{"id":"ac5c7934-ff51-427a-a9fe-108d064afc3b","arxiv_id":"2411.10534","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The chain of alignment method derives model behavior rules from publicly supported objectives and yields an automated reward that tracks expert ratings of response alignment (r=0.841).","lead":"This paper introduces a process called chain of alignment that lets members of the public set goals for AI behavior while experts translate those goals into concrete rules. It tests the method on mental health prompts and reports that the resulting automated score closely matches expert judgments of whether responses honor public goals.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central inference from an expert-validated RBR to a measure of public will is untested: the public voted on objective texts, not on model responses, so the RBR may only approximate expert judgments.","rationale":"I read the paper in good faith and find the CoA framework genuinely valuable: it separates normative objectives from empirical rules, uses public deliberation for the normative layer, and transparently reports a plausible small-scale validation. The strongest empirical result—RBR correlating with expert judgments at r=0.841—is a solid first step. My main concern is not the linear independence assumption flagged by the reader, because the ablation in Appendix A.5 shows that the weighted and unweighted versions perform nearly identically (r=0.842 vs r=0.833), so the exact aggregation formula is not what carries the result. The more load-bearing gap is the final inference: the paper validates the RBR against expert judgments of objective alignment, but claims it measures alignment with public will. The public support data are about objective texts, not about model responses, so there is a missing link between objective-level endorsement and response-level scoring. This is not an internal inconsistency; it is an external validity gap. The paper's own limitation statements support this reading. A direct public-rating study would settle whether the RBR tracks public will or only expert-consistency. For that reason I do not recommend changing the reader's conditional verdict, but I would sharpen the stated condition: the central claim requires evidence that public raters and experts agree on response-level objective achievement.","tokens_in":11919,"tokens_out":8127,"duration_ms":88950,"concrete_test":"Recruit a new representative sample of the US public (same strata as Appendix A.4) and have participants rate a subset of the 65 prompt/response pairs for alignment with the ratified normative objectives, using the same 5-point scale described in Section 2.3. Compute the Pearson correlation and AUC between the average public rating and the CoA RBR, and compare these with the expert-based r=0.841 and AUC=0.964. If the public–RBR correlation is substantially lower than the expert–RBR correlation (e.g., the lower bound of its 95% confidence interval falls below 0.7), or if public ratings diverge from expert ratings on the same responses, the final inference from expert-validated RBR to public-will alignment is not supported. If the correlations are statistically indistinguishable, the central claim gains direct empirical support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a CoA RBR provides an approximate measure of alignment between LM behavior and public will (Abstract). The evidence for this rests on two separate results: (1) the public supports the normative objectives at 96–98% (Section 2.1), and (2) the RBR correlates with expert assessments of response–objective alignment at r=0.841 (Section 2.3). But nothing in the paper connects these two levels. The public was never asked to judge model responses—they only voted on objective texts—and the RBR is validated exclusively against expert holistic ratings. The chain therefore requires the unstated assumption that expert operationalization of the ratified objectives faithfully captures what the public would regard as achieving those objectives. If, for example, the public and experts disagree about which responses help a user feel validated or avoid impulsive actions, the RBR could be a valid measure of expert-consistency while failing to measure public will. Section 3 acknowledges that 'public support is an imperfect measure of public will' and that the small number of expert-evaluated responses limited the work, but it does not address the response-level gap. This is load-bearing because the headline conclusion—alignment with public will—depends on it, not on the linearity of rule aggregation. The Appendix A.5 ablation further suggests that the rule-objective weights contribute little (r=0.833 unweighted vs r=0.842 weighted), so the reader's identified linearity concern is less central than the public-will inference gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces 'Chain of Alignment' (CoA), a method for building rule-based rewards (RBRs) for language model behavior by factoring the alignment problem into normative objectives endorsed by the public and empirical rules created by domain experts. The authors report a public-input pipeline (collective dialogues, bridging-based ranking, and ratification thresholds) that yields normative objectives with 96–98% US public support across three mental-health domains, and expert-created rules that, when converted into an RBR via weighted rule-adherence scores (Eq. 1), correlate with expert holistic ratings of response-objective alignment (Pearson's r = 0.841, AUC = 0.964). The paper's central claim is that because the objectives have near-unanimous public support and the RBR tracks expert judgments of objective alignment, the RBR approximates alignment between LM behavior and public will.","tokens_in":12222,"tokens_out":2751,"duration_ms":28662,"significance":"If fully validated, the CoA framework would be a valuable contribution to pluralistic alignment: it provides a legible, decomposable alternative to black-box reward models and a concrete procedure for integrating