Pith. sign in

REVIEW 4 major objections 5 minor 43 references

An LLM agent that escalates reasoning effort only when optimization stalls reaches deeper PPA improvements at equal dollar cost, and does so without relying on how its cross-design memory is written.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-07-31 23:43 UTC pith:SIA3CQBH

load-bearing objection A solid, honest empirical paper that makes cost-normalized comparison the standard and shows adaptive effort helps on a small test set; the proxy fit is unvalidated but the held-out and paired-branch results carry the claim. the 4 major comments →

arxiv 2607.27879 v1 pith:SIA3CQBH submitted 2026-07-30 cs.AR cs.AI

ARES: Adaptive Reasoning-Effort Steering for PPA- and Cost-Aware RTL Optimization with LLM Agents

Stef Cuyckens , Mihaela Jivanescu , Jun Yin , Chao Fang , Marian Verhelst This is my paper
classification cs.AR cs.AI
keywords LLM agentsRTL optimizationadaptive reasoning effortPPAinference costagent memorypatience counterelectronic design automation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that for LLM agents optimizing RTL designs, the quality-per-dollar is set by where reasoning effort is spent, not by how the long-term memory is constructed. It introduces a normalized dollar-cost-per-call metric, shows that an engineered memory brings no dependable gain over a plain concatenation of the same experience, and proposes a patience-counter policy that escalates from medium to high reasoning only when cheaper effort stalls. On three held-out designs, this policy lowers the figure of merit by 23–27% where the best fixed-effort setting reaches 16–23%, at equal normalized cost. If correct, adaptive per-call reasoning is a cheap, high-leverage improvement over fixed-effort agent design.

Core claim

Ares claims that the quality and cost of an LLM-based RTL optimizer are set by where its reasoning is spent, not by how its long-term memory is written. Concretely, a two-state policy that starts every design at medium effort, escalates to high only after a patience counter accumulates stalled or failed iterations, and discharges the counter on accepted improvements, reaches a 23–27% figure-of-merit reduction on three held-out designs while the best fixed-effort setting reaches 16–23%, at equal normalized cost. The paper also reports that an engineered memory—structured rules, deduplication, ranking—yields no dependable gain over a plain concatenation of the same experience, and that reporti

What carries the argument

The patience counter C is the central mechanism: stalled iterations add 1, failed iterations add w=2.8, and accepted improvements discharge C by max(0, 1−ΔFoM/κ) with κ=0.05. When C reaches the threshold p=3, the next call runs at high reasoning effort and the counter resets. The constants are fit once on 21 training runs against a proxy label for 'warranted' escalation. Cost is measured in 'high-calls,' the cumulative dollar spend divided by the design's mean cost of one high-effort call, which keeps comparisons robust to price changes.

Load-bearing premise

The policy's escalation constants are fit on 21 training designs using a proxy for 'warranted' escalation—no accepted improvement within the next three iterations—and the headline 23–27% advantage is measured on only three held-out designs, so the claim transfers only if that proxy, calibrated on the authors' training runs with a specific LLM, marks where deeper reasoning pays on unseen designs.

