acceptodds
Under review as a conference paper at ICLR 2027

ASM-RISE: Learning to Investigate Assembly under an Evidence Budget

Abstract

Assembly semantics is often distributed across control flow, call relations, and data dependencies rather than contained in the visible instruction sequence. Resolving a question therefore requires not only interpreting assembly, but also deciding which program evidence to inspect next. We introduce ASM-RISE, a framework that turns assembly understanding into budgeted program investigation. A structured environment exposes complementary static views through scoped actions with explicit evidence prices, while trajectory supervised fine-tuning teaches the interaction protocol and agentic reinforcement learning jointly improves evidence acquisition and semantic answering. We evaluate the learned policies on binary code similarity retrieval, tool-free program semantics, semantic answering under four evidence budgets, and general code generation. Across three model families, RL consistently improves retrieval and budgeted reasoning over the corresponding supervised initialization. On Qwen2.5-Coder-7B-Instruct, adaptive RL yields a 52.92% relative improvement in Recall@1 over SFT and outperforms both static-plan and no-tool RL on the evidence-dependent endpoints. Policy rankings differ between evidence-dependent tasks and tool-free semantics, showing that investigation quality and direct-answer proficiency are not interchangeable. ASM-RISE demonstrates that models can learn not only what assembly means, but how to investigate it.

Primary area:
reinforcement learning
Keywords:
Assembly Code Understanding, Agentic Reinforcement Learning, Adaptive Program Analysis, Sequential Evidence Acquisition, Cost-Aware Learning, Binary Code Similarity Detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4EFqOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ASM-RISE, a framework for learning to acquire and interpret program evidence under explicit budget constraints. To address the problem that key semantic information in assembly programs is distributed across different code regions, the authors first construct a structured evidence acquisition environment and formulate assembly understanding as a question-specific program evidence acquisition task under budget constraints. They then develop a training framework combining supervised fine-tuning and agentic reinforcement learning, enabling the model to autonomously select the program evidence to acquire, adjust subsequent investigation actions based on intermediate observations, and complete semantic understanding tasks within a limited budget. The authors conduct extensive experiments covering multiple tasks across three model families, Qwen2.5-Coder, Qwen3, and DeepSeek-Coder, and evaluate the learning gains, adaptivity, budget sensitivity, and complementary capabilities of the models and their variants.

Critical strengths
  1. Key semantic information in assembly programs may be distributed across different basic blocks, function calls, and data dependencies. How to effectively locate and acquire evidence relevant to the current question is a reasonable research problem with practical value.

  2. The authors construct a unified structured evidence acquisition environment that explicitly defines different evidence types, operation scopes, interaction procedures, and budget constraints, and integrate trajectory supervision and agentic reinforcement learning into a unified training framework.

Critical weaknesses
  1. One important contribution of this paper is the combination of SFT and agentic reinforcement learning to train the model's evidence acquisition and semantic understanding capabilities. However, existing studies have already explored reinforcement learning for multi-turn tool use and adaptive information acquisition, such as ToolRL, ReTool, and AutoSearch. The authors have not sufficiently clarified the incremental contributions of this work compared with existing methods.

  2. The authors compare three Qwen policies to evaluate the value of adaptive evidence acquisition, but the analysis mainly focuses on performance results and lacks an analysis of the model's own investigation behavior. Although Section 3.4 presents an example of adaptive evidence acquisition, the experimental analysis remains insufficient. Additional experiments to validate the underlying mechanism could examine whether the model has truly learned to adaptively change its next investigation location based on newly acquired evidence, and whether such adjustments help the model acquire more useful evidence and improve the quality of its final answers. This would provide more direct support for the paper's central conclusion regarding adaptive evidence investigation. In addition, the budget is quantified using manually defined costs, without comparing actual tool usage, token consumption, or latency across methods. Although acknowledged in Section 6, reporting these statistics would help demonstrate the practical value of the budget constraints.

Questions for authors
  1. Could the authors further compare ASM-RISE with existing methods such as ToolRL, ReTool, and AutoSearch, and clarify the incremental contributions of this work over these methods?

  2. I noticed that the baseline results in Table 2, except for Qwen3, are identical to those reported in Table 2 of the ASMA-Tune paper. However, the two papers report these results using Recall@1 and MRR, respectively. Could the authors confirm whether these two metrics have been confused?

  3. Could the authors provide additional experimental analyses of the model's own investigation behavior?

  4. If possible, could the authors provide statistics on tool usage, token consumption, and inference latency across different policies, and further analyze the relationship between the proposed method's performance and its actual runtime costs?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 4EFq

Thank you for your thoughtful feedback. We have incorporated the additional experiments and clarifications into the revised manuscript, with changes in the main text highlighted in blue.

