acceptodds
Under review as a conference paper at ICLR 2027

TRACS: Combining LLM-Generated Plan Trajectories with Classical Search for Numeric Planning

Abstract

Numeric planning requires long-horizon reasoning under logical and numerical constraints. Advances in model architecture and increases in model scale have strengthened the planning capabilities of frontier large language models (LLMs), but a single inapplicable action can invalidate an entire generated plan. We introduce TRACS, a solver-centered hybrid that exploits LLM planning capability without requiring the model to solve each problem independently. Stagnation signals trigger occasional requests for LLM continuations from selected states. Symbolic replay retains valid prefixes for integration into native eager or lazy search. Symbolic validation preserves plan validity, while bounded intervention preserves the backbone's conditional completeness. We instantiate TRACS with a 9B proposer trained on solver-verified residual continuations using supervised fine-tuning followed by policy optimization, rewarding executable progress and penalizing state revisits. Across Depots-Num and ZenoTravel-Num, the lazy-search variant of TRACS achieves full coverage and substantially improves solution quality over native planners under the reported deployments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
numeric planning, large language models, heuristic search, neuro-symbolic reasoning, automated planning

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_tx6VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a hybrid LLM-numeric search planning framework called TRACS. It essentially pairs a classical numeric planner with a fine-tuned LLM: planners run as usual, and cheap stagnation detectors flag when search is stuck. At those moments, an LLM proposer suggests trajectories from the current state. The planner replays each suggestion, keeps only the legal prefix, and feeds it into its eager/lazy heuristic search. This paper seeks to improve search performance based on coverage, solving success/plan quality (Q), and expansion reduction. The proposer (qwen3.5-9B) was trained on pieces of solver-verified solution with SFT and then with DAPO (GRPO-style) reinforcement learning. The paper tests 2 TRACS variants against 5 baselines: 2 native planners (TRACSEager against NLM-CutPlan and TRACS-Lazy against Count Downward) and 3 LLMs (kimi-k2.6, deepseek-v4.1-flash, and glm-4.7. Domains used were IPC DEPOTSNUM and ZEROTRAVEL-NUM.

Critical strengths
  • The experimental performance of TRACS-Lazy is strong. It achieves 100% coverage across both evaluation domains, compare to 88.2% for the baseline Count Downward method.
  • This paper fits well with the LLM-Modulo direction the planning community has been moving towards (Kambhampati et al., 2024), just in a different context. The LLM only proposes, and the planner replays each action and admits just the legal prefix. Using valid but incomplete prefixes as guidance is a smart design choice, and it applies more widely than methods that need complete LLM plans. The paper plays to complementary strengths instead of forcing an LLM into planning, which makes sense given the decades of planning tools / infrastructure we already have.
  • Empirical gains are large where they should be. TRACS-Lazy solves every test problem in both domains. At Depots-scale-30 (Table 1), it solves 50/50 against 39/50 for Count Downward, Q goes from 0.71 to 0.93. At scales 30-40 (Table 2) it needs roughly ~21-27% fewer expansions at matched plan quality. The pairedexpansion analysis reduces the dependence on hardware speed. The ablations also support the design choices.
Critical weaknesses
  • One of my main concerns is that the evaluation is narrower in domain count and may lack comparison on standardized public sets. The source of the Depots-num and Zenotravel-num domains have additional numeric domains in their public sets and results on several of these are reported to show generalization. The instances used in the paper are self-generated which is required in large quantity for LLM-proposer training. Per scale, the test sets are larger than the official IPC sets. However, there are distributional differences, e.g., in-solver evaluation stops at 40 entities, whereas the official sets reach 70-80 objects / entities (IPC 2002 Depots, IPC 2023 ZenoTravel). On the official ZenoTravel set, coverage is already saturated [5] to complete coverage. Thus reporting the performance of TRACS on public sets of these and additional domains could better contextualize the contribution.
  • The gains may come from the intervention, i.e. detect a plateau / stagnation, then inject a validated trajectory from an assisting mechanism. Injecting trajectories at stagnation points has some existing LLM-free techniques, e.g. random-walk explorations [1,2] in propositional planning domains structurally show almost the same design. Closer to numeric planning, random walks are also shown to work on integer-based problems [3] through numeric to propositional encoding. Where propositional encoding is not practical, TRACS own LLM-based trajectory generation and replay steps can be replaced with random walks (or relaxed-plan prefixes) utilizing the planner's numeric successor generator. Similarly, LLM-based search initializations or one-time LLM-generated heuristics are shown to have strong performance on domains including Zenotravel in numeric setting [4]. Without such controls, the paper does not propoerly contextualize the role and extent of LLM’s contribution.
  • TRACS and the native planner baselines ran on different CPUs and allocations (TRACS also used a GPU). The wall-clock results therefore are a bit unclear, and the expansion counts excluded inference and replay costs. The throughput and paired expansion cutoffs help, but overall, it might need more clarification.
  • There are no experiments injecting non-LLM guidance (e.g., random valid rollouts) at the same trigger points. It’s unclear how much really comes from the trained proposer and how much from injecting extra states. Have the authors tried this, for example with the untrained base model in the search?
  • Only 2 domains were tested, each with its own specific proposer (+ ablations are only on Depots). The scale-40 results rest on 13-15 problems, and the matched lazy comparison there has n=4. The seeded reruns vary a lot at scale-40, and the eager gains are small and mixed.