What would settle it

Run the adaptive policy with the same fitted constants on a larger held-out set (e.g., ten or more RTL designs from varied sources) and compare final FoM at matched cumulative cost against fixed medium and fixed high effort; if the adaptive arm does not consistently beat both fixed arms on every design, the central claim fails. Additionally, re-fit the constants with a different commercial LLM's reasoning-effort levels and check whether the advantage persists, which would test whether the proxy is model-specific.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Reporting normalized dollar cost alongside figure of merit becomes the standard for comparing RTL optimizers; FoM alone can mislead.
  • No fixed reasoning-effort level fits a whole run: adaptive escalation beats every fixed level on all three test designs at equal cost.
  • Engineering the structure of long-term memory—deduplication, ranking, abstraction—adds little value over plain experience at equal spend; effort allocation matters more.
  • The adaptive policy cuts run-to-run variance by 58%, making single runs more dependable rather than requiring repeated runs or population sampling.
  • The same optimizer closes up to 83% of the gap from an LLM-drafted multiply-accumulate unit to a hand-optimized counterpart.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the patience-counter policy is generic—any iterative LLM agent with a measurable progress signal (program repair, test generation, constraint solving) could adopt escalation-only-on-stall; the paper does not claim this transfer.
  • Editorial inference: the fitted constants (p=3, w=2.8, κ=0.05) are likely model- and domain-specific; retuning for other LLMs or design families may be needed, and an online adaptive scheme could remove the need for a training set.
  • Editorial inference: if memory construction truly does not matter, prior agent designs that invest heavily in skill libraries could simplify to plain experience logs and reallocate engineering effort to effort scheduling.
  • Editorial inference: the cost metric could be extended beyond LLM call cost to include synthesis tool runtime or human review time, which would broaden the fairness of comparisons.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces Ares, an LLM-agent framework for RTL PPA optimization. Its three contributions are: (1) a normalized per-call dollar cost metric that lets optimization quality be compared at equal spend; (2) an empirical comparison showing that an engineered cross-design long-term memory has no dependable advantage over a plain concatenation of the same experience; and (3) an adaptive reasoning-effort policy that escalates from medium to high effort only after a patience counter, fit on 21 training designs, signals that progress has stalled. The main reported result is that on three held-out test designs this adaptive policy reduces the normalized FoM by 23–27%, versus 16–23% for the best fixed effort level, at equal normalized cost. The paper also reports closing up to 83% of the gap between an LLM-drafted MX MAC and a hand-optimized version, and a deeper FoM than Dr. RTL at 12% of its tokens on the controller design.

Significance. If the central claim holds, the paper makes a useful contribution: it provides a cost-aware evaluation methodology for LLM-based RTL optimizers and presents evidence that adaptive reasoning-effort steering, rather than long-term memory construction, is the main driver of PPA-per-dollar. The paper is commendable for reporting a train/test split for the fitted escalation constants, for measuring costs in a normalized dollar unit, and for including a paired branch experiment on the MX MAC that directly compares fixed-medium and adaptive continuations from the same run. These design choices reduce, but do not eliminate, concerns about overfitting and stochasticity. The main significance depends on whether the escalation policy's proxy — 'stall predicts where deeper reasoning pays' — transfers to unseen designs and to other LLMs; the current evidence for that transfer is limited to three test designs with a small number of repetitions.

major comments (4)
  1. [§5.2] The escalation constants p, w, κ are fitted to maximize agreement with a proxy label: an iteration is 'warranted' if medium effort produces no accepted improvement within the next three iterations. This is a stall-prediction objective, not a direct measure of whether high effort would have produced a better FoM per dollar at that point. The paper never validates on the training runs that escalating at warranted points actually improves the optimization trajectory relative to staying at medium. Thus the fitted policy is, at best, a calibrated stall detector; the causal link to FoM-per-dollar is only tested on the three held-out designs. This is the load-bearing assumption of the paper and should be tested directly, e.g. by comparing, on a sample of warranted and non-warranted points, the outcome of a high-effort continuation versus a medium-effort continuation.
  2. [§3.2, §5.2] The headline result rests on three test designs (tv80, uart, controller), and the number of repetitions per design is not stated explicitly; Figures 2 and 4 show three faint runs per curve, implying nine test runs in total. With nine runs, the 23–27% versus 16–23% difference could plausibly arise from design selection or LLM stochasticity. The MX MAC branch experiment (§5.3) adds six paired runs, but on a single design not part of the held-out set. The authors should report exact run counts, per-design and per-arm results with confidence intervals, and ideally add more held-out designs or a bootstrap analysis to quantify the stability of the claimed advantage.
  3. [§3.2, §5.2] The conclusion that 'long-term memory construction matters little' is supported only by three test designs, three runs each, and the paper reports no statistical test. The two memory curves in Figure 2 are close, but the paper interprets a null result as evidence for the stronger claim that 'the quality and cost of such an optimizer are set not by how its long-term memory is written.' This is an overstatement given the small sample. The memory comparison should either be framed as 'no dependable gain was measured on these designs' or be backed by more designs and a power analysis.
  4. [§5.4] The escalation policy is only ever compared against fixed low, medium, and high effort. It is not compared against other adaptive baselines, such as a policy that escalates after a fixed number of stalls without a fitted patience-counter discharge, or a policy that alternates effort levels. The fitted constants (p=3, w=2.8, κ=0.05) are reported to catch 94% of warranted points on training runs, but the paper does not report how much of the held-out gain is due to the specific counter dynamics versus the simple principle of 'escalate after a stall.' Adding an ablation with a simpler heuristic would clarify whether the fitted parameters are actually necessary for the reported improvement.
minor comments (5)
  1. [§5.1] The paper says 'each configuration is repeated multiple times' without giving exact numbers. Please state the number of runs for every experiment, including the SotA comparison in §5.4, which currently reports single trajectories.
  2. [§5.2] The reported mean FoMs (0.76, 0.77, 0.73 for adaptive versus 0.79, 0.84, 0.79 for fixed high) would benefit from error bars or per-run endpoints. The current figures show only faint individual runs that are hard to read in grayscale.
  3. [§5.3] The paired-branch experiment on the MX MAC is the strongest evidence for the policy, but the caption and text do not specify how the 'first stall point' is determined and whether the medium continuation is a true control (same run history) or a separate run. Please clarify the branching protocol.
  4. [§3.1] The definition of 'high-calls' as 'the dollars spent divided by the design's mean computed cost of one high-effort call' uses a mean over presumably several high-effort calls. If the mean is computed from the adaptive runs themselves, the normalization could be affected by the policy. Please specify how the mean is computed and whether it is fixed per design before running the adaptive policy.
  5. [References] Several references are to arXiv preprints dated 2026; if these are not yet publicly available, the authors should confirm the citations are accessible. Also, the OpenRouter pricing links ([21], [26]) should be cited with accessed dates.