Full denotes A-RL; Original denotes Qwen3.5-122B-A10B-FP8; DS denotes DeepSeek-V4.1-Flash-1017; GLM denotes GLM-5.3-Flash.

Q1

ASM-RISE builds on SFT and tool-use RL. Its contribution is learning which static program view and location to inspect, and when the evidence suffices to answer. ToolRL studies rewards for tool selection/application; ReTool learns reasoning interleaved with code execution; AutoSearch learns retrieval depth. ASM-RISE instead combines typed, scoped program evidence, heterogeneous access prices, and an objective linking semantic quality to grounding, evidence utility, and cost (Section 2).

PlanOnceFrozen uses Full's weights and initial information to generate a question-conditioned plan before any tool outputs, then executes it without revision. This control and the reward ablations test adaptive acquisition and the evidence-aware objective (Section 5.5). Full exceeds QualityOnly under all three judges while reducing tokens by 25.1% and evidence cost by 19.9%. This method-level comparison clarifies our research focus, supported by controlled assembly experiments.

Q2

Thank you for catching this error. We apologize: an incorrectly defined metric-label variable displayed Recall@1, and we failed to catch it in the compiled manuscript. Our 0.3355 and the published ASMA-Tune reference of 0.257 are both MRR. ASMA-Tune's normalized Recall@1 is not the quantity used in this comparison. The revised Section 4.2 and Table 2 state MRR; Appendix D identifies the published aggregate.

Source: ASMA-Tune, Table 2 and Section 5.1.

Q3

PlanOnceFrozen uses Full's trained checkpoint, question, initial information, and test budget. It generates the complete tool-and-scope plan before seeing any tool outputs, then executes it without replanning. All methods use the same 200 tasks from 100 source functions at each budget in {1, 2, 3, 5}, with budgets weighted equally.

Each quality score is an integer grade from 0 to 4 divided by four. Failed answers remain in the fixed denominator with score zero; independent-judge scores are reported separately.

In the recorded observation replay, Full receives symbols including fgetws and system, inspects bb_22 (a buffer-capacity check), and identifies command-execution risk. PlanOnceFrozen executes its preselected CFG-summary and symbol requests, but its answer omits that risk.

A separate same-state control selects cfg_block at bb_22 when the first observation is visible, versus function-level cfg_summary when masked. This tests observation-dependent action selection; the matched comparison below measures overall quality. Figure 4 in the revised manuscript shows the traces and control; Appendix F gives case details.

MethodOriginalDSGLMValid (%)
Full0.63530.64060.719799.25
PlanOnceFrozen0.53250.56190.563497.38
S-RL0.57750.54250.557898.63
N-RL0.53500.51160.472585.13
SFT0.43910.42340.450090.25

Full exceeds PlanOnceFrozen by 0.1028/0.0787/0.1563 despite matched weights, initial information, and budget (Section 5.5; Appendix G.1). The complete reward also improves downstream retrieval: Full's MRR is 0.3355, versus 0.2774 for QualityOnly and 0.30514 for NoCost, with the retriever and candidate pools fixed (Appendix G.3).

The case starts from 40 of 90 assembly lines at budget 3. Full and PlanOnceFrozen receive GLM scores of 1.00 and 0.75, and DS scores of 0.75 and 0.50. The inspected block checks remaining buffer capacity, not command contents; the frozen answer mentions execution and stack-canary protection but omits injection risk.

The case costs 1.5 evidence units for Full and 1.0 for the plan.

Q4

Table 10 reports actual resource use:

MethodTokens/taskCost/taskMedian latency (s)
Full5431.952.08461.0682
PlanOnceFrozen4196.761.99500.7526
S-RL3776.452.12100.4397
N-RL2408.380.00000.5967
SFT9685.761.90251.1588

Compared with PlanOnceFrozen, Full uses 29.4% more tokens and 4.5% more evidence cost, with 41.9% greater median recorded latency, alongside higher quality under every judge. Relative to QualityOnly, Full instead reduces tokens by 25.1% and cost by 19.9% while improving quality. Thus, adaptation buys higher quality, while the complete reward improves resource use within the adaptive setting. Cost sums the prices of executed evidence requests.

Against NoCost, Full also reduces tokens by 21.2% and evidence cost by 16.5%, with higher quality under all three judges.

Thank you again for your thoughtful feedback. We welcome any remaining questions about the revised results.

