{"id":"3b26ce19-6e2d-43b1-a931-9b21aaa74944","arxiv_id":"2505.15503","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CBTL learns parameterized personalization constraints inside the safe solution space of robot planning CSPs, using entropy-based active queries to adapt quickly.","lead":"This paper introduces Coloring Between the Lines (CBTL), a method that lets a robot personalize its behavior by learning extra constraints that pick among the many safe solutions of a planning constraint satisfaction problem. The approach matters because it could let assistive robots adapt to individual users without giving up hard safety limits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Safety claim rests on an unverified assumption: base CSP constraints must fully encode safety, and no experiment or proof checks this, so 'safe' is a promise rather than a demonstrated result.","rationale":"Read in good faith, the paper contributes a clear framework: plan with a CSP whose constraints guarantee safety, then learn personalized constraints that select within that safe set. The empirical results are encouraging for personalization efficiency and generalization. However, the claim 'safe' in the abstract and introduction is not supported by the evidence. The paper's own limitation statement admits the safety guarantee depends on the CSP generators, but no experiment measures whether unsafe behavior can occur. I considered alternative concerns: LLM next-token-probability calibration affects learning efficiency but not safety; the single real-robot demonstration is a generality concern but not the central claim; nonstationary handling is appendix-only but not required for the main claim. The safety delegation is the most load-bearing because the entire method's value proposition—'color between the lines'—depends on the lines being correct. The reader's weakest_assumption identifies exactly this, and I agree. The correct response is to keep the paper under conditional acceptance, requiring either a safety audit of the generators or explicit reframing of the claim from 'safe' to 'safe conditional on correct base constraints.'","tokens_in":16115,"tokens_out":5735,"duration_ms":55287,"concrete_test":"Audit the base CSP generators for the three released simulation environments: sample (or exhaustively enumerate for small domains) a large number of solutions from each base CSP with personalized constraints disabled, and check every sampled solution against the environment's ground-truth safety predicates (collision with obstacles, self-collision, joint limits, static stability). If any sampled solution violates a safety predicate, the null space is unsafe and the central claim fails. Additionally, add a safety-violation counter to the existing simulation evaluation and report it for CBTL and all baselines; a nonzero count would directly contradict the claim of safe personalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CBTL's central claim of 'safe robot personalization' is inherited from the premise that the hand-written, non-personalized CSP constraints fully and correctly encode safety and general competency. Since personalized constraints only further restrict the solution set, the method cannot create unsafe behavior if that premise holds; if it fails, however, the null space itself contains unsafe solutions that CBTL will actively seek out during entropy-maximizing exploration and then enforce during exploitation. The paper explicitly concedes this in Section 6: 'the extent to which safety and competency are guaranteed by CBTL depends on the implementation of the CSP generators.' No formal verification, no safety metric, and no adversarial test of the generators is provided. The simulation experiments measure user satisfaction, not safety violations; the web study measures stated preference; the real-robot demonstration is a single participant without quantitative safety evaluation. Thus the strongest advertised property—safety—is exactly the one that is not established. This is not a circularity but an unexamined delegation: the method's contribution is sound only if the delegated component is correct, and that correctness is assumed rather than demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Coloring Between the Lines (CBTL), a framework for robot personalization that treats the solution set of a planning constraint satisfaction problem (CSP) as a 'null space' and adds learnable, parameterized constraints to select solutions tailored to an individual user. The base CSP constraints are assumed to encode safety and general competency, while personalized constraints are learned online from user feedback, either through supervised classification or through LLM-based natural-language summaries. Exploration is driven by an entropy-maximizing objective over the personalized constraints, constrained by the base CSP. The paper evaluates CBTL in three simulation environments against four baselines, with a 60-participant web study and a real-robot assisted-feeding demonstration, and reports that CBTL personalizes more effectively and with fewer interactions.","tokens_in":16380,"tokens_out":6702,"duration_ms":63300,"significance":"If the claims hold, CBTL offers a useful unification of constraint-based planning, online