{"id":"449f1689-c834-4d06-97ae-add067efd0d7","arxiv_id":"2507.03190","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A machine learning framework called CLP discovers, improves, and tailors algorithms by chaining computational tokens with MCTS and RL, with strong results on the Quadratic Assignment Problem and quantum search.","lead":"This paper introduces a system that treats algorithms as sentences made of small computational tokens, then uses game-playing AI methods to chain the tokens into better algorithms. It reports rediscovering standard heuristics and beating commercial solvers on a hard assignment problem, and finding a shorter version of Grover's quantum search circuit.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 389/390 and 94.8% QAP win rates are measured against weak, underspecified baselines (SA, simple BB, Gurobi-on-MIQP) without seeds or variance; they do not yet establish the central claim of substantial outperformance over existing methods.","rationale":"The reader's verdict is CONDITIONAL, and my read leaves that unchanged. The most load-bearing concern is not the internal mathematics: the Grover circuit is a correct basis change, and the paper discloses its equivalence. The grammar-expressivity issue identified by the reader is real, but it is secondary: even if the grammar is sufficiently expressive, the paper has not yet shown that the produced algorithms outperform genuinely strong baselines. I credit the clear explanations of primitives and A-BPE, the disclosed equivalence of the Grover circuit, and the enough detail in the QAP setup to make the results plausible. The QAP comparison is the weak point because the central claim's quantitative force depends entirely on beating credible existing methods, and the manuscript compares only against SA, a simple BB, and a generic MIQP Gurobi formulation, with no statistical reporting. The reader flagged the Gurobi comparison as an overclaim, which overlaps with this concern, but the reader's weakest_assumption focused on grammar expressivity and search budget; my concern is more directly about the empirical yardstick. If the code release and stronger baseline rerun confirm Table S4, the central claim would be much better supported. If they do not, the correct verdict is CONDITIONAL with the overclaims removed or REJECT, depending on how far the numbers move. For now, the existing CONDITIONAL verdict is appropriate.","tokens_in":22009,"tokens_out":7991,"duration_ms":107841,"concrete_test":"Release the exact token grammar, neural-network hyperparameters, seeds, and per-instance logs, and rerun the QAPLIB experiment on the same hardware with a matched wall-clock budget against (i) Gurobi 10 with tuned MIQP parameters, (ii) a state-of-the-art tabu or hybrid genetic QAP solver, and (iii) a branch-and-bound using a stronger bound. If CLP's median per-instance gap and win rate do not reproduce Table S4 and do not beat these baselines, the 'substantially outperform existing methods' claim should be withdrawn or restricted to the tested baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central QAP claim is empirical, so its load-bearing condition is that the comparison protocol is a fair and strong test of 'substantially outperform existing methods.' That condition is not met in the manuscript as written. SI §1.2.3 lists only simulated annealing, a simple branch-and-bound with the Gilmore-Lawler bound, and Gurobi on a MIQP formulation with 'runtime set to be the same as our method.' No specialized QAP state-of-the-art heuristics (tabu search, iterated local search, hybrid genetic/GRASP, FANT) and no exact solver with strong bounds are included. The paper reports aggregate win rates (389/390 PQAP/CQAP, 94.8% QAPLIB) but no per-instance standard deviations, seeds, hardware, solver parameters, or time-limit details. The phrase 'beats or equals' also includes ties on easy instances, so the headline 'beats Gurobi' overstates the tables. Unless CLP is compared with the best available QAP solvers under matched resources, the strong quantitative claims in the abstract are not supported even if the reported tables are internally correct. A separate localized overclaim in SI §2.6 — 'reducing the required number of qubits by a factor of two' — is inconsistent with the depth-reduction result, but it is not what the main argument rests on.