{"id":"7f593cfe-b065-431f-8ba6-58804c69976b","arxiv_id":"2411.15230","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For calibrated binary predictors, any collaboration rule that is guaranteed to be at least as accurate as the worst agent must essentially always defer to a single agent.","lead":"A new theorem shows that no deterministic rule for combining calibrated human and AI probability estimates can guarantee matching the worst individual performer, unless it essentially always trusts the same person. The result means complementarity in human-AI teams cannot be achieved for free; you need extra structure such as independence or learning how the two predict jointly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stated 'iff' in Theorem 1 is false: Definition 4's boundary freedom lets non-collaborative strategies be unreliable; forward impossibility proof remains valid.","rationale":"The central one-way impossibility result is well supported: Lemma 8 and Proposition 9 construct calibrated settings for any violating tuple, and Proposition 6 correctly combines them. The unrestricted-distribution assumption the reader flags as weakest is intentional for a no-free-lunch theorem; it is the source of the adversarial constructions, not an internal flaw. The genuine formal problem is that the paper's Definition 4 and restated Theorem 1 are inconsistent: non-collaborative permits arbitrary boundary behavior, but reliability does not. The n=1 counterexample is decisive and shows the 'iff' statement must be corrected. Since the forward proof is unaffected and the fix is localized, the conditional verdict remains appropriate; no change from the reader's recommendation is needed.","tokens_in":11638,"tokens_out":20208,"duration_ms":200126,"concrete_test":"Instantiate the counterexample and check both directions: verify that C(p)=floor(p) on (0,1), C(1)=0 is non-collaborative under Definition 4; verify reliability fails on the one-point calibrated setting (P_1=1, Y=1), giving acc_C=0 < acc_1=1. If confirmed, revise Theorem 1's 'iff' by either restating it as the one-way implication or adding the missing requirement that C defer to any agent with p_i∈{0,1}; then re-check that the earlier constructions (Lemma 8, Proposition 9) remain valid under the revised definition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 4 defines non-collaborative only on interior tuples (0,1)^n and leaves C(p) arbitrary when some p_i ∈ {0,1}. The restated Theorem 1 in Section 4.2 asserts reliability if and only if (i)-(ii), but the reverse implication is false. Take n=1: C(p)=floor(p) for p∈(0,1) with C(1)=0 and C(0)=1. This satisfies Definition 4 (k=1, alpha=1 at p=1/2). Consider X={x}, D the point mass with P_1(x)=1 and Y=1. P_1 is calibrated, acc_1=1, but C outputs 0 on the only input, so acc_C=0<1; C is not reliable. The paper's own Section 2 discussion says the certainty exception should be handled by deferring to a fully confident agent, but that mandatory deferral is absent from Definition 4 and from condition (ii) of the restated theorem. Consequently Theorem 1's characterization is false as written. The forward direction (reliability implies non-collaborative) appears sound: Lemmas 8 and Proposition 9 construct valid finite calibrated settings, and the linear-combination argument in Proposition 6 correctly glues them. Thus the core one-way no-free-lunch result survives, but the formal statement overstates it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies deterministic collaboration strategies C:[0,1]^n -> {0,1} that combine calibrated probabilistic predictions from n agents. A strategy is called reliable if, in every collaboration setting, its 0-1 accuracy is at least the minimum accuracy of the individual agents. The main result, Theorem 1, claims that every reliable strategy is non-collaborative: it must essentially always follow the threshold classification of a single fixed agent, with only a fixed tie-break at p_k=1/2 and an exception when some agent is fully certain. The proof proceeds by showing that any violation of the non-collaborative condition lets one construct finite collaboration settings in which the strategy is strictly worse than the violating agent and no better than the others, and then combines these settings using linear mixtures. The paper uses the theorem to argue that complementarity cannot be achieved for free from calibration alone, and that successful collaboration requires independence or learned joint information.","tokens_in":1272,"tokens_out":3366,"duration_ms":149557,"significance":"If the one-way theorem is correct, this is a valuable conceptual result. The proof is self-contained and gives explicit finite constructions (Lemma 8, Proposition 9, Proposition 6); it does not rely on fitted parameters or circular assumptions. The paper is also appropriately cautious in Section 5 about the unrestricted