acceptodds
Under review as a conference paper at ICLR 2027

PackAgent: Evolving Three Coordinated Skills for State-Adaptive 3D Bin Packing

Abstract

The three-dimensional bin packing problem aims to pack heterogeneous items into as few bins as possible while satisfying geometric and physical constraints, including containment, non-overlap, and support. Item selection and placement decisions must adapt to changes in the remaining items and the current spatial configuration, whereas fixed heuristics struggle to accommodate the decision requirements of different packing phases. To address this challenge, we propose PackAgent, a three-skill evolutionary framework for constrained multi-bin three-dimensional packing. PackAgent decomposes the core decisions into three skills: state reflection , item selection , and placement scoring . A large language model jointly evolves these executable skills offline. At each placement step, generates phase cues and factor weights from the current packing state, dynamically modulating the decision preferences of and . Candidate generation, rotation enumeration, and constraint verification are handled by a fixed solver to ensure geometric feasibility. Unlike automated heuristic design methods that search for a single fixed rule or rely on post-hoc selection among multiple strategies, PackAgent enables state-adaptive decisions through step-level factor modulation within a single packing trajectory. We evaluate PackAgent on the SSSCSP benchmark under offline, online, and semi-online settings, as well as across instance scales on the IVC dataset. PackAgent achieves the lowest average bin count among the compared single-trajectory methods in all three information-visibility settings. Deployment ablations further demonstrate the importance of dynamic modulation through state reflection for coordinating item selection and placement scoring.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Three-dimensional bin packing, Large language models, Automated heuristic design, State-adaptive decision making, Evolutionary computation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_TMCeOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper is an adaptation of "Evolution of Heuristic" (EoH) on the specific scenario of 3D bin packing. More specifically, instead of direct program search from current state sts_t to selected item and position (jt,ct)(j_t, c_t), it provides an "inductive bias" that decompose the process as follows:

  • State Reflection: from sts_t to hyperparameters of item and position selections hth_t
  • Item Selection: select the optimal next item jtj_t to fill from a set, given sts_t and hth_t
  • Position Selection: select the optimal position ctc_t of the next item jtj_t from a set, given sts_t and hth_t

In such a way, instead of searching for a large program from sts_t to (jt,ct)(j_t, c_t), the proposed approach searches for three smaller programs, which may be more efficient. And an explicit state reflection program allows the optimization algorithm to change "phase" or "style" during the sequential filling process.

Critical strengths
  • This paper addressed an important optimization problem (3D bin packing)
  • The proposed three skills do make sense, expecially the explicit introduction of "state reflection" that allow the optimization process to adapt to different "phase" during the sequential filling process
Critical weaknesses
  • The proposed approach looks like an inductive bias over the program search space, rather than a fundamental extension of EoH's expressiveness. Let's say, if we have an EoH with a pretty intelligent LLM, and prompt it to design a good optimization program from sts_t to (jt,ct)(j_t, c_t), it may evolve the state reflection mechanism, or something even more clever, by itself?
  • The performance improvement is not that impressive, especially considering that the proposed approach works specically on 3D bin packing, which is much less universal than the baseline approaches like EoH and ReEvo.
Questions for authors

See "weaknesses"

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

This paper presents PackAgent, a framework for 3D bin packing in which a large language model evolves three coordinated, executable skills: state reflection, item selection, and placement scoring. At each packing step, the state reflection skill reads the current layout and produces phase cues and factor weights that dynamically modulate the other two skills, while a fixed solver handles candidate generation, rotation enumeration, and geometric constraint verification. The framework supports offline, online, and semi-online settings. Evaluated on the SSSCSP benchmark and the industrial IVC dataset, PackAgent achieves the lowest average bin count among single-trajectory baselines, with deployment ablations showing that state-conditioned weight modulation drives the gains.

Critical strengths

The main strength is the step-level state-adaptive mechanism: a state reflection skill that modulates the factor weights of item selection and placement scoring at every packing step, rather than evolving a single fixed rule as in EoH/ReEvo/EoH-S. The deployment ablations are well-designed.

Critical weaknesses

The decisive weakness is the absence of strong external baselines, which undermines the scientific value of the results. All main comparisons are within the LLM-heuristic-evolution family (EoH, ReEvo, EoH-S — the latter weakened by using only its single best rule). The SSSCSP benchmark was published together with exact solutions, yet no optimality gap is reported, so absolute solution quality is uninterpretable. No classical strongholds (extreme-point heuristics, beam search, tuned metaheuristics) and no DRL state of the art for the online setting (PCT, constrained DRL, Attend2Pack, DMRL) are benchmarked in any of the three operating modes. Consequently, the evidence supports only "better than other LLM-evolved heuristics," not "competitive with strong 3D-BPP solvers," and the reported margins (0.07–0.32 bins) may be meaningless in absolute terms.

Questions for authors

Q1. The SSSCSP benchmark was introduced with exact solutions. What is PackAgent's optimality gap (in bins and percentage) relative to these exact results per instance, in each of the three settings? Without this anchor, it is impossible to judge whether an average of ~15 bins is strong in absolute terms. Q2. Why are no classical baselines included — at minimum a tuned extreme-point/DBLF-style constructive heuristic and one established metaheuristic or beam search (e.g., Araya & Riff, which you cite)? The compared methods all share the same LLM-evolution paradigm; what evidence suggests the gains persist against non-LLM heuristic classes? Q3. For the online and semi-online settings, why is no learning-based SOTA compared — e.g., the PCT line (Zhao et al., 2022/2026) or constrained DRL (Zhao et al., 2021), both of which you cite? If the multi-bin objective precludes direct reuse, could you adapt one such baseline (or its policy) to the fewest-bins objective so the comparison is possible? Q4. The framework depends on a fixed solver (candidate generation, orientation enumeration, pallet-opening policy). How sensitive are the results to this fixed component? In particular, does the evolved agent remain effective if the pallet-opening policy or EMS generator is replaced, or is part of the reported performance attributable to the fixed solver itself? A comparison against an ablation where the fixed solver uses a standard heuristic rule would clarify this.

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

This paper investigates LLM-based Automatic Heuristic Design (AHD) for the three-dimensional bin packing problem (3D-BPP), aiming to minimize the number of bins while maximizing space utilization. The goal is to adapt state-dependent weights through three components: State Reflection, Item Selection, and Placement Scoring. The LLM evolves three executable programs ((\varepsilon,\phi,\psi)) rather than the entire geometric solver. The State Reflection component dynamically generates weight vectors to control the decision preferences of the other two components.

The proposed method is relatively simple and lacks substantial novelty.

Critical strengths

Figure 3 provides a clear and intuitive visualization of the packing results across different methods.

Critical weaknesses
  1. The proposed framework mainly combines existing AHD with state-dependent weight adaptation.

  2. PackAgent does not demonstrate clear improvements over classical AHD methods such as EoH and ReEvo.

Questions for authors
  1. Could the authors compare against stronger classical or GP-based packing heuristics and report the actual execution-time overhead of the generated three-skill agent?

  2. Why is EoH-S restricted to a single heuristic rather than its original heuristic-set evaluation?

  3. The improvements on SSSCSP are relatively small, particularly in the offline setting. Could the authors report results over multiple independent evolutionary runs?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.