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces computational language processing (CLP), a framework in which algorithms are represented as token chains over a primitive vocabulary constrained by a grammar, and discovered by an ensemble variant of Monte Carlo tree search guided by reinforcement learning, with algorithmic byte-pair encoding used to grow the vocabulary. The authors report three application studies: the Quadratic Assignment Problem, quantum search (Grover), and QAOA. The reported highlights are optimal or near-optimal results on CQAP/PQAP instances, 41% optimality on a QAPLIB subset, a Grover circuit with roughly half the depth of the standard one, and an average 34.62% improvement over ADAPT-QAOA on 21 MAX-CUT graphs. The paper also shows that low-level primitives can be recombined to rediscover Frank-Wolfe and 2-OPT/TSP-style heuristics.","tokens_in":22352,"tokens_out":10801,"duration_ms":120361,"significance":"If the empirical claims are upheld, CLP would be a valuable contribution to automated algorithm discovery, particularly the combination of a token grammar, ensemble MCTS, and algorithmic BPE in a single framework, together with instance-adaptive policy generation. The paper ships detailed token decompositions and a supplementary table for QAPLIB comparisons, and the rediscovery of Frank-Wolfe and 2-OPT from elementary primitives is a concrete, checkable demonstration of the babbling-plus-BPE mechanism. I also note that the evaluation is not circular in the sense flagged by the reader: the QAP comparison is against external benchmarks with known optima, and the rediscovered algorithms are not used as predictions. The main weakness is that the strength of the reported evidence is not yet commensurate with the breadth and strength of the claims: the QAP baselines are weak and under-reported, one headline statement is contradicted by the paper's own supplementary table, and the quantum 'discovery' is a re-coordinatization of Grover's algorithm rather than a new algorithm.","major_comments":[{"comment":"The statement that 'CLP outperforms all baselines—including commercial solvers such as Gurobi' is contradicted by Table S4. On chr25a, chr22a, chr20a, chr18a, chr15c, and tai64c, Gurobi reports a strictly smaller relative gap than CLP (e.g., tai64c: CLP 1.2% vs. Gurobi 0.4%; chr15c: CLP 4.6% vs. Gurobi 0.0%). The SI statement that 'Gurobi beats our method in only one test problem of size larger than 25' is also incorrect because tai64c has size 64. The aggregate phrase 'beats or equals' also counts ties (many rows show CLP and Gurobi both at 0.0), so the abstract's wording overstates the tables. The authors should replace aggregate win rates with per-instance win/tie/loss counts, reconcile the main text with Table S4, and temper the abstract accordingly.","section":"Main text, 'Results on the QAP'; Table S4"},{"comment":"The comparison protocol is not strong enough to support 'substantially outperform existing methods' for the QAP. The baseline set is limited to simulated annealing, a simple branch-and-bound with the Gilmore-Lawler bound, and Gurobi on an MIQP formulation with runtime set equal to CLP; no state-of-the-art QAP heuristics (tabu search, iterated local search, hybrid genetic/GRASP, FANT) and no exact solvers with strong bounds are included. The SA baseline is described as run 'for many more iterations' (i.e., not resource-matched), while no seeds, per-instance run-to-run variance, hardware, solver parameters, or time limits are reported for any method. Because the central QAP claim is empirical, this is a load-bearing gap: the authors should add strong baselines under matched compute, report multiple seeds with standard deviations, and state the exact QAPLIB subset and success criterion used for the 41% and 94.8% figures.","section":"SI §1.2.3 and Table S4"},{"comment":"The quantum search result is a known equivalence rather than a newly generated algorithm. The SI analysis shows that the optimized circuit starts from |1>^⊗n and uses D' = H^⊗n(2|1><1|-I)H^⊗n, and that after each iteration the state differs from standard Grover only by a global phase (-1)^{|w|}. This is a valid circuit optimization, but it is a re-coordinatization of Grover's algorithm within the same gate set, not a new quantum search procedure. Moreover, SI §2.6 states that the method reduces 'the required number of qubits by a factor of two,' which is false—the circuit uses the same n qubits—and is inconsistent with the depth-reduction claim in the main text. The main text should either present this result as rediscovery/optimization of an existing algorithm (with correct metrics) or remove the 'new algorithms' language for this example.","section":"SI §1.3 and §2.6"},{"comment":"The QAOA result is reported without sufficient experimental detail to be verified. The 21 test graphs are not described (sizes, degree of the p-regular graphs, number