Reviewer_M9n7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ASM-RISE, a framework that formulates assembly code understanding as a budget-constrained evidence acquisition problem. The framework provides four structured program-analysis tools and adopts a two-stage training strategy combining trajectory SFT with agentic RL. Experiments on three model families demonstrate improvements in binary code similarity retrieval and budgeted assembly reasoning, suggesting the potential benefits of learning adaptive evidence acquisition for assembly understanding.

Critical strengths

1.The paper addresses an important and practical problem by introducing adaptive evidence acquisition into code understanding.

2.The proposed framework is clearly presented, and the experiments across three model families demonstrate promising performance improvements.

Critical weaknesses

W1. Insufficient evidence to isolate the effectiveness of adaptive evidence acquisition. The static-plan baseline follows a relatively simple heuristic strategy, while the no-tool baseline uses a different initial assembly context. Therefore, the improvements reported in Table 4 may not fully demonstrate the advantages of adaptive evidence acquisition itself.

W2. Lack of ablation studies on the reward design. The proposed reward combines answer quality, evidence grounding, evidence utility, and acquisition cost, but the individual contributions of these components are not evaluated. It remains unclear whether the improvements mainly come from semantic correctness optimization or from the proposed evidence-aware reward design.

W3. Limited validation of semantic evaluation and evidence efficiency. Budgeted Assembly Reasoning relies on an LLM-based judge that also provides training rewards, without independent validation of the evaluation reliability. Moreover, Figure 4 reports performance under allocated evidence budgets rather than actual evidence consumption, making it difficult to assess whether the proposed method achieves meaningful efficiency improvements.

Questions for authors

1.Could the authors compare ASM-RISE with a stronger question-conditioned static evidence selection baseline under matched evidence budgets?

2.Could the authors provide ablation studies to clarify the contributions of the grounding, evidence utility, and acquisition cost terms in the reward function?

1 comment
AuthorsOpenReview ↗

Response to Reviewer M9n7

Thank you for your thoughtful feedback. We have incorporated the additional experiments and clarifications into the revised manuscript, with changes in the main text highlighted in blue.

Full denotes A-RL; Original denotes Qwen3.5-122B-A10B-FP8; DS denotes DeepSeek-V4.1-Flash-1017; GLM denotes GLM-5.3-Flash.

Q1 / W1

PlanOnceFrozen uses Full's trained checkpoint, question, initial information, and test budget. It generates the complete tool-and-scope plan before seeing any tool outputs, then executes it without replanning. All methods use the same 200 tasks from 100 source functions at each budget in {1, 2, 3, 5}, with budgets weighted equally.

Each quality score is an integer grade from 0 to 4 divided by four. Failed answers remain in the fixed denominator with score zero; independent-judge scores are reported separately.

MethodOriginalDSGLMValid (%)
Full0.63530.64060.719799.25
PlanOnceFrozen0.53250.56190.563497.38
S-RL0.57750.54250.557898.63
N-RL0.53500.51160.472585.13
SFT0.43910.42340.450090.25

Full exceeds PlanOnceFrozen by 0.1028/0.0787/0.1563. This comparison removes the original position-based planning heuristic while matching weights and initial information (Sections 3.4 and 5.5).

As you note, N-RL receives a different initial assembly context. We therefore base the adaptive-execution comparison on PlanOnceFrozen, rather than the no-tool comparison alone.

In the recorded case, Full observes external symbols, inspects bb_22, and identifies command-execution risk; the frozen-plan answer omits that risk. A separate same-state control selects bb_22 when the observation is visible and a function-level CFG summary when it is masked. Figure 4 and Appendix F show the traces and case results.

The case starts from 40 of 90 assembly lines at budget 3. Full uses 6,762 tokens and costs 1.5, versus 5,173 tokens and cost 1.0 for the frozen plan. Their GLM scores are 1.00/0.75 and DS scores 0.75/0.50.

Q2 / W2 — Reward components

The ablations retain the adaptive interface. QualityOnly uses answer-quality reward; NoGrounding, NoUtility, and NoCost remove the corresponding term from the complete objective. Reward mappings and clipping are defined in Appendix E.5; controls and results appear in Section 5.5 and Appendix G.3.

MethodOriginalDSGLMTokensCost
Full0.63530.64060.71975431.952.0846
QualityOnly0.50660.50030.35387254.132.6025
NoGrounding0.54810.55130.70223223.362.1800
NoUtility0.62720.51500.70195466.142.1181
NoCost0.62030.55940.70596893.452.4961

Full improves quality under every judge while reducing tokens/cost by 25.1%/19.9% versus QualityOnly and 21.2%/16.5% versus NoCost. NoCost retains hard test-budget limits, demonstrating the additional role of the cost penalty within those limits. NoGrounding uses fewer tokens but has lower quality; removing utility also lowers quality, most strongly under DS.

