acceptodds
Desk rejected at ICLR 2027

Where to Think, What to Explore: Hierarchical Agentic Reasoning and Exploration for Coding Agents

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Coding agents work by interacting with a repository over many turns, so the compute they spend at inference time is as much a design choice as the model itself. That compute can go into reasoning inside a decision, more turns, more candidate actions, or parallel subagents, and these choices have not been compared under one accounting. We measure the four dimensions (reasoning depth , interaction horizon , action width , subagent fan-out ) and their interactions under one coding agent, model, benchmark, serving stack, and cost accounting. Scaled alone they saturate for different reasons: depth has a finite operating point, horizon reaches diminishing returns, width produces candidate diversity that naive selection fails to exploit, and delegation uptake limits passive subagents. They also interact: reasoning changes both the value of later interaction and the cost of parallel branching, and the depth of the step that reduces a fan-out to one action sets what that fan-out is worth. These measurements motivate HARE, Hierarchical Agentic Reasoning and Exploration, which treats and as a serial pair and and as a parallel pair. HARE reasons only at the turns that write to the repository or synthesize gathered evidence, and folds each fan-out back into one committed action with a single reduction step. On SWE-bench-Pro tasks with a Qwen3.8-27B model, HARE matches dense reasoning at 15% less wall-clock and raises the best measured resolve rate from 56.7 to 60.0%. Whole-trajectory sampling and sequential refinement stay below the frontier HARE reaches, at higher cost.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agentic execution, Compute allocation, Tool use, Tool calls, Reasoning, Coding agents

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_qPQjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a policy to better assign compute in the agentic coding tasks by exploring four different dimensionalities: the reasoning depth, the horizon, the alternative actions, and the read-only subagent fan-outs, and their combinations. The paper first run some pilot experiments and draw two observations to serve their HARE motivation. The paper then detailed their approach HARE, they select some turns mechanically for more reasoning depth, e.g., edit turns or first/last turns and they decide for the interaction horizon whether to continue.

Critical strengths
  • The paper did extensive experiments to show the performance of each configuration.
Critical weaknesses
  • The paper only studies the main question under one coding agent, model, benchmark, serving stack, and cost accounting. While I understand scale up the experiment to other models and benchmarks are expensive, but this posts a question and weakness on the generalizability of the approach.
  • Section 5.2 reports the negative results that the sparsity is the gain rather than where to select the reasoning depth. This is counter-intuitive, yet the paper does not further explain the reason of it. What about the control experiments of the context length because the sparsity also correlates with the context length, on which some research shows that longer context length (prefill) usually leads to worse results.

In general, I feel this is a paper with very extensive experiments, but lack what can be really drawn from the experiments. There are many learnings that we can derive from the results but the paper fails to deliver them.

Questions for authors
  1. in terms of the horizon, the context needs to equip with better context management, such as sliding window, llm summarization compaction, or other approaches like [1]. How does HARE's results interacts with these approaches? Does the policy need to be redesigned?
  2. Observation 2 should be mapped with concrete results, now it reads like without no support. For example, "spend depth at selected turns" because ...
  3. Figure 3 should be placed in page 3 to ensure readers understand what D,I,W,D are.
  4. Equation 2 formulation is confusing and the continue_t, D_t, W_t, S_t are left unexplained.
  5. How I controls the decision points and whether to continue, use another LLM? But it looks like in line 337, page 7, the I is also a constant cap. Then does HARE adds anything on top of the constant capping of I?
  6. For D, how do we know in advance that it is a end turn? It looks like it is entangled with the I which also contribute to decide whether to continue or to end.
  7. What is the reducer, another LLM?
  8. page 5, line 316-317 said that "never in the judge that picks among action proposals" However, Figure 2B shows the opposite that the judge reasoning iperforms the best in the reduction.
  9. why only two compositions, e.g., I think HARE-Scout, HARE-Width and HARE-Edit can be combined as well. And I don't see any definition of Compositions in Figure 4.
  10. page 7, line 364-365: "A misplaced reducer survives composition" -> However, lots of compositions perform better than the Judge stack.
  11. Section 5.3 still presents counter-intuitive results, which lead to further questions. for example, the judge/reducer is bad at width (action) but good at reading explorations reduction. does it mean the judge/reducer model is too weak to handle complex actions?

[1] https://arxiv.org/abs/2604.23069

Reviewer_sgR1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates how inference-time computation should be allocated in coding agents along four dimensions: reasoning depth (D), interaction horizon (I), action width (W), and subagent fan-out (S). Through controlled experiments with a Hermes-based coding agent, Qwen3.8-27B, and a 90-task subset of SWE-bench-Pro, the authors characterize the individual scaling behavior and interactions of these dimensions. With the analysis, the authors introduce HARE, a training-free framework that combines sparse reasoning at selected interaction steps with selective parallel action or context exploration, followed by a single reduction step.

Critical strengths
  1. The paper provides a unified investigation of different test-time scaling dimensions (reasoning depth, interaction horizon, action width, and subagent fan-out) under a consistent experimental configuration.
Critical weaknesses
  1. The analysis and experiments are all based on a single model, coding-agent framework, and one benchmark, SWE-bench-Pro, with the main comparison conducted on a 90-task subset. It is unclear whether the conclusions and results are generalizable to other models, agent architectures, or other tasks/benchmarks.
  2. Some design conclusions are also weaker than presented. For example, HARE-Rand nearly matches HARE-Edit, suggesting that sparse reasoning itself may be more important than the proposed write-trigger mechanism. Likewise, the rule against reasoning-based action judges is derived from a relatively narrow set of judge implementations and depth budgets rather than a broad characterization of possible reduction strategies.
Questions for authors

Please see weaknesses.

Reviewer_U6CGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores how compute can be most effectively allocated when Coding agents repair code to complete tasks. It starts by defining and categorizing compute allocation into four main types: reasoning depth, interaction horizon, action width, and subagent fan-out. The authors then run experiments to analyze how these different ways of increasing compute perform under various conditions, and propose their own approach featuring three mechanisms to help agents allocate compute more reasonably and solve problems more effectively.

Critical strengths

Strengths:

  1. The problem is valuable, and it provides a reasonable breakdown of compute usage. Based on the specific problem and this framework, it proposes a solution that enables Coding agent to solve problems more effectively.
Critical weaknesses

Weaknesses:

  1. Test coverage is relatively narrow. For example, it relies on just one Qwen 3.8-27B model, a single Harness, and only a subset of SWE-bench-pro.

  2. Some performance gains are not significant enough, making them less reliable under a single evaluation scenario. For instance, HARE-edit performs quite similarly to a strategy that randomly selects the reasoning rounds to enable. Also certain design choices within the HARE strategy lack sufficient evidence.. For example, why is a non-reasoning judge better, and why choose four candidate actions?

Questions for authors

Questions:

  1. Will these issues still occur with more powerful frontier models? What will happen if it switch to a different harness?

  2. Can the authors provide stronger experimental or theoretical evidence for the specific choices made by HARE? For example, why is it better for the non-reasoning judge? Why choose to think before making modifications when it performs similarly to thinking for a randomly selected turns?

  3. Are these three HARE strategies still effective on a new test set that wasn't involved in their design? What do the generalization results look like?

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

Related papers

Discussion (0)

Sign in to comment.