Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Compressing coding-agent observations so the next action stays the same cuts token use by about a third without losing task success.

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 · grok-4.5

2026-07-12 06:10 UTC pith:POVAN4I2

load-bearing objection Solid systems paper: NAP-guided observation compression delivers ~33% token savings with competitive PASS@1; the local-to-global proxy is soft but the empirics hold under the stated setup. the 3 major comments →

arxiv 2607.02911 v1 pith:POVAN4I2 submitted 2026-07-03 cs.SE cs.AI

CoACT: Action-Preserving Observation Compression for Coding Agents

Haorui Chen , Yuancheng Zhu , Yitong Zhang , Jia Li This is my paper
classification cs.SE cs.AI
keywords coding agentsobservation compressionnext-action preservationLLM inference costSWE-benchcontext managementefficiency-effectiveness trade-off
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.

Coding agents burn most of their inference budget on long environment observations that pile into the context after every tool call. Existing compressors either throw away information the agent still needs or leave too much redundant text, so the efficiency–effectiveness trade-off stays poor. This paper argues that a practical stand-in for final task success is next-action preservation: a compressed observation is acceptable only if the agent would still issue the same next command. CoACT uses that rule as a filter when a teacher produces many candidate compressions, then keeps the shortest survivors as training targets for a small compressor. On SWE-bench Verified with three agent models, the trained compressor cuts average total tokens by roughly one-third while PASS@1 stays close to—and on two models slightly above—the uncompressed baseline, and it combines cleanly with trajectory-level compressors that rewrite history.

Core claim

Observation compression for coding agents can be cast as a constrained optimization problem whose effectiveness constraint is next-action preservation rather than end-of-trajectory PASS@1. Training a lightweight compressor with an action-preservation reward (keep only candidates that induce the same next action) plus a length-reduction reward yields about 33% lower total token consumption on SWE-bench Verified while keeping task success comparable to the uncompressed agent, outperforming prior observation compressors on the efficiency–effectiveness trade-off.

What carries the argument

Next-action preservation (NAP): a compressed observation is accepted only when the agent’s immediate next action matches the action taken under the raw observation; NAP supplies a cheap, step-level proxy for the PASS@1 constraint and drives the action-preservation reward that filters teacher candidates before length-based selection.

Load-bearing premise

If every compression keeps the agent’s next action the same, the whole action sequence—and therefore final task success—will stay the same, even when information not needed for the immediate next step is still required later.

What would settle it

Run the same agent on SWE-bench with CoACT but deliberately keep compressions that pass NAP yet drop later-needed facts (for example long error traces used only after several steps); if PASS@1 then falls while next-action match rates stay high, NAP is not a sufficient proxy.

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

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

3 major / 5 minor

Summary. The paper proposes CoACT, an observation-compression method for LLM coding agents that trains a lightweight compressor under a next-action preservation (NAP) constraint. NAP requires that a compressed observation induce the same immediate next action as the raw observation; the authors treat this as a practical proxy for the end-of-trajectory PASS@1 constraint in a constrained efficiency–effectiveness optimization (Eqs. 1–2). Supervision is built by filtering teacher-generated compression candidates with an action-preservation reward and a length-reduction reward, then training via offline bootstrap and online alignment. On SWE-bench Verified with three agentic models, CoACT reports an average 33.0% reduction in total tokens while keeping PASS@1 close to (and on two models above) the uncompressed baseline, outperforming LLMLingua-2, LongCodeZip, and SWE-Pruner on the efficiency–effectiveness trade-off and combining usefully with trajectory compression.

Significance. If the empirical trade-off holds under broader agent settings, CoACT is a useful systems contribution: observation compression that preserves KV-cache reuse is practically important, and explicitly tying compression supervision to agent next-action behavior is a clearer design principle than pure token-importance or code-structure heuristics. Strengths include multi-model evaluation, direct observation-compression baselines, reward ablations (Table IV), offline-vs-online ablation (Table V), trajectory-combination cost results (Table III), and distributional checks on observation lengths and steps. The work is reproducible in intent (code/data link) and the training pipeline is concrete enough for follow-on systems work.

