises reason objects directly; interactive decomposition elicits low-burden local judgments that function as implicit reasons proxies. Both are grounded in the weak-to-strong setting: humans need not solve the full task if the system presents the task as critiques of critiques, or as a sequence of bounded decisions over an interpretable decomposition.

4. Weak signals, triage, and operational substrates

Some work in this area is not reasons-based in the strong sense, but it supplies the routing and infrastructure that reasons-based supervision requires. “Arguing Machines” pairs a primary black-box AI system with an independently trained secondary system, computes disagreement, and invokes a human when disagreement exceeds a threshold. In ImageNet, disagreement between ResNet-50 and VGG-16 flags $A^0$1 of images; sending those cases to perfect human review reduces top-5 error from $A^0$2 to $A^0$3. In driving, disagreement between Tesla Autopilot and a separately trained end-to-end steering model predicts $A^0$4 of system disengagements judged challenging and needing human supervision. This framework does not generate reasons, but it supplies a principled escalation trigger for when reasons are needed (Fridman et al., 2017).

“Scalable Oversight via Partitioned Human Supervision” uses complementary labels as structured negative evidence. Instead of asking a human for the correct class, it asks whether class $A^0$5 is correct; a “no” yields a complementary label $A^0$6, meaning class $A^0$7 is known to be wrong. Under the paper’s uniform complementary-label assumption, the unbiased estimator of top-1 accuracy is

$A^0$8

and matching the variance of ordinary labels requires

$A^0$9

The paper is not framed as reasons-based supervision, but it explicitly motivates complementary labels as narrow-expert disqualifications such as “this is not related to cardiology.” A plausible implication is that minimal negative reasons can already support principled evaluation when full verification is unavailable (Yin et al., 26 Oct 2025).

Operational deployment raises a separate systems problem: how to insert human review without stalling expensive computation. CIF, the Collaborative Innovation Framework, addresses asynchronous human-AI collaboration across HPC clusters, local machines, and cloud platforms. Workflows are specified declaratively in TOML, include explicit HITL checkpoints with hitl.enabled = true, and follow the operational cycle

$C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$0

CIF-CLI, TaskScheduler, TaskExecutor, WorkflowExecution, Watcher, and JobRegistry provide a substrate for non-blocking supervision, runtime restructuring, and persistence of decision artifacts such as hitl_decision.json. The paper explicitly states that CIF does not yet implement substantive reasons capture, but it offers the workflow checkpoints and audit hooks through which richer reason payloads could be attached (Mendoza et al., 5 May 2026).

A medical analogue appears in HiLWS, a cascaded human-in-the-loop weak supervision framework for neurological video curation. HiLWS first aggregates five expert severity annotations into probabilistic labels, trains a LightGBM classifier and a Random Forest classifier, and then combines the five experts and two models in a second weak-supervision stage. Uncertainty is quantified by entropy,

$C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$1

and high-entropy cases are returned to experts for review. On the gold-standard balanced test set, HiLWS Full attains MAE $C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$2, F1 $C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$3, accuracy $C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$4, and $C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$5, outperforming single-rater labels and majority vote. The framework is label-centric rather than explicitly reasons-based, but it shows how ambiguity-triggered human escalation and multi-source aggregation can be embedded into a curation pipeline under domain shift (Irani et al., 9 Sep 2025).

5. Automated driving as a direct instantiation

Automated driving is the domain in which Human Reasons-Based Supervision appears most directly and explicitly. One contribution is taxonomic. “Principles and Reasons Behind Automated Vehicle Decisions in Ethically Ambiguous Everyday Scenarios” reports interviews with $C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$6 AV experts and identifies $C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$7 categories of reasons organized across normative, strategic, tactical, and operational levels. These include regulatory compliance, social legitimacy, environmental responsibility, fairness and equality, cultural and norm adaptation, efficiency and trip optimisation, user and public acceptance, comfort and user experience, risk minimisation and safety assurance, human interaction management, continuous vehicle control, control transition, and driver/system vigilance and readiness. The paper’s strongest prioritization claim is that safety is primary and “non-negotiable,” while legality is a default baseline that can be conditionally defeasible when strict compliance undermines higher-priority or legitimate practical values (Suryana et al., 18 Jul 2025).

The direct framework paper then operationalizes those ideas into a supervisory layer over planning and control. “A Framework for Ethical Decision-Making in Automated Vehicles through Human Reasons-based Supervision” models three stakeholders in a cyclist-overtaking scenario: policymakers, a vulnerable road user, and the driver. For each stakeholder $C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$8, it defines a reason function

$C^1(Q, R_1, R_2) \rightarrow (T^1, A^1),$9

uses a shared threshold $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$0, and triggers replanning when

$C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$1

The system combines an A* global motion planner, an MPC controller, and the Human Reasons-Based Supervision Framework, which monitors reason scores in real time and can lower the planner’s traffic-rule penalty during replanning. In simulation, the baseline controller does not overtake and reaches the goal in $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$2 seconds, with the driver score falling to zero by the end. With the supervisory layer active, the system detects driver-score misalignment at $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$3 seconds, replans into a brief opposite-lane overtake, and reaches the goal in $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$4 seconds; after returning to lane, all reason values recover to $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$5 (Suryana et al., 31 Jul 2025).