of qubits), the maximum circuit depth and optimization protocol are not specified, and no variance or number of independent runs is given. The 'average improvement of 34.62%' is a point estimate over a single table with no confidence interval or statistical test. Since the abstract's claim of substantial improvement over 'foundational quantum computing approaches' rests in part on this result, the authors should provide the full protocol and error bars, or explicitly mark the result as preliminary.","section":"SI §1.4 and Table S2"},{"comment":"The grammar that defines valid token chains is never formally specified. The text describes constraints (e.g., special primitives operate on a single primitive, input/output type compatibility) and gives examples, but no production rules or type system are provided. Since the grammar is a core component of CLP and directly determines the expressivity of the search space, its absence makes the framework difficult to reproduce or extend. The authors should provide a formal grammar (or a precise algorithmic description of the type-checking step) in the supplementary material.","section":"SI §1.2.1"}],"minor_comments":[{"comment":"The phrase 'serves as the the alphabet' contains a duplicated article; also correct 'independant' in SI §1.1 and 'Additionaly' in the main text.","section":"Main text, 'Computational Language Processing' section"},{"comment":"The x-axis label 'Iterations' is ambiguous; it should state that one round of BPE occurs every 10 self-play iterations and clarify whether the success rate is computed over a fixed set of instances or over a distribution of instances.","section":"Figure 4 and SI §1.2.1"},{"comment":"The sentence 'our achieves 1% optimality gap' is missing a word; more importantly, the exact time limits for Gurobi and branch-and-bound are not reported numerically, so the resource-matching claim cannot be checked.","section":"SI §1.2.3"},{"comment":"The 'scaling quadratically' statement and the SI §2.6 estimates of '10,000-fold reduction in FLOPS' and 'thousand-fold quality improvement' are asserted without a derivation or a benchmark protocol; they should be removed or substantiated with profiling data.","section":"Main text, 'Results on the QAP'; SI §2.6"},{"comment":"The paper does not include a data/code availability statement. For an empirical algorithm-discovery claim, releasing the token vocabularies, trained checkpoints, and evaluation scripts would substantially strengthen verifiability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a strong central idea and the QAP rediscoveries are convincing, but the current evidence does not support the abstract's strong claims. I would ask the authors to (1) correct the Gurobi contradiction and the qubit-reduction error, (2) add a serious QAP baseline comparison with matched resources and variance, and (3) soften the 'new algorithms' language for the Grover circuit. These issues are fixable within the manuscript's scope, so I do not recommend rejection; however, the claims in the abstract and conclusion currently outrun the data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2507.03190. First, the genuinely new capability is instance-adaptive algorithm discovery by token chaining with ensemble MCTS + RL and byte-pair vocabulary expansion. Second, the central quantitative claims are not yet supported by the evaluation as written: the baselines are too weak, there are no seeds or variance estimates, and the \"beat Gurobi\" headline overstates the paper's own tables.\n\nWhat the paper does well: the A-BPE mechanism is a real idea. Starting from eight low-level primitives and rediscovering Frank-Wolfe and k-OPT, plus the randomized-restart variant that reaches 93% on CQAP-40, is a genuinely interesting demonstration. The step-size experiment discovering cyclical schedules is also a nice proof of concept. The SI is detailed enough that a determined reader could reimplement the machinery. The framework's output is a token chain, not a generated code blob, which is a meaningful difference from AlphaEvolve-style LLM agents.