{"id":"bab0c0ce-2b74-441c-90f4-22e15f0fa288","arxiv_id":"1906.08809","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Deep RL agent trained on generated global routing instances outperforms sequential A* search.","lead":"This paper proposes a deep reinforcement learning method to solve global routing by training an agent in a simulated environment to learn routing policies. A smart generalist might read it to see how RL techniques could improve automated design in electronics and similar path-planning domains.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Evaluation confined to synthetic generator instances; no results reported on standard ISPD global routing benchmarks","rationale":"The reader's weakest assumption (simulation fidelity and generalization beyond generated instances) is precisely the point at which the empirical claim is least secure; the absence of standard-benchmark results makes that assumption untested. The proposed check directly falsifies or supports the transfer claim without requiring changes to the method itself.","tokens_in":1756,"tokens_out":322,"duration_ms":12156,"concrete_test":"Apply the released or re-trained policy to at least three ISPD 2007/2008 global routing benchmark instances; compute total wirelength, via count, and overflow against both the paper's sequential A* implementation and the best published results for those instances. If the DRL policy does not improve on A* or falls behind published routers, the headline outperformance claim does not generalize.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that the fine-tuned DRL policy outperforms sequential A* on routing solutions and rewards. This claim is load-bearing on the assumption that performance on the authors' parameterized generator transfers to realistic global routing instances. The manuscript describes a custom problem-set generator and presents results exclusively on instances drawn from it; no experiments appear on established contest benchmarks (ISPD 2007/2008 global routing suites) that encode real netlists, layer counts, and design-rule constraints. Consequently, any observed advantage could be an artifact of the generator's distribution rather than a general solution to the routing problem.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript presents a deep reinforcement learning (DRL) method for global routing in circuit design. It introduces a custom parameterized problem-set generator and reports that a fine-tuned DRL policy produces better routing solutions and rewards than a sequential A* baseline when evaluated on instances drawn from the generator.","tokens_in":1851,"tokens_out":424,"duration_ms":16969,"significance":"If the empirical claims hold under more rigorous evaluation, the work would demonstrate a data-driven alternative to heuristic routing methods and supply a generator useful for training and benchmarking future routing algorithms. The conjoint optimization aspect of DRL is presented as a conceptual contribution.","major_comments":[{"comment":"Results section: All reported comparisons are performed exclusively on synthetic instances from the authors' custom generator; no experiments appear on standard ISPD 2007/2008 global routing benchmarks containing realistic netlists, layer counts, and design-rule constraints. This directly undermines the load-bearing claim that the DRL approach outperforms A* for global routing.","section":"Results"},{"comment":"Abstract and Results section: The outperformance claim is stated without quantitative metrics (e.g., wirelength, via counts, overflow), statistical significance tests, problem-size ranges, or ablation studies on network architecture or reward components.","section":"Abstract and Results"}],"minor_comments":[{"comment":"The description of the state representation and action space in the DRL formulation could be made more precise with explicit equations or pseudocode.","section":"Method"},{"comment":"Figure captions and axis labels in the results plots lack sufficient detail on the exact metrics plotted and the number of instances averaged.","section":"Results"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an early arXiv preprint whose central empirical claim rests on an unvalidated generator; the absence of standard benchmarks is a scope issue that would need to be addressed before consideration at a venue focused on electronic design automation."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make.","responses":[{"response":"We agree that all quantitative comparisons are performed on instances drawn from the custom generator. This choice was deliberate to enable controlled variation of problem parameters (grid size, net count, layer constraints) and to isolate the effect of conjoint optimization in DRL. The generator is explicitly presented as a contribution for creating training and evaluation data. We acknowledge that the absence of results on ISPD 2007/2008 benchmarks limits direct claims about performance on industrial netlists with full design-rule sets. In the revised manuscript we will add a limitations paragraph that (a) states this scope explicitly and (b) describes the additional engineering steps (multi-layer encoding, via-cost modeling, overflow handling) needed to apply the same policy network to those benchmarks. We will not claim general superiority beyond the generated distribution.","revision_made":"partial","referee_comment":"[Results] Results section: All reported comparisons are performed exclusively on synthetic instances from the authors' custom generator; no experiments appear on standard ISPD 2007/2008 global routing benchmarks containing realistic netlists, layer counts, and design-rule constraints. This directly undermines the load-bearing claim that the DRL approach outperforms A* for global routing."},{"response":"We will revise the abstract and the results section to report concrete metrics: average reward, total wirelength, via count, and overflow where measured; the range of grid sizes and net counts used (e.g., 8×8 to 32×32 grids with 10–100 nets); and the number of evaluation instances. We will also add error bars across multiple random seeds and note any statistical tests performed. In addition, we will include ablation results comparing network architectures (MLP vs. CNN variants) and reward-component weightings. These quantitative details and ablations will be inserted into the revised manuscript.","revision_made":"yes","referee_comment":"[Abstract and Results] Abstract and Results section: The outperformance claim is stated without quantitative metrics (e.g., wirelength, via counts, overflow), statistical significance tests, problem-size ranges, or ablation studies on network architecture or reward components."