Appendix F provides a socket-input case: Full obtains symbols and a CFG summary with two calls, cost 1.0 and 8,429 tokens; QualityOnly inspects two setup blocks before symbols, using three calls, cost 2.5 and 12,009 tokens. Both identify injection risk.

The ablation names denote reward components, not an unweighted sum of normalized evaluation scores. Removing grounding lowers Original/DS quality by 0.0872/0.0893; removing utility lowers DS quality by 0.1256. These distinct effects support evaluating grounding and utility separately rather than treating evidence use as one undifferentiated reward.

W3

DS and GLM rescore the generated answers separately and both preserve Full's advantage over PlanOnceFrozen and the reward ablations. The gains therefore extend beyond the training evaluator.

BCSD provides a judge-independent check with the retriever and candidate pools fixed: Full achieves 0.3355 MRR, versus 0.2774 for QualityOnly and 0.30514 for NoCost (Appendix G.3). Full's MRR gains are 0.0581 over QualityOnly and 0.03036 over NoCost.

Actual resource use is reported separately from allocated budgets (Table 10):

MethodTokens/taskCost/taskMedian latency (s)
Full5431.952.08461.0682
PlanOnceFrozen4196.761.99500.7526
S-RL3776.452.12100.4397
N-RL2408.380.00000.5967
SFT9685.761.90251.1588

Full's quality gains over PlanOnceFrozen accompany 29.4% more tokens, 4.5% more weighted evidence cost, and 41.9% greater median latency. The full reward, however, improves both quality and resource use over QualityOnly and NoCost. Evidence cost sums the prices of executed requests.

Thank you again for your thoughtful feedback. We welcome any remaining questions about the revised results.

Reviewer_vSn5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

ASM-RISE trains models to investigate assembly code under an evidence budget. Models use four static-analysis tools to choose what information to inspect and when to answer. Training combines teacher-trajectory SFT with GRPO rewards for answer quality, evidence use, and acquisition cost.

The paper evaluates three model families on binary similarity retrieval, two Juliet-based assembly evaluations, and HumanEval. Adaptive RL improves retrieval and budgeted answering over SFT and outperforms the fixed-plan and no-tool policies on these endpoints.

Critical strengths

The retrieval experiments provide useful evidence of learning gains. The authors fix the candidate pools and retriever and compare policies starting from the same SFT checkpoint. Qwen2.5-Coder-7B improves from 0.2194 to 0.3355 Recall@1, with adaptive RL outperforming both alternatives on all seven software projects. All three model families improve over SFT. These results do not depend on an LLM judge.

Critical weaknesses

1.The budget constraint lacks practical justification. Assembly semantics can require selective inspection, but the paper does not show that evidence acquisition is a substantial resource bottleneck in this setting. Tool prices are assigned manually, without measurements of token usage, latency, or actual analysis costs. Moreover, training uses a budget of 6, while five acquisitions costing at most 1 each can spend only 5. The hard budget therefore cannot bind during training. Results at smaller test budgets are useful, but do not establish a practical efficiency benefit. This weakens the motivation for a method built largely from existing SFT and RL techniques.

2.The controls do not fully establish the value of adaptive investigation. The fixed plan follows a hand-written order and selects blocks by position rather than relevance to the question. Adaptive RL therefore differs in information selection, replanning, and stopping, making their individual contributions unclear. Budgeted answering also uses the same model for training rewards and final grading, without independent validation. Each RL condition has only one training run, and there are no reward ablations under the same adaptive tool interface. These gaps limit confidence in the mechanism and robustness of the reported gains.

Questions for authors

1.Given the five-acquisition limit and maximum cost of 1 per acquisition, can the training budget of 6 ever bind? Are there additional costs or budget settings missing from the description?

2.Could you compare adaptive investigation with a model that creates a question-specific plan before seeing tool outputs and then follows it without revision? Reporting actual calls, tokens, and latency would also help establish a practical efficiency benefit.

3.Can an independent judge or expert audit confirm the budgeted-answering gains? With the adaptive interface unchanged, how does performance change when using answer-quality rewards alone or removing the evidence and cost terms?

1 comment
AuthorsOpenReview ↗

Response to Reviewer vSn5

Thank you for your thoughtful feedback. We have incorporated the additional experiments and clarifications into the revised manuscript, with changes in the main text highlighted in blue.