nonexpert public input with expert knowledge. The paper is commendably transparent: it includes detailed appendices on the participant pipeline (A.2), demographic limitations (A.4), an ablation of the weighting scheme (A.5), a technical analysis of the normative-empirical conflation (A.7), and an explicit list of limitations in Section 3. The main validation step—correlating the RBR against expert judgments on 65 responses—is a genuine, non-circular check of the rule-weighted scheme, since the rule-objective weights were fixed before the comparison. However, the headline inference from objective-level public support to response-level public will is not directly tested, and the ablation results weaken the link between the RBR and the normative objectives that carry the public-will content.","major_comments":[{"comment":"The central claim that the CoA RBR 'provide[s] an approximate measure of alignment between LM behavior and public will' rests on an unvalidated link. The public voted on objective texts (Section 2.1), while the RBR was validated against expert holistic ratings of response-objective alignment (Section 2.3). Nothing in the paper measures public judgments of model responses or tests whether expert operationalization of the objectives matches how the public would apply them. The paper presents the expert evaluations as 'ground truth' for objective alignment, but that conflates expert judgment with public will. Section 3 acknowledges that 'public support is an imperfect measure of public will' but does not address this response-level gap. To support the headline claim, the authors need either a direct public-rating study or a clearly stated and defended argument that expert response judgments are a faithful proxy for the public's will—absent that, the conclusion should be reframed as measuring expert-consistency rather than public-will alignment.","section":"Abstract and Section 2.3"},{"comment":"The ablation in Appendix A.5 reports that removing the rule-objective alignment weights ϕ(r,J) from Eq. (1) changes the Pearson correlation with ground truth from r=0.842 to r=0.833, and the authors note the improvement is not statistically significant at N=65. Since the ϕ(r,J) weights are the only component of the RBR that connects rule adherence to the normative objectives (and hence to public will), this result indicates that the RBR's predictive power is essentially carried by the raw rule-adherence scores, not by the objectives-alignment weighting. The paper should either demonstrate that the weighting effect is statistically significant with a larger sample or acknowledge that the RBR is effectively measuring adherence to expert-authored rules, not alignment with the public-endorsed objectives.","section":"A.5 and Eq. (1)"},{"comment":"The quantitative evidence for both the public-support claim and the expert-correlation claim lacks appropriate uncertainty quantification. The public sample is described in A.4 as 'reasonably well balanced' on gender, age, and political party but skewed on ethnicity, education, religion, household income, and AI opinion; despite this, the abstract reports support as '96% ± 2%' without stating whether the quoted margin is a sampling margin or a demographic-bounding range. For the central correlation, N=65 with no confidence intervals on r or AUC; a reported r=0.841 could have a wide interval at this sample size. The authors should provide bootstrap confidence intervals or equivalent, and should report the demographic-adjusted range of public support rather than a simple margin of error.","section":"A.4 and Section 2.3"},{"comment":"The RBR formula Eq. (1) inherits the linear, independent rule-aggregation assumption stated in Eq. (15) of A.7: each rule's impact on objective probabilities is assumed independent and proportional to rule adherence. The paper labels this a limitation in Section 3, but it is load-bearing because the RBR is the only quantitative bridge from model behavior to the public-endorsed objectives. If rules interact or have nonlinear effects—plausible in mental-health contexts where multiple constraints may compound—the weighted sum in Eq. (1) can misrepresent objective achievement even when individual rules are well chosen. The authors should either provide evidence that the linear independence assumption holds for their rule sets (e.g., a residual analysis or a comparison with a learned aggregation on a larger ground-truth set) or restrict the paper's claims to the specific, non-interacting regime they assume.","section":"A.7, Eq. (15) and Section 3 (Rule-based reward)"}],"minor_comments":[{"comment":"The word 'liklelihood' is a typo; it should be 'likelihood'.","section":"Figure 1 caption"},{"comment":"'princlpled' should be 'principled'.","section":"Section 3"},{"comment":"There are several typos: 'sucide' should be 'suicide', and 'domian' should be 'domain'.","section":"A.2.1"},{"comment":"'we're assessed' should be 'were assessed', and 'Person's r' should be 'Pearson's r'.","section":"A.5"},{"comment":"The footnote describing the ratification thresholds as 'somewhat arbitrary' and 'handwavey' is candid but undercuts the precision of the high public-support numbers; the authors should either justify the thresholds more rigorously or soften the presentation of the ratification as a validated procedure.","section":"A.2.3"},{"comment":"The description of the three domains is clear, but the table in A.1 duplicates content from the main text; consider consolidating to avoid redundancy.","section":"A.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid workshop-level contribution with an honest appendix, but the gap between objective-level public support and response-level public will is the central scientific issue. The authors should be encouraged to either collect public response-level ratings or explicitly reframe the contribution as measuring alignment with expert-interpreted objectives, not with public will directly. The ablation result (A.5) should be reported more prominently, as it suggests the RBR's performance is largely independent of the objective-alignment weights that carry the public-will content."