}],"tokens_in":1308,"tokens_out":492,"duration_ms":31926,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The punchline is that this paper introduces a custom problem generator for global routing and shows a DRL policy outperforming sequential A* on instances drawn from it. The generator is the clearest addition: it can produce problems with varying sizes and constraints, which is useful for creating training data and testing different algorithms. They also walk through the RL setup, try different network structures, and analyze the learned policies and rewards on their instances. That part is straightforward and gives a concrete example of applying existing deep RL techniques to a routing task. The main soft spot is the evaluation scope. All reported results stay inside the synthetic generator; there are no experiments on the standard ISPD 2007 or 2008 global routing benchmarks that reflect real netlists and design rules. This leaves open whether the observed advantage is tied to the generator's distribution rather than a general improvement. The comparison is only against sequential A*, with no ablations or statistical details visible in the abstract, though the full paper presumably supplies more. The work is honest about its simulation setting and does not overclaim transfer. Readers working on RL for path planning or EDA tools could pick up the generator idea or the training approach. It is not a foundational advance but supplies a usable data-generation tool and an empirical data point. I would send it to peer review so referees can check the full experimental protocol and push for ISPD results if the authors want broader claims.","headline":"They built a parameterized generator for global routing problems and trained DRL that beats sequential A* on those synthetic cases, but the evaluation never leaves the generator.","tokens_in":2336,"tokens_out":359,"would_cite":false,"duration_ms":11568,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"DRL routing on synthetic grids; unrelated to recognition-cost forcing chain","alignment":"orthogonal","rationale":"Paper applies standard DQN (state vector of coordinates+capacities, reward +100 on target, sequential two-pin decomposition via MST) to a custom generator of 8x8x2/16x16x2 grids; results compared only to A* on those instances. No J-cost, phi-ladder, 8-tick periodicity, ratio-symmetric functional equation, or distinction-to-spacetime derivation appears. Domain is applied RL for path planning; RS has no opinion on it.","tokens_in":54224,"confidence":"high","tokens_out":145,"duration_ms":5173,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A deep reinforcement learning agent can outperform sequential A* in finding global routing solutions for circuit designs.","keywords":["global routing","deep reinforcement learning","circuit design","path planning","A* algorithm","problem generation","electronic design automation"],"falsifier":"Evaluating the trained routing agent on actual global routing instances from commercial electronic design automation software and comparing total wire length, via count, and design rule violations against A* results.","tokens_in":2662,"feed_emoji":"🔌","tokens_out":406,"duration_ms":20065,"temperature":0.7,"pith_summary":"This paper applies deep reinforcement learning to the global routing problem, where wires must connect circuit components without violating design rules. The method trains an agent in a simulated environment to learn a routing policy that works across many different problem instances through conjoint optimization. It also introduces a generator that creates routing problems with adjustable sizes and constraints for training and testing. If the learned policy generalizes, it offers a flexible alternative to traditional greedy algorithms and heuristics that often yield suboptimal results. Experiments show the fine-tuned model achieves higher rewards than a sequential A* benchmark on the generated problems.","feed_headline":"Reinforcement learning beats A* for global routing","feed_subtitle":"A trained agent finds higher-reward wire paths than sequential search on varied simulated circuit problems.","key_machinery":"The deep reinforcement learning agent that uses a conjoint optimization mechanism to learn a policy for connecting circuit components with wires in a simulated environment.","core_discovery":"The deep reinforcement learning approach enables an agent to produce an optimal routing policy based on the variety of problems presented, and the fine-tuned model outperforms the sequential A* method in routing solutions and rewards within the simulated global routing environment.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Deep RL outperforms A* in global routing simulation","DRL agent outperforms sequential A* on routing tasks","Reinforcement learning policy outperforms A* in circuits","DRL model finds better routing solutions than A*"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The simulated environment and reward structure must accurately capture the constraints and objectives of real global routing problems so that the policy generalizes beyond the training instances.","fun_headline_variants_meta":{"raw":{"variants":["Deep RL outperforms A* in global routing simulation","DRL agent outperforms sequential A* on routing tasks","Reinforcement learning policy outperforms A* in circuits","DRL model finds better routing solutions than A*"]},"model":"grok-4.3","cost_usd":0.007985,"raw_usage":{"total_tokens":3637,"prompt_tokens":671,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":79849500,"prompt_tokens_details":{"text_tokens":671,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2907,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":671,"tokens_out":59,"duration_ms":22577,"temperature":1.0,"reasoning_tokens":2907,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-25T19:28:34.445860+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Evaluating the trained routing agent on actual global routing instances from commercial electronic design automation software and comparing total wire length, via count, and design rule violations against A* results.","supporting_citations":[],"review_version":1}