Circularity Check

0 steps flagged

No significant circularity: the adaptive policy is fit on training runs and evaluated on held-out designs; the stall-based proxy is a validity concern, not a definitional reduction.

full rationale

The central adaptive-effort claim is not circular. The escalation constants p=3, w=2.8, kappa=0.05 are fit once on 21 training designs to a stall-based 'warranted' proxy, and the policy is then run on three held-out test designs where the 23-27% FoM reduction is measured against fixed-effort arms. This is a genuine train/test split: the fitted constants do not by construction determine the held-out FoM. The 'warranted' label ('an escalation is warranted at an iteration where medium produces no accepted improvement within the next three iterations') is a proxy for high-effort payoff, but the paper does not present that proxy as the derivation of the headline result; it reports the headline as an out-of-sample empirical measurement. The memory-construction finding is an ablation at equal spend with all other factors held fixed, which is an empirical negative result rather than a self-definition. The only author-overlapping citation is [7], the published MX MAC design used as a benchmark target in Sec. 5.3; that is an external, falsifiable artifact and is not used to justify the adaptive-policy mechanism or the main held-out comparison. The principal risks are external validity: the proxy is calibrated on the authors' own Opus-4.6 training traces, the test set is only three designs with three runs each, and the 'warranted' label is never validated on training data against what high effort would actually achieve. Those are correctness and reproducibility concerns, not circularity, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claim rests on three fitted escalation constants (p, w, κ), the assumed transfer of a heuristic stall proxy from training to test designs, the two-state medium-to-high policy form, and the definitions of FoM (Eq. 1) and LLM-only cost (Eq. 2). No new physical or conceptual entities are postulated; the patience counter is a policy mechanism accounted for under free parameters, and it has an evaluation handle inside the paper (held-out designs) rather than an external falsifiable prediction.