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this paper has a genuinely new organizing idea—split normative objectives from empirical rules—and a transparent, small-scale validation. But the abstract's leap from 'objectives have near-unanimous public support' to 'RBR measures alignment with public will' is not supported by the data. The public voted on objective statements; the RBR was validated against expert ratings of model responses. Nothing shows experts and the public would judge responses the same way. That is the load-bearing gap, and it is bigger than the linearity issue the authors flag.\n\nWhat's actually new: the normative-empirical factoring. Prior RBR and Constitutional AI work either wrote rules from experts or sourced principles from the public without cleanly separating what we want (objectives) from what behaviors achieve it (rules). CoA gives nonexperts authority over ends and experts authority over means. The process is described in enough detail to reproduce: collective dialogues, bridging-based ranking, expert refinement, ratification thresholds, and an explicit technical derivation in A.7. The ablation in A.5 is honest—weights barely move the correlation (0.833 vs 0.842)—which suggests the RBR is robust but also that the weighting scheme isn't doing much.\n\nSoft spots, in proportion. First, the response-level gap above. The paper says in Section 3 that public support is an imperfect measure of public will, but that is not the issue. The issue is that the second link—expert judgments of response-objective alignment—is never calibrated against public judgments of responses. If experts and the public disagree about what 'feels validated' means in a concrete response, the RBR measures expert consistency, not public will. Second, the headline correlation rests on 65 responses, no confidence intervals, and subjective expert ground truth. Third, the sample skews on ethnicity, education, religion, income, and AI opinions; the authors acknowledge this but still report 96-98% support as 'US public support.' Fourth, the ratification thresholds are admittedly arbitrary. None of these are fatal on their own, but together they support a conditional reading.\n\nWho this is for: people working on pluralistic alignment, participatory AI, and rule-based reward design. A serious referee should engage with it because the framework is novel and the limitations are mostly stated, albeit with an overclaiming abstract. My recommendation: send it to peer review, but require the authors to either temper the public-will claim or add a study where the public rates responses. The linear aggregation concern is real but secondary.\n\n— [name]","headline":"Genuinely new normative-empirical factoring, but the public-will measurement claim outruns the evidence: the RBR is validated against experts, not the public.","tokens_in":12749,"tokens_out":3496,"would_cite":true,"duration_ms":30124,"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":"Public will can be turned into a measurable target for language model alignment.","keywords":["language model alignment","public will","rule-based rewards","normative objectives","collective dialogue","mental health","normative-empirical conflation","alignment measurement"],"falsifier":"Take a set of mental-health prompts in which two rules are chosen so that following one reverses the other's effect (for example, a rule demanding immediate acknowledgment and a rule demanding brevity in crisis triage). Have experts rate the full set of responses holistically, compute the CoA RBR on the same set, and test whether the RBR's ranking matches the experts. If the independence assumption is materially false, the RBR should misrank the interaction-heavy cases even though each rule alone is aligned with the objectives.","tokens_in":11736,"feed_emoji":"🗳️","tokens_out":5290,"duration_ms":47953,"temperature":0.7,"pith_summary":"This paper introduces a method, called chain of alignment (CoA), to measure whether a language model's behavior matches what the public actually wants. The method splits the problem in two: the public expresses normative objectives (desired outcomes and values) through collective dialogues, and mental health experts convert those objectives into concrete behavioral rules. A rule-based reward (RBR) then scores any model response by how well it follows the rules, weighted by how much each rule serves the objectives. Across three mental-health prompt domains, the objectives drew support from 96 to 98 percent of a representative US sample, and the RBR's scores tracked mental-health experts' holistic judgments closely (Pearson r = 0.841, AUC = 0.964). If this holds, the same machinery could supply fine-tuning signals, online oversight, and pre-release safety checks.","feed_headline":"Public will becomes a measurable AI alignment target","feed_subtitle":"Chain of Alignment builds objectives from public votes, rules from experts, and a reward that matches expert ratings.","key_machinery":"The chain of alignment factorization and its reward formula. The paper defines normative objectives J (good outcomes, bad outcomes to avoid, and deontological values) as the link to public will, and empirical rules R as the link to behavior. For each domain, experts rate each rule's expected effect on each objective, producing a rule–objective alignment φ(r, J); the RBR in Eq. (1) is the weighted average of an LM grader's rule-adherence scores φ({x,y}, r), with those alignments as weights. The normative–empirical decoupling is what lets nonexperts set goals and experts pick behaviors without asking the public to predict outcomes.","core_discovery":"The central claim is that alignment between public will and language model behavior can be factored into a chain of three links—public will to normative objectives, normative objectives to empirical rules, and empirical rules to observable model behavior—and that the resulting rule-based