preference learning, and active exploration for continual robot personalization. The paper is empirically broad: three simulation environments, a 60-participant web study, a real-robot demonstration, and released code for reproducing the results. The compositional structure of personalized constraint generators and the entropy-based active-learning criterion are the most novel elements, and the demonstration that a learned occlusion preference transfers from feeding to drink placement is a meaningful generalization result. The main weakness is that the advertised safety property is not directly established by any experiment, and the empirical comparisons would be stronger with explicit statistical testing.","major_comments":[{"comment":"The paper advertises 'safe robot personalization' as a central contribution, but no experiment directly measures safety. No metric counts constraint violations, collisions, near-misses, or failures caused by incomplete or incorrect non-personalized CSP generators; the real-robot demonstration is n=1 and reports no quantitative safety evaluation. Section 6 explicitly concedes that 'the extent to which safety and competency are guaranteed by CBTL depends on the implementation of the CSP generators,' which means that the entropy-maximizing exploration will actively seek out any unsafe solution that the base constraints fail to exclude. This is acceptable as a conditional contribution, but the claim as stated overreaches. Please either soften the safety language throughout (e.g., 'safety-preserving given correct base CSP generators') or add a targeted experiment that varies the completeness/correctness of the generators and measures safety violations.","section":"Abstract, §1, §5, §6"},{"comment":"The headline empirical claim that CBTL 'consistently personalizes better and faster' is supported only by mean learning curves with standard errors; no pairwise significance tests, confidence intervals, or effect sizes are reported for the three environments. The Cooking plot raises a further question: the caption notes that Epsilon-Greedy and Exploit Only obtain equal performance, and Appendix D attributes this to both being trapped in the same local minima under controlled seeding, which suggests the comparison may be sensitive to the evaluation protocol. Please report per-environment statistical comparisons over the 10 seeds and clarify whether the CBTL advantage in Cooking is significant and robust.","section":"§5.1, Figure 4"},{"comment":"The description of uncertainty quantification for LLM-based constraints is internally inconsistent. The text states that LLM next-token probabilities 'do not provide a useful measure of uncertainty' because the model is overconfident, but then defines the expected constraint probability as \\sum_{i=0}^{10}(i/10)p_i, where p_i is the next-token probability. If these p_i are the same overconfident probabilities, the resulting entropy will be artificially low and the active-learning objective in Section 4 may not reflect true uncertainty. Please clarify whether p_i are raw token probabilities or a recalibrated or softened distribution, and specify whether the scalar response (e.g., '8') is converted directly to p=0.8 (in which case the formula with p_i is unnecessary) or used as a sample from a calibrated distribution.","section":"Appendix B, Section 4"}],"minor_comments":[{"comment":"The y-axis is not labeled, and the caption says units are deliberately omitted; however, the plotted quantity should at least be named (e.g., 'average user satisfaction' or 'held-out task score') so that readers can interpret the scale.","section":"Figure 4"},{"comment":"The p<0.005 results are reported per meal without correction for multiple comparisons across five meals; please state whether these are pre-specified hypotheses or exploratory analyses.","section":"§5.2"},{"comment":"The real-robot demonstration is described as 'validates' and 'demonstrates real-world applicability'; with n=1 and no quantitative measurements of task success or safety, the wording is stronger than the evidence. Consider describing it as an illustrative pilot.","section":"§5.3"},{"comment":"The LEARNINGPROMPT for Books asks the LLM to return multiple possible summaries of the user's preferences, and the entire response is used as θ; please explain how the CONSTRAINTPROMPT interprets a parameter containing several alternative summaries when deciding whether a book is enjoyable.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the overclaimed safety property: the method's key selling point rests on an unvalidated delegation to hand-written CSP generators, and the paper's own limitations section acknowledges this. If the authors reframe the contribution as 'safety-preserving under correct base constraints' and add a targeted safety-failure experiment, the paper would be much stronger. The LLM uncertainty-quantification inconsistency in Appendix B should also be resolved before publication, as it bears directly on the active-learning mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Name],\n\nRead the CBTL paper. My take: this is a solid, well-executed paper that delivers on its main claim—personalization in the null space of CSP constraints works in simulation, in a web study, and in a real feeding demo. The core idea—treat the solution set of a planning CSP as the safe operating envelope, and learn parameterized constraints inside it—is genuinely useful and unifying. It connects constraint-based TAMP with active learning and continual adaptation in a way I haven't seen done cleanly before. The experiments are broad: three sim environments with 10 seeds and sensible baselines, a 60-person web study, and an n=1 real robot demo showing cross-task generalization. The paper is also refreshingly honest in its limitations section.