free parameters (4)
  • p (patience threshold) = 3
    Counter value at which effort escalates; grid-searched on 21 training runs under the 'warranted escalation' label (§3.3).
  • w (failure weight) = 2.8
    Weight added to the counter on failed iterations; fitted jointly with p and κ (§3.3).
  • κ (discharge scale) = 0.05
    Relative FoM gain that fully clears the patience counter; fitted jointly (§3.3).
  • testbench length = 10^4 random vectors, fixed seed
    Chosen verification depth (§4); longer testbenches cost more synthesis time and could change which candidates pass, but it is not tuned to the result.
axioms (6)
  • domain assumption Post-synthesis PPA on the Nangate 45nm library with Synopsys DC/PrimeTime is a faithful measure of design quality for fair cross-optimizer comparison
    All FoM claims rest on the commercial flow (§4); the paper motivates commercial over open-source tools but cannot validate that this flow is bias-free toward any optimizer.
  • domain assumption On Claude Opus 4.6/Claude Code, a 'high' effort call produces better edits than a 'medium' call when medium progress has stalled
    Central mechanism of §3.3; indirectly validated only by the 3-design held-out results and the MX MAC branch experiment.
  • ad hoc to paper The 'warranted escalation' label (no accepted improvement within the next three iterations, defined on the authors' own training traces) is a valid proxy for where deeper reasoning pays
    The fitted targets of §3.3 are defined internally from medium-only runs; its transfer to unseen designs is the paper's core empirical bet.
  • ad hoc to paper The patience-counter functional form (add 1 on stall, add w on failure, multiplicative discharge by ΔFoM/κ) adequately captures staleness dynamics
    Constants are fitted, but the functional form is a modeling choice that is not derived or ablated (§3.3).
  • domain assumption LLM call cost (OpenRouter prices, Eq. 2) is the relevant optimization cost; synthesis and verification compute are excluded from the cost axis
    All cost comparisons use only LLM token dollars (§3.1); EDA tool effort is not priced, so the 'equal cost' comparisons are partial.
  • domain assumption FoM = (area/area0)(power/power0)(delay/delay0) is the correct, comparable objective
    Eq. 1 defines the objective; inherited from prior RTL optimization work, and the Dr. RTL comparison is made under this power-aware FoM even though Dr. RTL does not optimize power.

pith-pipeline@v1.3.0-daily-deepseek · 12201 in / 21339 out tokens · 198921 ms · 2026-07-31T23:43:42.498080+00:00 · methodology

0 comments
read the original abstract

Large language model (LLM) agents optimize the power, performance, and area (PPA) of register-transfer-level (RTL) designs by iterating over edits, synthesis, and PPA analysis, paying a dollar cost for every LLM call. Prior agents report the quality reached without its normalized cost, attribute that quality to an engineered cross-design memory, and hold the reasoning effort of every call fixed. We propose Ares with three corresponding innovations. (1) We introduce a normalized dollar cost per LLM call reported alongside the figure of merit (FoM), enabling fair comparison across effort levels and optimizers. (2) Using this accounting, we find the construction of the long-term memory matters little. An engineered memory brings no dependable gain over a plain concatenation of the same experience. (3) We instead adapt the per-call reasoning effort by escalating to deeper reasoning only once progress at a lower effort stalls, via a patience counter fit on 21 training designs, allocating reasoning where it pays rather than uniformly across all iterations. On three test designs unseen during training, the effort policy lowers the FoM by 23-27% where the best fixed effort reaches 16-23%, at equal normalized cost. Ares closes up to 83% of the gap from an LLM-drafted multiply-accumulate unit to its highly hand-optimized counterpart, and reaches a 25% deeper FoM than state-of-the-art Dr. RTL at 12% of its tokens.

Figures

Figures reproduced from arXiv: 2607.27879 by Chao Fang, Jun Yin, Marian Verhelst, Mihaela Jivanescu, Stef Cuyckens.

Figure 1
Figure 1. Figure 1: Overview of Ares. The LLM agent edits the running-best RTL; each candidate is verified against the input design 𝑣0, synthesized on the commercial flow, and kept only if it lowers the FoM. Our contributions: 1 the per-call cost metric enabling proper comparison, 2 the finding that the construction of the cross-design long-term memory matters little, and 3 an adaptive reasoning-effort policy that escalates o… view at source ↗
Figure 3
Figure 3. Figure 3: The adaptive-effort policy, with constants (𝑝 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: The three memory constructions of Sec. 3.2 on the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The three fixed effort levels and the adaptive policy [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: An LLM-created MX MAC optimized toward a [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ares with adaptive effort, REvolution, and Dr. RTL optimizing controller from the same 𝑣0 on the same flow, 25 candidates each, dots marking the verified ones. Dr. RTL’s timing-focused selection returns the annotated 0.923 and discards its ringed best candidate at 0.909. the same effort. Dr. RTL optimizes timing and area without power, and under the power-aware FoM the power its rewrites add cancels much o… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 6 linked inside Pith