\n\nWhere it softens: the QAP comparison in Table S4 is against simulated annealing, a simple branch-and-bound, and Gurobi on a MIQP formulation with the same wall-clock budget. Those are not the state-of-the-art QAP solvers; tabu search, iterated local search, and hybrid GRASP variants are standard in this literature and are absent. So \"substantially outperform existing methods\" is not established. Second, the paper gives no seeds, no standard deviations, no hardware or hyperparameter details for the baselines; \"beats or equals in 389/390\" cannot be evaluated as a statistical claim. Third, the text says CLP \"outperforms all baselines\" while Table S4 shows Gurobi with lower gaps on chr25a, chr22a, chr20a, chr18a, and tai64c. The \"beats or equals\" formulation gets stretched into \"outperforms\" in the summary. Fourth, the quantum result is a basis change they explicitly prove equivalent in Sec 1.3; calling it a discovered algorithm is generous. And SI Sec 2.6 contains a localized overclaim that the method \"reduces the required number of qubits by a factor of two\", which is inconsistent with the depth-reduction result.\n\nNone of this is fatal to the core idea, but the paper needs code/data release, a fair baseline suite, and a rewritten abstract before the claims carry weight.\n\nWho it's for: researchers in automated heuristic design and anyone working on QAP heuristics. It deserves a serious referee, not a desk reject, but it needs major revision. I'd send it out, with a strong request for reproducibility artifacts.","headline":"A genuinely interesting framework for instance-adaptive algorithm discovery, but the headline QAP claims rest on weak baselines and missing variance, so the paper needs major revision and code/data release before its central claims carry weight.","tokens_in":22889,"tokens_out":2179,"would_cite":false,"duration_ms":24541,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C27","68Q12","68T20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that treating algorithms as grammatically chained tokens lets a reinforcement-learning-guided search rediscover, improve, and invent algorithms that match or beat Gurobi on 389 of 390 hard instances, halve Grover's…","keywords":["algorithm discovery","computational language processing","Monte Carlo tree search","reinforcement learning","quadratic assignment problem","quantum search","quantum approximate optimization algorithm","byte-pair encoding"],"falsifier":"Run CLP on a held-out family of quadratic assignment instances with known optima that comes from neither the CQAP nor the PQAP generator, for example instances built from uniformly random flow and distance matrices, with Gurobi given the same wall-clock budget, and then inspect the learned token chains: if the optimality-gap advantage over Gurobi disappears on this distribution, or if removing the A-BPE-created composite tokens leaves performance unchanged, the claim that the framework generates genuinely new and generalizable algorithms rather than tuned recombinations of the supplied primitives would be settled against it.","tokens_in":21791,"feed_emoji":"⚙️","tokens_out":23220,"duration_ms":227911,"temperature":0.7,"pith_summary":"The paper sets out to show that algorithm discovery can be automated by treating algorithms as sequences of elementary operations, or tokens, chained together under a grammar — in effect, sentences in a computational language. It claims that an ensemble Monte Carlo tree search guided by reinforcement learning, paired with a byte-pair-style step that merges frequent token sequences into new composite tokens, can explore this language well enough to rediscover classical heuristics (Frank–Wolfe, k-OPT), improve them with instance-specific adaptations (random restarts, a new cyclic step-size schedule), and generate novel circuits and solvers. On the Quadratic Assignment Problem, the discovered policies match or beat all tested baselines, including the commercial solver Gurobi, on 389 of 390 instances and solve 41% of QAPLIB benchmarks to optimality; for quantum search the framework finds a Grover circuit whose depth is reduced by $(\\pi/2)\\sqrt{N}-1$ layers, and for QAOA it beats ADAPT-QAOA by an average of 34.62% on MAX-CUT instances. The wider interest is that the approach abstracts away from code and from the specific problem, so the same machinery could in principle be pointed at any domain with a tokenizable operation set — the authors' own conclusion points toward systems engineering and mathematical discovery.","feed_headline":"389 of 390: token-chain search tops Gurobi","feed_subtitle":"The same token-chain search rediscovers heuristics, nearly halves Grover depth, and improves QAOA by 34.6%.","key_machinery":"The load-bearing machinery is the computational language itself. Its alphabet is a set of primitives — for the QAP, matrix-level operations (identity, gradient, negative, linear-sum assignment via the Hungarian method) plus special control-flow wrappers (k-iteration loops, residual updates, parallel random permutations, parallel 2-city swaps, and an early-stopping token), all constrained by a grammar that keeps special primitives acting