\n\nThe soft spots are real but not fatal. The safety claim is inherited from the assumption that the hand-written non-personalized constraints are correct. The paper acknowledges this, but the abstract and framing still say 'safe personalization' without the qualifier. That's worth pushing back on: either verify the generators for a given task, or explicitly scope the safety claim to 'safe only to the extent that the base constraints are correct.' The real robot evidence is a single participant—fine for a demo, but don't oversell it. Also, the code release is mentioned but no link appears in the text; that's a minor reproducibility issue.\n\nOverall, this deserves serious review. It's a strong framework paper with empirical support. The safety caveat should be addressed with more explicit language and ideally a stress test, but it doesn't invalidate the contribution.\n\nFor peer review: yes, send it out.","headline":"Solid framework paper for robot personalization via CSP null-space constraints; safety claim needs scoping but the mechanism and evaluation hold up.","tokens_in":16844,"tokens_out":1768,"would_cite":true,"duration_ms":16156,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes that a robot can personalize to an individual user by learning parameterized constraints that select among the already-safe solutions of its planning constraints, rather than by changing those constraints.","keywords":["personalization","constraint satisfaction problems","null space","active learning","robot planning","assistive feeding","online learning","compositional constraints"],"falsifier":"Give CBTL a CSP generator that omits a collision constraint between the robot arm and the user's torso during feeding, then run it in a physical or simulated feeding task; if the robot ever executes a torso-colliding configuration while satisfying the stated constraints and the user-adaptive objective, the claim that personalization stays in a safe null space is refuted.","tokens_in":15947,"feed_emoji":"🤖","tokens_out":6492,"duration_ms":55698,"temperature":0.7,"pith_summary":"This paper argues that the tension between safety and flexibility in robot personalization can be dissolved by keeping the robot's original planning constraints fixed and personalizing only within the set of solutions those constraints already allow. The proposed method, Coloring Between the Lines (CBTL), adds learnable parameterized constraints to a constraint-satisfaction-planning pipeline, learns their parameters from online interaction, and actively chooses configurations that reduce uncertainty about the user. If the base constraints correctly encode safety and general competence, then every personalized behavior remains inside the safe solution set. The paper reports that CBTL personalizes better and faster than four baselines across three simulation environments, is preferred by users in a web study, and generalizes on a real assistive-feeding robot.","feed_headline":"Robots personalize safely inside the null space of planning constraints","feed_subtitle":"CBTL learns user preferences from a few interactions without leaving the safe solution set of the base planner.","key_machinery":"The central object is the CSP null space: the set of assignments satisfying the robot's non-personalized constraints. CBTL's machinery is a CSP generator augmented with personalization generators, each of which emits parameterized constraints; parameters are learned online from interaction, and an entropy-based mathematical program selects solutions with maximum uncertainty about the user. A sampling-based solver finds solutions, with a fallback that re-plans without personalized constraints if they are unsatisfiable. This setup lets the robot personalize by reducing the null space rather than moving outside it.","core_discovery":"The central claim is that the multiplicity of solutions of a planning constraint satisfaction problem is not a nuisance but a resource: it is a null space within which a robot can safely adapt to an individual user. CBTL operationalizes this by generating parameterized personalized constraints alongside the original constraints, where each constraint is produced by a generator with an initiation condition and a learnable parameter, and by updating those parameters from feedback history through supervised classification for numeric parameters and through natural-language summaries for large-language-model-implemented constraints. Active learning then solves an entropy-maximization program over the CSP's null space to pick the most informative next behavior. The paper demonstrates that this yields continual, flexible, active, and safe personalization in simulation, in a 60-participant web study, and on a real