  1. [1]

    Rosenstein, and Sherief Reda

    Manar Abdelatty, Maryam Nouh, Jacob K. Rosenstein, and Sherief Reda. 2025. Pluto: A Benchmark for Evaluating Efficiency of LLM-generated Hardware Code. arXiv preprint arXiv:2510.14756(2025)

    arXiv 2025

  2. [2]

    Mohammad Ali Alomrani, Yingxue Zhang, Derek Li, Qianyi Sun, Soumyasundar Pal, Zhanguang Zhang, Yaochen Hu, Rohan Deepak Ajwani, Antonios Valkanas, Raika Karimi, Peng Cheng, Yunzhou Wang, Pengyi Liao, Hanrui Huang, Bin Wang, Jianye Hao, and Mark Coates. 2025. Reasoning on a Budget: A Survey of Adaptive and Controllable Test-Time Compute in LLMs.arXiv prepr...

    Pith/arXiv arXiv 2025

  3. [3]

    Anthropic. 2025. Claude Code. https://code.claude.com/docs/en/overview

    2025

  4. [4]

    Cadence Design Systems. 2025. JasperGold Formal Verification Platform. https://www.cadence.com/en_US/home/tools/system-design-and-verification/ formal-and-static-verification/jasper-verification-platform.html

    2025

  5. [5]

    Lingjiao Chen, Matei Zaharia, and James Zou. 2024. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance.Transactions on Machine Learning Research (TMLR)(2024)

    2024

  6. [6]

    Yuxi Chen, Junming Chen, Chenyu He, Yiwei Li, Yicheng Ji, Yifan Wu, Dingyu Yang, Lansong Diao, Lidan Shou, Hongliang Zhang, Huan Li, and Gang Chen

  7. [7]

    Stef Cuyckens, Xiaoling Yi, Nitish Satya Murthy, Chao Fang, and Marian Verhelst

  8. [8]

    Wenji Fang, Mengming Li, Min Li, Zhiyuan Yan, Shang Liu, Hongce Zhang, and Zhiyao Xie. 2025. AssertLLM: Generating and Evaluating Hardware Verification Assertions from Design Specifications via Multi-LLMs. InProceedings of the Asia and South Pacific Design Automation Conference (ASP-DAC)

    2025

  9. [9]

    Wenji Fang, Yao Lu, Shang Liu, Jing Wang, Ziyan Guo, Junxian He, Fengbin Tu, and Zhiyao Xie. 2026. Dr. RTL: Autonomous Agentic RTL Optimization through Tool-Grounded Self-Improvement.arXiv preprint arXiv:2604.14989(2026)

    Pith/arXiv arXiv 2026

  10. [10]

    Shubham Gandhi, Manasi Patwardhan, Lovekesh Vig, and Gautam Shroff. 2024. BudgetMLAgent: A Cost-Effective LLM Multi-Agent System for Automating Machine Learning Tasks. InProceedings of the 4th International Conference on AI-ML Systems (AIMLSystems)

    2024

  11. [11]

    Pengfei Gao and Chao Peng. 2026. More with Less: An Empirical Study of Turn- Control Strategies for Efficient Coding Agents. InProceedings of the IEEE/ACM International Conference on Software Engineering (ICSE)

    2026

  12. [12]

    Siyuan Guo, Cheng Deng, Ying Wen, Hechang Chen, Yi Chang, and Jun Wang

  13. [13]

    Chia-Tung Ho, Haoxing Ren, and Brucek Khailany. 2025. VerilogCoder: Au- tonomous Verilog Coding Agents with Graph-based Planning and Abstract Syntax Tree (AST)-based Waveform Tracing Tool. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI)

    2025

  14. [14]

    Bowen Jin, TJ Collins, Donghan Yu, Mert Cemri, Shenao Zhang, Mengyu Li, Jay Tang, Tian Qin, Zhiyang Xu, Jiarui Lu, Guoli Yin, Jiawei Han, and Zirui Wang