on a single primitive at a time. The search engine is an ensemble variant of the AlphaZero-style MCTS in which tree edges are tokens, each node stores an ensemble of computational states, and lightweight transformer policy and value networks score sequences of actions; a computational-cost budget is fed into the search and into the networks. The vocabulary grows through Algorithmic Byte-Pair Encoding, which merges frequently co-occurring token pairs into single composite tokens, expanding 8 primitives into more than 100 tokens over successive rounds and thereby making higher-complexity algorithms such as Frank–Wolfe expressible. Alternating babbling (MCTS rollouts) with word-formation (BPE merges) is what lets the framework climb from elementary steps to full algorithms.","core_discovery":"On the paper's own terms, the discovery is that algorithm design can be reframed as search over a computational language, and that this reframing pays off across very different domains. An algorithm is a sentence: a grammatically valid chain of primitives — identity, gradient, negative, linear-sum assignment (Hungarian), and special control-flow wrappers such as loops, residual updates, parallel random permutations, parallel 2-city swaps, and early stopping. New words are formed by Algorithmic Byte-Pair Encoding, which merges frequently co-occurring token pairs into single composite tokens, growing the initial vocabulary of 8 primitives to more than 100 tokens and thereby making higher-complexity algorithms expressible. The search engine is an ensemble variant of the AlphaZero-style MCTS in which edges are tokens, nodes hold ensembles of computational states (Gibbs-averaged losses) to cope with stochastic actions, and lightweight transformer networks predict policy and value from the action history under an explicit compute budget. With this machinery the paper reports rediscovering Frank–Wolfe and k-OPT, learning random-restart strategies that push success rates above 90% on CQAP instances of size 40, discovering a cyclic step-size schedule that outperforms the scipy line-search baseline, and producing instance-adapted policies that reach global optima on all tested CQAP instances with $n \\le 80$, stay within a 1% optimality gap on PQAP, match or beat every tested baseline on 389 of 390 instances, and solve 41% of QAPLIB instances to optimality. In the quantum domain it finds a Grover circuit, mathematically equivalent to the standard one up to a global phase, whose depth is reduced by $(\\pi/2)\\sqrt{N}-1$ layers (initializing in $|1\\rangle^{\\otimes n}$ and implementing the diffusion step as $H^{\\otimes n}\\,\\mathrm{MCZ}\\,H^{\\otimes n}$, removing the per-iteration X gates), and QAOA mixer circuits that beat ADAPT-QAOA by an average 34.62% on unseen MAX-CUT instances.","pith_inferences":["A test the paper does not run is whether A-BPE tokens learned on QAP instances transfer to other problems; an editor would expect the high-level composite tokens to transfer while the instance-specific adaptations do not, and checking that would show where the genuine discovery actually happens.","The paper proves the discovered Grover circuit is the standard algorithm up to a global phase introduced by a different initial state, so the quantum result is best read as compressibility of a known strategy; a sharper novelty test, not in the paper, would be to run CLP on a quantum search task where no such symmetry exists.","The paper notes that the QAP specializes to the Traveling Salesman Problem and to graph matching, so a natural extension is to reuse the same grammar and tokens on those problems; success there without modification would strengthen the claim that this is a universal computational language.","The discovered cyclic step-size schedule echoes the cyclical learning-rate phenomenon in deep learning that the paper cites, which suggests CLP's token search could be pointed at optimizer and learning-rate schedule discovery for neural network training, with validation loss playing the role of the QAP objective."],"forward_implications":["Practitioners facing a strongly NP-hard problem could generate instance-adapted solvers on demand instead of hand-tuning heuristics; the paper reports quadratic scaling and roughly a 10,000-fold reduction in computational effort (in FLOPS) for size-100 QAP instances relative to tested baselines.","Any application that uses Grover's algorithm as a subroutine, including brute-force-resistance analyses in cryptography, would inherit a circuit with depth reduced by $(\\pi/2)\\sqrt{N}-1$ layers built from only H, X, and oracle gates, which the paper argues is