major comments (3)
  1. Section I, rewrite of Eq. (1) into Eq. (2): the claim that preserving a_{t+1} at every compression step implies the same action trajectory and therefore the same PASS@1 is incomplete. An observation can be irrelevant to the immediate next action yet still be required later (e.g., a stack trace or file snippet used only after intermediate edits). In that case NAP can hold locally while the final trajectory diverges. The paper should either (i) weaken the formal implication to a practical proxy with explicit failure modes, or (ii) measure how often NAP-passing compressions still force recovery steps, different later actions, or different patches on held-out trajectories.
  2. Section III-B Eqs. (11)–(14) and Section VI-C: the action-preservation reward depends on a custom field-based action-similarity function with threshold θ=0.6 that agrees with human binary labels on only 80% of 100 pairs. That agreement rate is modest for a load-bearing training filter. Please report sensitivity of final PASS@1/Tokens to θ (and to M/K), and quantify how often candidates accepted under the automatic score would be rejected by human operational-intent labels—or how often accepted compressions still change the agent’s subsequent tool sequence despite high sim.
  3. Tables I–II (n=200 SWE-bench Verified instances): several headline deltas are small (e.g., Deepseek PASS@1 76.5%→75.0%; Qwen 57.0%→60.5%) and no confidence intervals, bootstrap estimates, or paired significance tests are reported. Given instance-level variance typical of SWE-bench, the claim of “maintaining” or “improving” effectiveness needs uncertainty quantification so readers can judge whether gains/drops are noise.
minor comments (5)
  1. Fig. 1 pipeline labels include “iiii” / “Num = K” style artifacts; clean the stage numbering and reward notation for camera-ready readability.
  2. Section IV-C: training uses 50 SWE-smith instances per model; briefly justify that this scale is enough for the 4B LoRA compressor and whether performance saturates with more data.
  3. Clarify whether Comp tokens in Tables I–II are always included in Tokens for every baseline (especially LLMLingua-2 / LongCodeZip / SWE-Pruner) so total-cost comparisons are apples-to-apples.
  4. Section VI-A teacher-model table: Deepseek teacher drops PASS@1 to 52.5% under Qwen agent—discuss whether this is candidate quality, reward mismatch, or over-compression so readers know how sensitive CoACT is to teacher choice.
  5. Minor wording: abstract/intro “unsatisfactory efficiency-effectiveness trade-off” is strong relative to LongCodeZip’s milder degradation; tone claims to match the quantitative gaps shown.

Circularity Check

0 steps flagged

No circular derivation: NAP is a training-time proxy; headline PASS@1 and token metrics are independent end-of-trajectory measurements on held-out data.

full rationale

CoACT is an empirical methods paper, not a first-principles derivation. Equation (1) states the desired constrained optimization (minimize tokens subject to PASS@1 within ε of the uncompressed agent). Equation (2) replaces the PASS@1 constraint with next-action preservation as a practical proxy for training, because full-trajectory PASS@1 is expensive and sparse. That rewrite is an incomplete sufficiency argument (if every next action is preserved then the action sequence matches), not a definition of the evaluation metric in terms of the training signal. Supervision is built by filtering teacher candidates with an action-preservation reward and a length-reduction reward, then SFT/LoRA-training a lightweight compressor; deployment uses only that compressor. The reported claims—33.0% average total-token reduction and PASS@1 close to (sometimes above) Vanilla on SWE-bench Verified across three agentic models—are measured independently of NAP: PASS@1 is end-of-trajectory test-pass rate on held-out instances, and Tokens is the sum of cached/uncached input, output, and compressor tokens. Ablations (Table IV) and online-alignment results (Table V) further treat AP/LR and offline/online stages as empirical levers, not tautologies. There is no self-definitional loop, no fitted parameter renamed as a prediction of the same quantity, no load-bearing uniqueness theorem imported from overlapping authors, and no renaming of a known empirical pattern as a forced result. Concerns about delayed-use information or imperfect action-similarity (θ=0.6, 80% human agreement) are correctness/proxy-validity risks, not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

The central efficiency-effectiveness claim rests on a small set of domain assumptions about agent trajectories, a methodological proxy (NAP) that rewrites the true but intractable PASS@1 constraint, several hand-chosen filtering hyperparameters, and the reliability of a custom action-similarity function. No new physical entities are postulated; the invented pieces are algorithmic constructs whose only external check is the final PASS@1 / token tables.