reward yields a usable, scalable approximation of will-alignment. The paper argues that eliciting preferences directly on model outputs conflates what people want (normative judgments) with how they predict outputs will affect the future (empirical judgments), and that this conflation hides underlying consensus. Instead, the public votes only on objectives, experts assess which observable rules best achieve those objectives, and an LM grader checks whether a response follows the rules. The validation uses mental-health experts' averaged ratings of 65 responses across three risk-graded domains as ground truth, reporting Pearson's r = 0.841 and AUC = 0.964 for the RBR's ordering and classification.","pith_inferences":["My inference: the linear independence assumption is the natural pressure point; a version that learns rule-interaction weights from a large ground-truth set would settle how much the simplicity costs.","My inference: the public-support numbers likely depend on the specific dialogue design and ratification thresholds (75% overall, 66% bridging), so re-running the process with different deliberation formats would reveal how robust the near-unanimity is.","My inference: the approach's usefulness for pre-release safety checks depends on the domain taxonomy being stable; cross-domain prompts that straddle MH1–MH3 would test whether the RBR degrades when the domain label is ambiguous."],"forward_implications":["The CoA RBR can be evaluated at scale, so it can generate fine-tuning datasets, provide online oversight by blocking low-scoring outputs, and serve as a pre-release safety check.","Because the objectives carry near-unanimous public support (96–98% overall, 92–96% bridging support), the RBR doubles as an approximate measure of alignment with public will, not just with expert preferences.","The normative–empirical split means the public does not need to predict how a behavior will affect outcomes; experts supply that prediction, which makes agreement easier to find.","Ablation shows rule–objective weighting outperforms a uniform-weight version (r = 0.833 vs 0.842), suggesting the expert alignments add signal even when all rules are positively aligned.","The same structure could be extended to AI systems whose behavior outruns human understanding by swapping human experts for more capable models, since the public agency is carried by the objectives."],"supporting_citations":[{"why":"Constitutional AI supplies the baseline comparison for LM self-evaluation against normative principles and the alternative that CoA aims to improve on.","marker":"[13]"},{"why":"Rule-based rewards for language model safety supply the RBR methodology and the interaction-weight learning approach the paper's limitation section points to.","marker":"[15]"},{"why":"The collective-dialogue policy development process that CoA's objective-creation pipeline builds on and whose 76% support figure CoA compares against.","marker":"[18]"},{"why":"Collective Constitutional AI is prior art for sourcing normative principles from collective input, which CoA extends by separating objectives from rules.","marker":"[14]"},{"why":"Inverse constitutional AI is cited as the future direction for compressing preferences into principles more rigorously during rule creation.","marker":"[19]"},{"why":"Prolific is the participant recruitment platform that supplied the representative US public samples for the collective dialogues.","marker":"[22]"},{"why":"Elicitation inference is used to predict missing support votes and compute bridging agreement across demographic segments in objective generation.","marker":"[20]"}],"fun_headline_variants":["Chain of Alignment ties public will to AI rules","Public votes, experts craft rules, AI learns alignment","New metric measures AI alignment with public will","From public goals to expert rules: AI alignment chain","Aligning AI with public will via expert-built rules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole reward rests on assuming that each rule's effect on the chance of achieving an objective is independent of all other rules and linear in how well the response follows it; if rules interact or the dose–response is nonlinear, the weighted sum in Eq. (1) can be wrong even when every rule individually is well chosen.","fun_headline_variants_meta":{"raw":{"variants":["Chain of Alignment ties public will to AI rules","Public votes, experts craft rules, AI learns alignment","New metric measures AI alignment with public will","From public goals to expert rules: AI alignment chain","Aligning AI with public will via expert-built rules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1302,"prompt_tokens":947,"completion_tokens":355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":282}},"tokens_in":563,"tokens_out":355,"duration_ms":3414,"temperature":1.0,"reasoning_tokens":282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:35:29.624032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of mental-health prompts in which two rules are chosen so that following one reverses the other's effect (for example, a rule demanding immediate acknowledgment and a rule demanding brevity in crisis triage). Have experts rate the full set of responses holistically, compute the CoA RBR on the same set, and test whether the RBR's ranking matches the experts. If the independence assumption is materially false, the RBR should misrank the interaction-heavy cases even though each rule alone is aligned with the objectives.","supporting_citations":[{"cited_title":"Rule- based rewards for language model safety, 2024","cited_arxiv_id":null,"evidence_quote":"Rule-based rewards for language model safety supply the RBR methodology and the interaction-weight learning approach the paper's limitation section points to."},{"cited_title":"Qiu, Michael Varga, and Aviv Ovadya","cited_arxiv_id":null,"evidence_quote":"Elicitation inference is used to predict missing support votes and compute bridging agreement across demographic segments in objective generation."}],"review_version":1}