  15. [15]

    Yushu Li, Wenlong Deng, Jiajin Li, and Xiaoxiao Li. 2026. Spend Less, Rea- son Better: Budget-Aware Value Tree Search for LLM Agents.arXiv preprint arXiv:2603.12634(2026)

    arXiv 2026

  16. [16]

    Mingjie Liu, Teodor-Dumitru Ene, Robert Kirby, Chris Cheng, Nathaniel Pinck- ney, Rongjian Liang, Jonah Alben, Himyanshu Anand, Sanmitra Banerjee, Ismet Bayraktaroglu, Bonita Bhaskaran, Bryan Catanzaro, Arjun Chaudhuri, Sharon Clay, Bill Dally, Laura Dang, Parikshit Deshpande, Siddhanth Dhodhi, Sameer Halepete, Eric Hill, Jiashang Hu, Sumit Jain, Ankit Ji...

    Pith/arXiv arXiv 2023

  17. [17]

    Mingjie Liu, Nathaniel Pinckney, Brucek Khailany, and Haoxing Ren. 2023. Ver- ilogEval: Evaluating Large Language Models for Verilog Code Generation. In Proceedings of the IEEE/ACM International Conference on Computer-Aided Design (ICCAD)

    2023

  18. [18]

    Controlling Performance and Budget of a Centralized Multi-agent LLM System with Reinforcement Learning.arXiv preprint arXiv:2511.02755(2025)

    arXiv 2025

  19. [19]

    Yao Lu, Shang Liu, Qijun Zhang, and Zhiyao Xie. 2024. RTLLM: An Open- Source Benchmark for Design RTL Generation with Large Language Model. In Proceedings of the Asia and South Pacific Design Automation Conference (ASP- DAC)

    2024

  20. [20]

    Yao Lu, Shang Liu, Hangan Zhou, Wenji Fang, Qijun Zhang, and Zhiyao Xie. 2026. A New Benchmark for the Appropriate Evaluation of RTL Code Optimization. arXiv preprint arXiv:2601.01765(2026)

    arXiv 2026

  21. [21]

    Elias Lumer, Faheem Nizar, Akshaya Jangiti, Kevin Frank, Anmol Gulati, Man- dar Phadate, and Vamse Kumar Subbiah. 2026. Don’t Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks.arXiv preprint arXiv:2601.06007(2026)

    arXiv 2026

  22. [22]

    Shang Liu, Wenji Fang, Yao Lu, Qijun Zhang, Hongce Zhang, and Zhiyao Xie

  23. [23]

    In2024 IEEE LLM Aided Design Workshop (LAD)

    RTLCoder: Outperforming GPT-3.5 in Design RTL Generation with Our Open-Source Dataset and Lightweight Solution. In2024 IEEE LLM Aided Design Workshop (LAD). 1–5

  24. [24]

    OpenCores. 2026. OpenCores: Open Source IP Cores. https://opencores.org

    2026

  25. [25]

    OpenRouter. 2026. OpenRouter Documentation: Reasoning Tokens. https:// openrouter.ai/docs/use-cases/reasoning-tokens

    2026

  26. [26]

    OpenRouter. 2026. OpenRouter Models: Model and Pricing Index. https:// openrouter.ai/models

    2026

  27. [27]

    Kyungjun Min, Kyumin Cho, Junhwan Jang, and Seokhyeong Kang. 2026. REvolu- tion: An Evolutionary Framework for RTL Generation Driven by Large Language Models. InProceedings of the Asia and South Pacific Design Automation Conference (ASP-DAC)

    2026

  28. [28]

    Nangate Inc. 2011. NanGate FreePDK45 Open Cell Library. https://si2.org/ open-cell-library/

    2011

  29. [29]

    Vaishnavi Pulavarthi, Deeksha Nandal, Soham Dan, and Debjit Pal. 2025. As- sertionBench: A Benchmark to Evaluate Large-Language Models for Assertion Generation. InFindings of the Association for Computational Linguistics: NAACL

    2025

  30. [30]