free parameters (3)
  • action-preservation threshold θ = 0.6
    Candidates are kept only if mean top-M action similarity ≥ θ; set to 0.6 by the authors and used both for training filters and for the human-agreement check.
  • candidate / reference counts (N, K, M, k) = N=8, K=8, M=3, k=4
    Teacher generates N=8 candidates; K=8 reference actions; top M=3 similarities averaged; top k=4 length-selected survivors become supervision. These control the supervision distribution.
  • LoRA rank and alpha = rank=64, alpha=128
    Compressor is adapted with LoRA rank 64, alpha 128; capacity of the learned compressor depends on these choices.
axioms (4)
  • ad hoc to paper If a compressed observation induces the same next action as the raw observation at every step, the compressed agent follows the same action trajectory and therefore achieves the same PASS@1.
    Stated in Section I and used to replace the intractable PASS@1 constraint (Eq. 1) with the NAP constraint (Eq. 2). It is a sufficient but not necessary condition and ignores information needed only for later steps.
  • ad hoc to paper A field-based action-similarity function (target file, search pattern, viewed line range, operation type) with threshold 0.6 adequately decides whether two agent actions express the same operational intent.
    Defined in Section IV-F and validated only by 80% agreement with human binary labels on 100 pairs (Section VI-C).
  • domain assumption Observations returned by the environment are a major and compressible source of token consumption in coding-agent trajectories.
    Motivating measurement in the introduction (45.7% on SWE-bench Verified, up to 67.8% on Terminal-Bench); standard in the cited agent-efficiency literature.
  • domain assumption Prefix KV-cache reuse is valuable and should be preserved; therefore observation compression (append-only) is preferable to trajectory rewriting when possible.
    Section II-B; used to motivate focusing on observation rather than trajectory compression.
invented entities (3)
  • next-action preservation (NAP) no independent evidence
    purpose: Dense, cheap proxy for the end-of-trajectory PASS@1 effectiveness constraint during observation compression.
    Core methodological construct of the paper; defined in Section I and operationalized via the action-preservation reward. Independent evidence is only the final empirical tables, not an external falsifiable prediction.
  • action-preservation and length-reduction rewards no independent evidence
    purpose: Filter teacher candidates that change next action, then prefer shorter survivors as compressor supervision.
    Section III-B; jointly define the training objective that implements Eq. 2.
  • two-stage offline-bootstrap / online-alignment compressor training no independent evidence
    purpose: Reduce train–deploy mismatch by first training on raw trajectories then continuing on trajectories whose earlier observations are already compressed.
    Section III-C and Algorithm 1; ablation in Table V shows the online stage improves PASS@1.

pith-pipeline@v1.1.0-grok45 · 22611 in / 3502 out tokens · 44762 ms · 2026-07-12T06:10:03.832304+00:00 · methodology

0 comments
read the original abstract

LLM-based coding agents solve software-engineering tasks through iterative interactions with development environments, where returned observations accumulate in the context and become a major source of inference cost. Observation compression reduces this cost by shortening observations before they are appended to the context. However, existing methods still exhibit an unsatisfactory efficiency-effectiveness trade-off, as they do not explicitly model how compression affects the agent's subsequent behavior. This paper proposes CoACT, an action-preserving observation compression method for coding agents. CoACT is built on next-action preservation (NAP), which requires a compressed observation to induce the same next action as the raw observation. By checking the agent's immediate next action, NAP provides a practical signal for whether a compression preserves the information needed for continued task solving. During training, a teacher model first generates multiple compressed candidates of each observation. CoACT then uses an action-preservation reward based on NAP to filter out candidates that would change the agent's next action, and uses a length-reduction reward to choose compact candidates as supervision for a lightweight compressor. Experiments on SWE-bench Verified with three agentic models show that CoACT reduces average total token consumption by 33.0% while maintaining task-solving effectiveness close to the uncompressed agent.

Figures

Figures reproduced from arXiv: 2607.02911 by Haorui Chen, Jia Li, Yitong Zhang, Yuancheng Zhu.

Figure 1
Figure 1. Figure 1: Overview of CoACT. difficult to solve directly because PASS@1 can be measured only after the full trajectory finishes. As discussed in Section I, CoACT addresses this difficulty by using next-action preservation (NAP) as the proxy for the PASS@1 constraint. Replacing the PASS@1 constraint in Equation (1) gives Equation (2). To solve Equation (2), CoACT trains a lightweight compressor with reward-selected s… view at source ↗
Figure 2
Figure 2. Figure 2: Observation-token distributions of Vanilla and [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Trajectory-step distributions of Vanilla and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] 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

46 extracted references · 10 linked inside Pith

  1. [1]

    Swe-agent: Agent-computer interfaces enable automated software engineering,

    J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. Narasimhan, and O. Press, “Swe-agent: Agent-computer interfaces enable automated software engineering,”Advances in Neural Information Processing Systems, vol. 37, pp. 50 528–50 652, 2024