feeding robot.","pith_inferences":["The safety guarantee reduces to the correctness of the hand-written CSP generator, so automated verification or synthesis of these generators would be a natural, high-impact next step that the paper leaves open.","The same null-space mechanism could serve as a general interface between human functional diversity and robot behavior, covering not just preferences but accessibility parameters such as reach envelopes and occlusion tolerance.","Because the active-learning program only needs probabilistic personalized constraints, other probabilistic preference models, such as Gaussian processes, could replace the supervised and LLM constraints used here and potentially yield sharper uncertainty estimates.","The framework suggests a direct extension to cost-based personalization, where preferences like speed could be optimized as soft objectives inside the same null space instead of hard constraints."],"forward_implications":["A single method can personalize across qualitatively different preferences including dip choice, ready signal, verbosity, occlusion sensitivity, book taste, and functional reach, because each is expressed as a parameterized constraint in the same CSP.","Personalization is sample-efficient: by actively maximizing entropy, CBTL needs fewer user interactions than unguided exploration or pure exploitation, as measured by user satisfaction in all three simulated environments.","Learned personalized constraints transfer across scenarios: the occlusion-sensitivity constraint learned during feeding is reused to reposition plate and drink when a new object appears on the table.","When personalized constraints cannot all be satisfied, CBTL can fall back to a non-personalized solution and still complete the task, maintaining task success at the cost of unmet preferences.","By keeping the original CSP constraints fixed, CBTL's safety and competency guarantees coincide with the correctness of those constraints."],"supporting_citations":[{"why":"Supplies the compositional constraint-based planning viewpoint that motivates the CSP-null-space framing.","marker":"[18]"},{"why":"Provides the stream-based planning abstraction from which the paper's CSP generators are derived and justified.","marker":"[59]"},{"why":"Introduces the entropy-based active constraint learning that CBTL's exploration objective is built on.","marker":"[30]"},{"why":"Offers the occlusion model used to implement the personalized occlusion-free constraint in feeding experiments.","marker":"[27]"},{"why":"Contributes user functionality embeddings that motivate and implement the reachability personalization used in the Books environment and feeding.","marker":"[1]"},{"why":"Presents long-horizon bite preferences that seed the dip-preference personalization examples and user-study design.","marker":"[26]"}],"fun_headline_variants":["Safe personalization via planning constraint null spaces","Robots personalize in the null space of safety constraints","CBTL: Few-interaction safe robot personalization","Learning user preferences without leaving safe solution space","Null-space constraints let robots adapt safely to users"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The robot's safety and everyday competence are entirely contained in the hand-written constraints of the CSP generator; if a constraint is missing or wrong, CBTL can freely explore behavior that is unsafe while still appearing to personalize.","fun_headline_variants_meta":{"raw":{"variants":["Safe personalization via planning constraint null spaces","Robots personalize in the null space of safety constraints","CBTL: Few-interaction safe robot personalization","Learning user preferences without leaving safe solution space","Null-space constraints let robots adapt safely to users"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1529,"prompt_tokens":885,"completion_tokens":644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":583}},"tokens_in":501,"tokens_out":644,"duration_ms":6281,"temperature":1.0,"reasoning_tokens":583,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:15:41.641456+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give CBTL a CSP generator that omits a collision constraint between the robot arm and the user's torso during feeding, then run it in a physical or simulated feeding task; if the robot ever executes a torso-colliding configuration while satisfying the stated constraints and the user-adaptive objective, the claim that personalization stays in a safe null space is refuted.","supporting_citations":[{"cited_title":"Li and T","cited_arxiv_id":null,"evidence_quote":"Introduces the entropy-based active constraint learning that CBTL's exploration objective is built on."},{"cited_title":"Belkhale, E","cited_arxiv_id":null,"evidence_quote":"Offers the occlusion model used to implement the personalized occlusion-free constraint in feeding experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes user functionality embeddings that motivate and implement the reachability personalization used in the Books environment and feeding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents long-horizon bite preferences that seed the dip-preference personalization examples and user-study design."}],"review_version":1}