CARE-Drive extends this line from control to evaluation of vision-LLMs. It compares baseline and reason-augmented decisions under controlled contextual variation and formalizes the decision model as

$C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$6

In the cyclist-overtaking case, the baseline condition $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$7 yields an overtaking rate of $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$8 over $C^2(Q, R_1, R_2, C_1, C_2) \rightarrow (T^2, A^2),$9 runs across all tested model and thought-strategy combinations, whereas explicit reasons induce overtaking and materially change behavior. Stage-2 perturbation fits the logistic model

$$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$0

showing strong positive responsiveness to safety margin $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$1 and rear-vehicle pressure $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$2, weak or counterintuitive responsiveness to urgency $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$3 and following time $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$4, and a strong suppression effect from constrained explanation length $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$5. CARE-Drive explicitly frames these results as behavioral evidence of reason-responsiveness rather than proof of internal causal use of reasons (Suryana et al., 17 Feb 2026).

6. Principle induction, constitutional grounding, and governance

Reasons-based supervision can also operate at the level of principles rather than instance-level action review. GCAI builds constitutions from two distinct supervisory channels: reasons for pairwise preferences and direct statements of values. From HelpSteer 2, human-written justifications are distilled into contextual principles; from PRISM, direct value statements are decomposed into general principles. Candidate principles are embedded with text-embedding-3-small, clustered, summarized, scored, and assembled into a final constitution with $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$6, split evenly between contextual and general principles. In constitution-level human evaluation, GCAI defeats ICAI by a $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$7 win-rate on moral grounding and $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$8 on both personal preference and governing preference. The resulting constitution includes principles about factuality, safety, impartiality, and respectful interaction that are underrepresented in preference-only induction. This moves reasons-based supervision from local judgment toward principle elicitation and constitutional governance (Bell et al., 26 Jan 2026).

A governance framework is still needed to decide how such reasons should constrain deployment. The risk-based oversight literature provides precisely that outer layer. By evaluating model influence and decision consequence, then mapping the resulting risk tier to HOTL, HITL, or HIC, it offers a repeatable procedure for deciding when reasons-based review should be periodic monitoring, case-level validation, or ultimate human command. The same framework also recommends Fundamental Rights Impact Assessment for high-risk systems and insists that overseers be given sufficient authority and independence to question, challenge, and critically assess AI outputs rather than defer to them as infallible experts. A plausible implication is that reasons-based supervision is most complete when paired with governance rules specifying whose reasons count, at what risk level, and under what override authority (Kandikatla et al., 10 Oct 2025).

7. Limitations, error floors, and open questions

The literature is explicit that reasons-based supervision is not solved. Recursive critique is promising, but current models “struggle to surpass the accuracy levels achieved at the Response stages” in most AI-AI settings, and the same work warns about self-consistent but wrong reasoning, collusion-like self-critique, verbosity bias, adversarial persuasion, unresolved faithfulness, and strategic deception. It also notes that recursive critique is not guaranteed to improve monotonically at every depth and that optimal recursion depth remains open (Wen et al., 7 Feb 2025).

A broader theoretical limitation is that human supervision itself may be an information bottleneck. “Human Supervision as an Information Bottleneck” argues that whenever the human supervision channel is not sufficient for a latent evaluation target $$C^n(Q, R_1, R_2,..., C^{n-1}_1, C^{n-1}_2) \rightarrow (T^n, A^n).$$9, it induces a strictly positive excess-risk floor for any learner dominated by that channel. In that account, persistent alignment errors arise from annotation noise, preference distortion, and semantic compression, and scaling alone cannot eliminate them. The paper further argues that sufficiently informative auxiliary non-human signals such as retrieval, program execution, and tools can collapse the floor by restoring information about the latent target. This places an upper bound on what reasons-based supervision can achieve if the reasons channel itself is too compressed or too lossy (Dominguez, 26 Feb 2026).

The boundaries of the concept are also important. Some adjacent approaches enrich human supervision without becoming fully reasons-based. R-Few stabilizes self-evolving LLMs by injecting $A$0–$A$1 of human anchor examples into a challenger-solver loop; on Qwen3-8B-Base it improves by $A$2 points over R-Zero on math tasks and matches General-Reasoner with roughly $A$3 times less human data, but its human signal is exemplar-based grounding rather than explicit human reasons. Weak human preference supervision for reinforcement learning similarly replaces fixed categorical choices with a scalar preference $A$4 and reduces human input cost by up to $A$5, yet it still collects graded preference intensity rather than rationales, critiques, or justifications (Yu et al., 2 Dec 2025, Cao et al., 2020).

The open problems therefore have a common structure. The field still lacks a generally accepted method for eliciting stakeholder reasons, validating that formal proxies faithfully represent those reasons, balancing conflicting reasons without arbitrary weight assignment, distinguishing genuine reasons-responsiveness from post hoc rationalization, and converting human-authored reasons into scalable training signals without losing the information they were supposed to preserve. The existing literature nevertheless converges on a stable core idea: alignment becomes more inspectable, and often more scalable, when supervision targets not only what systems output, but also the reasons, critiques, values, and disqualifying considerations through which humans judge those outputs.