    2024

  2. [2]

    Openhands: An open platform for ai software developers as generalist agents,

    X. Wang, B. Li, Y . Song, F. F. Xu, X. Tang, M. Zhuge, J. Pan, Y . Song, B. Li, J. Singhet al., “Openhands: An open platform for ai software developers as generalist agents,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 65 882–65 919

    2025

  3. [3]

    Trae agent: An llm-based agent for software engineering with test-time scaling,

    P. Gao, Z. Tian, X. Meng, X. Wang, R. Hu, Y . Xiao, Y . Liu, Z. Zhang, J. Chen, C. Gaoet al., “Trae agent: An llm-based agent for software engineering with test-time scaling,”arXiv preprint arXiv:2507.23370, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Agentless: Demystifying llm-based software engineering agents,

    C. S. Xia, Y . Deng, S. Dunn, and L. Zhang, “Agentless: Demystifying llm-based software engineering agents,”arXiv preprint arXiv:2407.01489, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Ai-driven self-evolving soft- ware: A promising path toward software automation,

    L. Cai, Y . Ren, Y . Zhang, and J. Li, “Ai-driven self-evolving soft- ware: A promising path toward software automation,”arXiv preprint arXiv:2510.00591, 2025

    arXiv 2025

  6. [6]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,”arXiv preprint arXiv:2210.03629, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Environmental injection attacks against gui agents in realistic dynamic environments,

    Y . Zhang, X. Li, L. Cai, and J. Li, “Environmental injection attacks against gui agents in realistic dynamic environments,” 2026. [Online]. Available: https://arxiv.org/abs/2509.11250

    arXiv 2026

  8. [8]

    Reducing cost of llm agents with trajectory reduction,

    Y .-A. Xiao, P. Gao, C. Peng, and Y . Xiong, “Reducing cost of llm agents with trajectory reduction,”arXiv preprint arXiv:2509.23586, 2025

    arXiv 2025

  9. [9]

    Swe-effi: Re-evaluating software ai agent system effectiveness under resource constraints,

    Z. Fan, K. Vasilevski, D. Lin, B. Chen, Y . Chen, Z. Zhong, J. M. Zhang, P. He, and A. E. Hassan, “Swe-effi: Re-evaluating software ai agent system effectiveness under resource constraints,”arXiv preprint arXiv:2509.09853, 2025

    arXiv 2025

  10. [10]

    AI | 2025 Stack Overflow Developer Survey

    “AI | 2025 Stack Overflow Developer Survey.” [Online]. Available: https://survey.stackoverflow.co/2025/ai

    2025

  11. [11]

    How do ai agents spend your money? analyzing and predicting token consumption in agentic coding tasks,

    L. Bai, Z. Huang, X. Wang, J. Sun, R. Mihalcea, E. Brynjolfsson, A. Pentland, and J. Pei, “How do ai agents spend your money? analyzing and predicting token consumption in agentic coding tasks,”arXiv preprint arXiv:2604.22750, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    More with less: An empirical study of turn-control strategies for efficient coding agents,

    P. Gao and C. Peng, “More with less: An empirical study of turn-control strategies for efficient coding agents,”arXiv preprint arXiv:2510.16786, 2025

    arXiv 2025

  13. [13]

    Swe-bench: Can language models resolve real-world github issues?

    C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. Narasimhan, “Swe-bench: Can language models resolve real-world github issues?” inInternational Conference on Learning Representations, vol. 2024, 2024, pp. 54 107–54 157

    2024

  14. [14]

    Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces,

    M. A. Merrill, A. G. Shaw, N. Carlini, B. Li, H. Raj, I. Bercovich, L. Shi, J. Y . Shin, T. Walshe, E. K. Buchananet al., “Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces,” arXiv preprint arXiv:2601.11868, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    Longcodezip: Compress long context for code language models,

    Y . Shi, Y . Qian, H. Zhang, B. Shen, and X. Gu, “Longcodezip: Compress long context for code language models,”arXiv preprint arXiv:2510.00446, 2025

    arXiv 2025

  16. [16]

    Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression,

    Z. Pan, Q. Wu, H. Jiang, M. Xia, X. Luo, J. Zhang, Q. Lin, V . R ¨uhle, Y . Yang, C.-Y . Linet al., “Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 963–981

    2024

  17. [17]