Full denotes A-RL; Original denotes Qwen3.5-122B-A10B-FP8; DS denotes DeepSeek-V4.1-Flash-1017; GLM denotes GLM-5.3-Flash.

Q1 — Training budget and constrained training

You are correct: five acquisitions costing at most one unit each cannot exhaust the original training budget of 6. Cost sensitivity is learned through the reward penalty; smaller test budgets impose hard access limits. Section 3.3 now distinguishes these roles. Evidence prices specify access allowances rather than wall-clock or monetary costs.

We also trained TightBudgetFull with budgets sampled uniformly from {1, 2, 3, 5}, using seed 11 for both conditions. Mean cost decreases from 2.0846 to 1.9936 (4.4%), and valid answers reach 100% at every test budget. Original/DS/GLM quality is 0.6135/0.6378/0.7122, versus Full's 0.6353/0.6406/0.7197. Thus, tighter-budget training retains most quality at lower expenditure. Section 5.2 and Appendix G.4 report the results, including each budget.

At B = 1, TightBudgetFull improves DS/GLM quality from 0.5675/0.6663 to 0.5775/0.6975. At B = 5, GLM favors Full (0.7475 versus 0.6938). The per-budget results show where tighter training constraints help and where quality decreases.

Q2

PlanOnceFrozen uses Full's trained checkpoint, question, initial information, and test budget. It generates the complete tool-and-scope plan before seeing any tool outputs, then executes it without replanning. All methods use the same 200 tasks from 100 source functions at each budget in {1, 2, 3, 5}, with budgets weighted equally.

Each quality score is an integer grade from 0 to 4 divided by four. Failed answers remain in the fixed denominator with score zero; independent-judge scores are reported separately.

MethodOriginalDSGLMValid (%)
Full0.63530.64060.719799.25
PlanOnceFrozen0.53250.56190.563497.38
S-RL0.57750.54250.557898.63
N-RL0.53500.51160.472585.13
SFT0.43910.42340.450090.25

Full exceeds the question-conditioned frozen planner by 0.1028/0.0787/0.1563 across the three judges, supporting adaptive execution beyond the original hand-written static plan (Section 5.5; Appendix G.1).

In the recorded case, Full observes external symbols, inspects bb_22, and identifies command-execution risk; the frozen-plan answer omits that risk. A separate same-state control selects bb_22 when the observation is visible and a function-level CFG summary when it is masked. Figure 4 and Appendix F show the traces and case results.

Full uses 5431.95 versus 4196.76 tokens/task, costs 2.0846 versus 1.9950, and has median recorded latency of 1.0682 versus 0.7526 s: increases of 29.4%, 4.5%, and 41.9%. This quantifies the quality–cost trade-off; Table 10 reports all five methods. Cost is the weighted sum of executed evidence requests.

The frozen-plan comparison tests adaptive acquisition and stopping jointly.

Q3

Yes. DS and GLM independently rescore the generated answers; both favor Full over PlanOnceFrozen and every reward ablation. The ablations retain the adaptive interface and remove components other than answer quality individually or jointly:

MethodOriginalDSGLMTokensCost
Full0.63530.64060.71975431.952.0846
QualityOnly0.50660.50030.35387254.132.6025
NoGrounding0.54810.55130.70223223.362.1800
NoUtility0.62720.51500.70195466.142.1181
NoCost0.62030.55940.70596893.452.4961

QualityOnly uses the quality component; NoGrounding, NoUtility, and NoCost remove the named term. Appendix E.5 defines reward mappings and clipping; Appendix G.3 describes the controls. Full reduces tokens/cost by 25.1%/19.9% versus QualityOnly and 21.2%/16.5% versus NoCost, with higher quality under all judges. NoGrounding uses fewer tokens but yields lower quality.

The socket-input case in Appendix F illustrates efficient evidence selection: Full uses two calls, cost 1.0 and 8,429 tokens, versus QualityOnly's three calls, cost 2.5 and 12,009 tokens; both identify injection risk.

NoCost retains hard test-budget limits. Its higher expenditure therefore shows that the reward penalty further regulates acquisition within those limits.

W2

The supplementary training comparisons use seed 11 and do not estimate cross-seed variability. The same-checkpoint planner, independent judges, and reward ablations provide complementary checks on the observed gains.

For clarity, retrieval performance 0.3355 is MRR; the revised manuscript corrects the earlier Recall@1 label.

Thank you again for your thoughtful feedback. We welcome any remaining questions about the revised results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.