    Zhengyuan Shi, Jingxin Wang, Tairan Cheng, Changran Xu, Weikang Qian, and Qiang Xu. 2026. CktEvo: Repository-Level RTL Code Benchmark for Design Evolution.arXiv preprint arXiv:2603.08718(2026)

    arXiv 2026

  31. [31]

    Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling LLM Test-Time Compute Optimally Can be More Effective than Scaling Parameters for Reasoning. InThe Thirteenth International Conference on Learning Representations (ICLR)

    2025

  32. [32]

    Jingyu Pan, Guanglei Zhou, Chen-Chia Chang, Isaac Jacobson, Jiang Hu, and Yiran Chen. 2025. A Survey of Research in Large Language Models for Electronic Design Automation.ACM Transactions on Design Automation of Electronic Systems (TODAES)30, 3 (2025)

    2025

  33. [33]

    Heng Ping, Peiyu Zhang, Zhenkun Wang, Shixuan Li, Anzhe Cheng, Wei Yang, Paul Bogdan, and Shahin Nazarian. 2026. POET: Power-Oriented Evolutionary Tuning for LLM-Based RTL PPA Optimization.arXiv preprint arXiv:2603.19333 (2026)

    arXiv 2026

  34. [34]

    Xufeng Yao, Yiwen Wang, Xing Li, Yingzhao Lian, Ran Chen, Mingxuan Yuan, Hong Xu, and Bei Yu. 2024. RTLRewriter: Methodologies for Large Models aided RTL Code Optimization. InProceedings of the IEEE/ACM International Conference on Computer-Aided Design (ICCAD)

    2024

  35. [35]

    Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, and Wenya Wang. 2026. MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents.arXiv preprint arXiv:2602.02474(2026)

    Pith/arXiv arXiv 2026

  36. [36]

    Zixi Zhang, Balint Szekely, Pedro Gimenes, Greg Chadwick, Hugo McNally, Jianyi Cheng, Robert Mullins, and Yiren Zhao. 2025. LLM4DV: Using Large Language Models for Hardware Test Stimuli Generation. In2025 IEEE 33rd Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM). 133–137

    2025

  37. [37]

    Yun-Da Tsai, Mingjie Liu, and Haoxing Ren. 2024. RTLFixer: Automatically Fixing RTL Syntax Errors with Large Language Models. InProceedings of the ACM/IEEE Design Automation Conference (DAC)

    2024

  38. [38]

    Yaoxiang Wang, Qi Shi, ShangZhan Li, Qingguo Hu, Xinyu Yin, Bo Guo, Xu Han, Maosong Sun, and Jinsong Su. 2026. VeriAgent: A Tool-Integrated Multi-Agent System with Evolving Memory for PPA-Aware RTL Code Generation.arXiv preprint arXiv:2603.17613(2026)

    arXiv 2026

  39. [42]

    Huichi Zhou, Yihang Chen, Siyuan Guo, Xue Yan, Kin Hei Lee, Zihan Wang, Ka Yiu Lee, Guchun Zhang, Kun Shao, Linyi Yang, and Jun Wang. 2025. Me- mento: Fine-tuning LLM Agents without Fine-tuning LLMs.arXiv preprint arXiv:2508.16153(2025)

    Pith/arXiv arXiv 2025

  40. [43]

    Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiangbin Liu, Xinlei Yu, Jianmin Zhou, Na Wang, Chunyang Sun, and Jun Wang. 2026. Memento-Skills: Let Agents Design Agents.arXiv preprint arXiv:2603.18743 (2026). 7

    arXiv 2026

  41. [2024]

    InProceedings of the International Conference on Machine Learning (ICML)

    DS-Agent: Automated Data Science by Empowering Large Language Models with Case-Based Reasoning. InProceedings of the International Conference on Machine Learning (ICML)

  42. [2025]

    InProceedings of the IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED)

    Efficient Precision-Scalable Hardware for Microscaling (MX) Processing in Robotics Learning. InProceedings of the IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED)

  43. [2026]

    Token Economics for LLM Agents: A Dual-View Study from Computing and Economics.arXiv preprint arXiv:2605.09104(2026)

    Pith/arXiv arXiv 2026