    Swe-pruner: Self-adaptive context pruning for coding agents,

    Y . Wang, Y . Shi, M. Yang, R. Zhang, S. He, H. Lian, Y . Chen, S. Ye, K. Cai, and X. Gu, “Swe-pruner: Self-adaptive context pruning for coding agents,”arXiv preprint arXiv:2601.16746, 2026

    Pith/arXiv arXiv 2026

  18. [18]

    Qwen/Qwen3.5-35B-A3B · Hugging Face,

    “Qwen/Qwen3.5-35B-A3B · Hugging Face,” Mar. 2026. [Online]. Available: https://huggingface.co/Qwen/Qwen3.5-35B-A3B

    2026

  19. [19]

    Deepseek-v4: Towards highly efficient million- token context intelligence,

    A. Xu, B. Lin, B. Xue, B. Wang, B. Xu, B. Wu, B. Zhang, C. Lin, C. Dong, C. Linget al., “Deepseek-v4: Towards highly efficient million- token context intelligence,”arXiv preprint arXiv:2606.19348, 2026

    arXiv 2026

  20. [20]

    Gemini 3 Flash: frontier intelligence built for speed,

    “Gemini 3 Flash: frontier intelligence built for speed,” Dec. 2025. [Online]. Available: https://blog.google/products-and-platforms/products/ gemini/gemini-3-flash/

    2025

  21. [21]

    What papers don’t tell you: Recovering tacit knowledge for automated paper reproduction,

    L. Li, R. Wang, H. Song, Y . Mao, T. Zhang, Y . Wang, J. Fan, Y . Zhang, J. Ye, C. Zhanget al., “What papers don’t tell you: Recovering tacit knowledge for automated paper reproduction,”arXiv preprint arXiv:2603.01801, 2026

    arXiv 2026

  22. [22]

    The complexity trap: Simple observation masking is as efficient as llm summarization for agent context management,

    T. Lindenbauer, I. Slinko, L. Felder, E. Bogomolov, and Y . Zharov, “The complexity trap: Simple observation masking is as efficient as llm summarization for agent context management,”arXiv preprint arXiv:2508.21433, 2025

    arXiv 2025

  23. [23]

    Scaling long-horizon llm agent via context-folding,

    W. Sun, M. Lu, Z. Ling, K. Liu, X. Yao, Y . Yang, and J. Chen, “Scaling long-horizon llm agent via context-folding,”arXiv preprint arXiv:2510.11967, 2025

    arXiv 2025

  24. [24]

    Agentfold: Long-horizon web agents with proactive context management,

    R. Ye, Z. Zhang, K. Li, H. Yin, Z. Tao, Y . Zhao, L. Su, L. Zhang, Z. Qiao, X. Wanget al., “Agentfold: Long-horizon web agents with proactive context management,”arXiv preprint arXiv:2510.24699, 2025

    arXiv 2025

  25. [25]

    Scaling llm multi-turn rl with end-to-end summarization-based context management,

    M. Lu, W. Sun, W. Du, Z. Ling, X. Yao, K. Liu, and J. Chen, “Scaling llm multi-turn rl with end-to-end summarization-based context management,” arXiv preprint arXiv:2510.06727, 2025

    arXiv 2025

  26. [26]

    Compass: Enhancing agent long-horizon reasoning with evolving context,

    G. Wan, M. Ling, X. Ren, R. Han, S. Li, and Z. Zhang, “Compass: Enhancing agent long-horizon reasoning with evolving context,”arXiv preprint arXiv:2510.08790, 2025

    arXiv 2025

  27. [27]

    Acon: Optimizing context compression for long-horizon llm agents,

    M. Kang, W.-N. Chen, D. Han, H. A. Inan, L. Wutschitz, Y . Chen, R. Sim, and S. Rajmohan, “Acon: Optimizing context compression for long-horizon llm agents,”arXiv preprint arXiv:2510.00615, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Agentswing: Adaptive parallel context management routing for long-horizon web agents,

    Z. Feng, L. Su, Z. Zhang, X. Wang, X. Zhang, X. Wang, R. Fang, Q. Zhang, B. Li, S. Caiet al., “Agentswing: Adaptive parallel context management routing for long-horizon web agents,”arXiv preprint arXiv:2603.27490, 2026

    arXiv 2026

  29. [29]

    Context as a tool: Context management for long-horizon swe-agents,

    S. Liu, B. Jiang, J. Yang, Y . Li, J. Guo, X. Liu, and B. Dai, “Context as a tool: Context management for long-horizon swe-agents,” inFindings of the Association for Computational Linguistics: ACL 2026, 2026, pp. 20 604–20 617