distribution over X × {0,1}. The broad message, that calibration alone does not justify deferring to a confident-but-not-certain agent, is important for human-AI collaboration. However, the manuscript currently overstates the result by presenting an 'if and only if' characterization that is false, and this must be corrected before publication.","major_comments":[{"comment":"The restated Theorem 1 in Section 4.2 is an 'if and only if' characterization, but the reverse implication is false as written. Definition 4 constrains C only on the interior (0,1)^n, leaving the boundary behavior free, so a strategy satisfying the stated interior conditions can be unreliable. For example, for n=1 define C(p)=1 for p>1/2, C(p)=0 for p<1/2, C(1)=0, C(0)=1, and C(1/2)=1. This satisfies conditions (i) and (ii) with k=1 and alpha=1. Now take X={x}, D the point mass with P_1(x)=1 and Y=1. P_1 is calibrated and acc_1=1, but C outputs 0 at the only input, so acc_C=0<1. Thus reliability does not follow from the stated conditions. The forward direction, proved via Lemma 8 and Proposition 9, appears sound, but the theorem should be restated as a one-way implication. Alternatively, Definition 4 should require C(p)=p_i whenever p_i is 0 or 1, which is what the informal discussion in Section 2 promises.","section":"4.2 (Theorem 1)"},{"comment":"The informal claim that the only exception is deferring to a fully confident agent is not reflected in the formal definition, which permits arbitrary boundary choices when some p_i is 0 or 1. Because a calibrated predictor with P_i(x)=1 implies the label is 1 almost surely, any reliable strategy must output 1 on such points. The current definition allows boundary choices that break reliability, and this is exactly what makes the 'if' direction false. The formal definitions and the theorem should be aligned so that the promised certainty exception is part of the characterization, or the theorem should be stated without the converse direction.","section":"2 (Definition 4 and the discussion after Theorem 1)"}],"minor_comments":[{"comment":"In the proof of Proposition 6, 'first randomly sampling m with probability λℓ' should read 'with probability λ_m'; the subscript typo makes the mixture construction harder to follow.","section":"Proposition 6 proof"},{"comment":"In the second case of the proof, 'agent i is incorrect on x=0 and correct on x=1' should read '... and correct on x=i', matching the table and the input space X={0,1,...,n}.","section":"Lemma 8 proof"},{"comment":"The important qualification that the result places no restrictions on the distribution over X × {0,1} appears only in the open problems. Consider stating this assumption earlier, since it is central for interpreting the theorem as a no-free-lunch result.","section":"5 (Open Problems)"}],"recommendation":"major_revision","confidential_remarks":"The one-way no-free-lunch theorem is the paper's core contribution and appears sound, with careful finite constructions. The false 'iff' statement is localized and fixable: the authors should either state the theorem as a one-way implication or add a boundary-deferral condition to Definition 4 and the restated theorem. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new impossibility result, and the main forward theorem is proved carefully. The restated “iff” in Section 4.2 is false as written, but it is a fixable overstatement; the one-way no-free-lunch claim survives.\n\nWhat’s new: I don’t know of a prior result showing that a deterministic aggregator of calibrated probabilities that is reliable for all distributions must be a single-agent dictator on interior tuples. The proof is self-contained and constructive, using finite input spaces and calibrated predictors induced by partitions. That is formal and reproducible in the sense that the constructions are explicit. I also buy the interpretation: calibration is a per-agent marginal property; it doesn’t constrain joint behavior, so complementarity needs independence or learned joint structure. That’s a useful way to read the empirical human-AI collaboration literature.\n\nSoft spots: the stated iff in Section 4.2 is wrong. Definition 4 lets C behave arbitrarily when some p_i is 0 or 1, and condition (ii) of the restated theorem only covers interior tuples. As a result, a non-collaborative C can fail reliability by mishandling certain inputs. The n=1 example in the stress test is enough: C(p)=floor p on (0,1), C(1)=0, C(0)=1 is non-collaborative by Definition 4 but fails on a point mass at label 1. The fix is to require deferring to any fully confident agent, which the paper’s own Section 2 discussion says is the intended design. That should be made part of the formal characterization. Also Eq (1) calls an expectation of 0-1 error “accuracy”; it is an error rate, so the inequality directions in the definitions need a careful sign check. Minor, but confusing.