easier to realize on hardware and exponentially more error-tolerant since noise scales with depth.","The automatic rediscovery of Frank–Wolfe and k-OPT from eight low-level primitives, together with the discovered cyclic step-size schedule that beats the scipy line-search baseline, suggests that composing and tuning known algorithmic building blocks is itself automatable and not limited to the QAP.","Because CLP operates at the computational level rather than the code level, its discovered token chains are platform-independent descriptions that downstream code generators could compile into running programs; the paper's conclusion points toward automated engineering-system design and agent-based real-time algorithm development."],"supporting_citations":[{"why":"Gage's data-compression Byte-Pair Encoding; A-BPE adapts its pair-merging routine to grow the computational vocabulary from primitives into composite tokens.","marker":"(12)"},{"why":"The AlphaZero/MuZero self-play MCTS plus neural-network paradigm that the paper's ensemble variant adapts as its token-chaining search engine.","marker":"(14–16)"},{"why":"Burkard, Karisch, and Rendl's QAPLIB library and the framing of the QAP as one of the hardest NP-hard problems; supplies the benchmark instances and difficulty baseline for the QAP experiments.","marker":"(19)"},{"why":"Kuhn's Hungarian method; supplies the linear-sum-assignment primitive that maps continuous matrices to permutations and appears throughout the discovered algorithms.","marker":"(21)"},{"why":"Drugan's composite QAP generator; supplies the CQAP benchmark instances with known optima used to train and evaluate CLP.","marker":"(23)"},{"why":"Palubeckis' QAP generator; supplies the harder PQAP benchmark instances with known optima used to train and evaluate CLP.","marker":"(24)"},{"why":"Vogelstein and coauthors' fast approximate QAP implementation behind scipy; its step-size schedule is the baseline that CLP's discovered cyclic schedule outperforms.","marker":"(25)"},{"why":"Grover's quantum search algorithm; the baseline whose circuit depth and gate count the discovered circuit reduces by $(\\pi/2)\\sqrt{N}-1$ layers.","marker":"(26)"},{"why":"Farhi, Goldstone, and Gutmann's QAOA; defines the hybrid quantum-classical algorithm whose mixer circuits CLP designs.","marker":"(27)"},{"why":"Zhu and coauthors' ADAPT-QAOA; the adaptive variant, as implemented in CUDA-Q, that CLP's mixer circuits beat by an average 34.62% on MAX-CUT instances.","marker":"(28)"}],"fun_headline_variants":["Token-chain search beats Gurobi on 389/390 instances","Token chaining rediscovers Grover and improves QAOA by 34.6%","Algorithm sentences: token chains outperform classic heuristics","Computational language processing writes winning algorithms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-chosen primitive vocabulary and the grammar governing valid chains are expressive enough to contain a genuinely better algorithm, and that ensemble MCTS with a lightweight transformer can find it within the compute budget — the paper itself notes this trade-off when it restricts special primitives to act on a single primitive at a time, which leaves algorithms like Frank–Wolfe unexpressible until Byte-Pair expansion adds new tokens.","fun_headline_variants_meta":{"raw":{"variants":["Token-chain search beats Gurobi on 389/390 instances","Token chaining rediscovers Grover and improves QAOA by 34.6%","Algorithm sentences: token chains outperform classic heuristics","Computational language processing writes winning algorithms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00069,"raw_usage":{"total_tokens":3197,"prompt_tokens":1091,"completion_tokens":2106,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":2036}},"tokens_in":707,"tokens_out":2106,"duration_ms":19515,"temperature":1.0,"reasoning_tokens":2036,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:16:25.137554+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CLP on a held-out family of quadratic assignment instances with known optima that comes from neither the CQAP nor the PQAP generator, for example instances built from uniformly random flow and distance matrices, with Gurobi given the same wall-clock budget, and then inspect the learned token chains: if the optimality-gap advantage over Gurobi disappears on this distribution, or if removing the A-BPE-created composite tokens leaves performance unchanged, the claim that the framework generates genuinely new and generalizable algorithms rather than tuned recombinations of the supplied primitives would be settled against it.","supporting_citations":[],"review_version":1}