    2026

  30. [30]

    Active context compression: Autonomous memory manage- ment in llm agents,

    N. Verma, “Active context compression: Autonomous memory manage- ment in llm agents,”arXiv preprint arXiv:2601.07190, 2026

    arXiv 2026

  31. [31]

    Claude Code by Anthropic | AI Coding Agent, Terminal, IDE

    “Claude Code by Anthropic | AI Coding Agent, Terminal, IDE.” [Online]. Available: https://claude.com/product/claude-code

  32. [32]

    Cursor — cursor.com,

    “Cursor — cursor.com,” https://cursor.com/

  33. [33]

    A self-evolving framework for efficient terminal agents via observational context compression,

    J. Ren, S. Wu, Y . Li, K. Zhu, S. Xu, B. Feng, R. Yuan, W. Zhang, R. Batista-Navarro, J. Yanget al., “A self-evolving framework for efficient terminal agents via observational context compression,”arXiv preprint arXiv:2604.19572, 2026

    Pith/arXiv arXiv 2026

  34. [34]

    Compressing code context for llm-based issue resolution,

    H. Jia, E. T. Barr, and S. Mechtaev, “Compressing code context for llm-based issue resolution,”arXiv preprint arXiv:2603.28119, 2026

    arXiv 2026

  35. [35]

    Llmlingua: Com- pressing prompts for accelerated inference of large language models,

    H. Jiang, Q. Wu, C.-Y . Lin, Y . Yang, and L. Qiu, “Llmlingua: Com- pressing prompts for accelerated inference of large language models,” inProceedings of the 2023 conference on empirical methods in natural language processing, 2023, pp. 13 358–13 376

    2023

  36. [36]

    Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression,

    H. Jiang, Q. Wu, X. Luo, D. Li, C.-Y . Lin, Y . Yang, and L. Qiu, “Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024, pp. 1658–1677

    2024

  37. [37]

    Provence: efficient and robust context pruning for retrieval-augmented generation,

    N. Chirkova, T. Formal, V . Nikoulina, and S. Clinchant, “Provence: efficient and robust context pruning for retrieval-augmented generation,” arXiv preprint arXiv:2501.16214, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Prompt compression with context-aware sentence encoding for fast and improved llm inference,

    B. Liskavets, M. Ushakov, S. Roy, M. Klibanov, A. Etemad, and S. K. Luke, “Prompt compression with context-aware sentence encoding for fast and improved llm inference,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 23, 2025, pp. 24 595–24 604

    2025

  39. [39]

    Codepromptzip: Code-specific prompt compression for retrieval-augmented generation in coding tasks with lms,

    P. He, S. Wang, and T.-H. Chen, “Codepromptzip: Code-specific prompt compression for retrieval-augmented generation in coding tasks with lms,”arXiv preprint arXiv:2502.14925, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Qwen/Qwen3.5-4B · Hugging Face,

    “Qwen/Qwen3.5-4B · Hugging Face,” Mar. 2026. [Online]. Available: https://huggingface.co/Qwen/Qwen3.5-4B

    2026

  41. [41]

    Swe-smith: Scaling data for software engineering agents,

    J. Yang, K. Lieret, C. Jimenez, A. Wettig, K. Khandpur, Y . Zhang, B. Hui, O. Press, L. Schmidt, and D. Yang, “Swe-smith: Scaling data for software engineering agents,”Advances in Neural Information Processing Systems, vol. 38, 2026

    2026

  42. [42]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” Iclr, vol. 1, no. 2, p. 3, 2022

    2022

  43. [43]

    Command similarity measurement using nlp,

    Z. Hussain, J. K. Nurminen, T. Mikkonen, and M. Kowiel, “Command similarity measurement using nlp,” in10th Symposium on Languages, Applications and Technologies (SLATE 2021). Schloss Dagstuhl–Leibniz- Zentrum f ¨ur Informatik, 2021, pp. 13–1

    2021

  44. [44]

    Models & Pricing | DeepSeek API Docs

    “Models & Pricing | DeepSeek API Docs.” [Online]. Available: https://api-docs.deepseek.com/quick start/pricing

  45. [45]

    Pricing

    “Pricing.” [Online]. Available: https://platform.claude.com/docs/en/ about-claude/pricing

  46. [46]

    OpenAI API Pricing

    “OpenAI API Pricing.” [Online]. Available: https://openai.com/api/ pricing/