\n\nI don’t see a deeper flaw. The forward impossibility proof uses valid finite calibrated settings and gluing via Proposition 6; the cases in Lemma 8 hold. The paper also openly flags the arbitrary-distribution assumption and suggests restricted distributions, multi-class problems, and other losses as open problems, so the scope is stated honestly.\n\nWho for: people working on human-AI complementarity, learning to defer, ensemble methods, and impossibility results in prediction aggregation. It won’t change practice by itself, but it clarifies why calibration alone is insufficient. I’d send it to a serious referee; the iff fix is a minor revision, not a desk reject.","headline":"New one-way impossibility result is real and worth engaging; the restated iff is false as written but fixable.","tokens_in":12384,"tokens_out":1665,"would_cite":true,"duration_ms":16583,"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":"The only reliable way to combine calibrated human and AI judgments is to let one agent decide.","keywords":["human-AI collaboration","complementarity","calibrated predictions","no free lunch theorem","binary classification","decision aggregation","learning to defer","impossibility theorem"],"falsifier":"A concrete falsifier would be a non-collaborative strategy $C$ that is not just deferring to one agent, together with a proof or exhaustive computation that $C$ meets the worst-agent accuracy on every collaboration setting. In particular, take $C$ to round the average of two calibrated probabilities and search over finite input spaces with partition-induced calibrated predictors; the paper's Lemma 8 construction implies the search will always find a setting where $C$ underperforms the worst agent, so exhibiting one such setting where it does not would contradict the proof.","tokens_in":1548,"feed_emoji":"🤝","tokens_out":4416,"duration_ms":92899,"temperature":0.7,"pith_summary":"This paper proves a no-free-lunch result for human-AI collaboration in binary classification. If each agent produces calibrated probability predictions, no deterministic rule for combining those probabilities into binary classifications can be guaranteed to match the accuracy of the worst individual agent unless that rule always follows one fixed agent. The result matters because the hoped-for benefit of human-AI teams—complementarity, beating both members alone—therefore cannot be achieved for free. It requires extra structure, such as independence between agents' predictions or learned knowledge of how their judgments jointly relate to outcomes. The only automatic exception is when another agent is entirely certain, which leaves a narrow but practical model of collaboration through overriding obvious errors.","feed_headline":"Reliable AI-human teams can only defer to one agent","feed_subtitle":"Any true fusion of calibrated predictions can underperform the worst member; complementarity needs extra structure.","key_machinery":"The proof constructs adversarial collaboration settings from finite input spaces and partitions of that space. A partition $\\mathcal A_i$ induces a calibrated predictor $P_i(x) = \\Pr[Y=1 \\mid X \\in A_i]$, so a setting can be specified by $(D, \\mathcal A_1, \\dots, \\mathcal A_n)$. For any tuple on which $C$ disagrees with agent $k$, the paper builds a distribution over $X = \\{0,1,\\dots,n\\}$ so that $C$ is wrong at a distinguished point while agent $k$ is at least as accurate; a linear-combination lemma (Proposition 6) then averages these per-agent settings into a single setting where $C$ is worse than every agent. The same construction is reused to force a fixed tie-break at $p_k = 1/2$.","core_discovery":"The central claim is Theorem 1: for binary classification with $n$ calibrated agents and a deterministic collaboration strategy $C:[0,1]^n \\to \\{0,1\\}$, if $C$ is reliable—meaning it achieves accuracy at least the minimum of the agents' accuracies for every collaboration setting—then $C$ is non-collaborative. In the paper's terms, there exists an agent $k$ such that for every interior tuple of probabilities, $C$ equals the threshold classification $\\lfloor p_k \\rceil$, with a fixed tie-break at $p_k = 1/2$ and exceptions only when some predictor is exactly $0$ or $1$. A collaboration strategy that genuinely combines agents—by averaging probabilities, taking majority votes, or deferring to the most confident agent—will sometimes perform strictly worse than the least accurate agent.","pith_inferences":["A practical upshot the paper only states implicitly: displaying an AI confidence score to a human is unlikely to improve joint accuracy unless the system also learns, from paired human-AI decisions, when the human's judgment is more reliable than the AI's.","The certainty-override exception suggests a testable design: give an AI veto power only on inputs where a trustworthy model outputs probability exactly $0$ or $1$, with the human deciding everywhere else; the theorem predicts this is the only guaranteed-safe protocol under calibration alone.","The result transfers to ensembles of machine-learning models: averaging calibrated probability outputs is not universally reliable, so ensembles must either decorrelate their members or learn a weighting from validation data, mirroring the paper's independence-versus-learning dichotomy.","Because the theorem relies on unrestricted input distributions, a natural research direction is to characterize restricted families—real feature geometries, independence structures, or known correlations—under which non-deferring rules such as averaging become reliable; the paper explicitly leaves this open."],"forward_implications":["Averaging probabilities, majority vote, and deferring to the most confident agent are all unreliable: each can perform worse than the least accurate calibrated agent on some setting.","Complementarity in prediction requires extra structure beyond calibrated individual predictions, specifically either independence among agents' signals or learned knowledge of the joint distribution of predictions and outcomes.","The only guaranteed collaboration protocol under the paper's assumptions is to give one agent primary authority and let another agent override only when it is fully certain, i.e. predicts probability exactly $0$ or $1$.","Common human-AI implementations that simply show a human a calibrated algorithmic probability lack both enabling conditions, so they should not be expected to deliver complementarity on their own.","Guarantees for ensemble and expert-combination methods trace to exactly these two structures, which explains why those methods succeed while general calibrated aggregation cannot."],"supporting_citations":[{"why":"Supplies the no-free-lunch framing that motivates reading the theorem as a trade-off in performance across collaboration settings.","marker":"Wolpert and Macready, 1997"},{"why":"Provides the majority-vote positive result that the theorem contrasts with, identifying independence as the enabling condition.","marker":"de Condorcet, 1785"},{"why":"Represents ensemble methods whose success relies on independence among individual predictors.","marker":"Breiman, 2001"},{"why":"Represents expert-combination methods that work by learning which experts to trust over time.","marker":"Blum, 2005"},{"why":"Introduces a learning-to-defer approach that requires joint-distribution data, the additional ingredient Theorem 1 says is necessary.","marker":"Madras et al., 2018"},{"why":"Shows complementarity can be achieved when agents have distinct regions of advantage, a condition the theorem says cannot be inferred from calibrated predictions alone.","marker":"Donahue et al., 2022"}],"fun_headline_variants":["There is no free lunch for human-AI collaboration","Fusing AI and human predictions can underperform the worst agent","AI-human teams must defer to one agent to avoid failure","Complementary AI-human collaboration is not achievable for free","Combining calibrated predictions can yield worse accuracy than any single agent"],"cache_read_input_tokens":14592,"weakest_assumption_plain":"The theorem's impossibility depends on letting the joint distribution of inputs and labels be completely arbitrary and chosen after the collaboration rule is fixed; if the distribution were restricted—say to real features, independence, or known correlation between agents—then some non-deferring rule could be reliable.","fun_headline_variants_meta":{"raw":{"variants":["There is no free lunch for human-AI collaboration","Fusing AI and human predictions can underperform the worst agent","AI-human teams must defer to one agent to avoid failure","Complementary AI-human collaboration is not achievable for free","Combining calibrated predictions can yield worse accuracy than any single agent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000826,"raw_usage":{"total_tokens":3563,"prompt_tokens":848,"completion_tokens":2715,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":2634}},"tokens_in":464,"tokens_out":2715,"duration_ms":19661,"temperature":1.0,"reasoning_tokens":2634,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:14:12.464002+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier would be a non-collaborative strategy $C$ that is not just deferring to one agent, together with a proof or exhaustive computation that $C$ meets the worst-agent accuracy on every collaboration setting. In particular, take $C$ to round the average of two calibrated probabilities and search over finite input spaces with partition-induced calibrated predictors; the paper's Lemma 8 construction implies the search will always find a setting where $C$ underperforms the worst agent, so exhibiting one such setting where it does not would contradict the proof.","supporting_citations":[],"review_version":1}