[1] Nakhost et al. Monte-Carlo Exploration for Deterministic Planning. IJCAI 2009.

[2] Xie et al. Adding Local Exploration to Greedy Best-First Search in Satisficing Planning. AAAI 2014.

[3] Nakhost et al. Resource-Constrained Planning: A Monte Carlo Random Walk Approach. ICAPS 2012.

[4] Tuisov et al. Successor-Generator Planning with LLM-generated Heuristics. AAAI 2026.

[5] Scala et al. On Using Lazy Greedy Best-First Search with Subgoaling Relaxation in Numeric Planning Problems. ICAPS 2025.

Questions for authors
  • Does the gain come from the trained proposer or from injecting extra valid states at the trigger points? I did not find a control that injects non-LLM guidance. I also did not see the untrained qwen3.5-9B base used inside the search. Could you add either one? “Windowed-random” still queries the trained model.
  • The proposer is trained on states from solved solution paths but queried at stagnation states. Do you have an analysis of how different these two distributions are (for example, by comparing usable-prefix rates on actual stagnation queries?)
  • Could the authors explain why they specifically chose DEPOTS-NUM and ZENOTRAVEL-NUM as the 2 IPC domains for experiments/training? They partially explain in Appendix D, but it only supports the theory and not selection rationale. It can give insight on what other domains can be used for future work.
  • With the same detectors, budgets (B, H, K) and priority queue, how does TRACS perform when the injected trajectories come from other exisitng trajectory proposing mechanisms like random applicable walks without state revisits or relaxed-plan lookahead prefixes. A clear advantage for the trained proposer would substantially raise my score.
  • Can you report results on the official instances (IPC 2002 Depots-numeric, IPC 2023 ZenoTravel), and potentially on at least a few more numeric domains, under the IPC protocol? Where coverage is saturated, please include additional metrics that are already reported in current work e.g. plan quality.
1 comment

Thank you for the feedback. We added experiments in Section 5.3 to clarify when trajectory guidance helps and how far its benefits extend.

Q1 and Q4. Transferring the DEPOTS-NUM-trained proposer to ZENOTRAVEL-NUM yielded no observed gain over Count Downward (Table 4). This suggests that the gains depend on the guidance the proposer can supply in the target domain. In the DEPOTS-NUM offline evaluation, DAPO reached the goal more often than both SFT and Qwen3.5-9B Base (Table 3(a)). In the separate seed-101 solver comparison, DAPO achieved higher average plan quality than SFT over the same search expansion intervals (Appendix C.7). The partial-only control retained most of the coverage gain after filtering out all goal-reaching proposals. Under the same per-phase sampling cap, adaptive TRACS achieved higher coverage than the initial-only control, which queried only the original initial state at the start of each search phase (Table 3(b)). Comparisons with the base model or non-LLM trajectory generators within the same search setup remain open.

Q2. On stagnation-triggered queries at scales 30–40, about 79–89% of TRACS–Lazy generations yielded a nonempty, executable prefix without state revisits across the two domains (Table 10). This shows that training on states sampled from verified solution paths can yield usable guidance at search-time stagnation states. The difference between the two state distributions remains unquantified.

Q3. The choice was pragmatic. We selected two established numeric planning domains with different demands. DEPOTS-NUM couples stack rearrangement with truck capacity constraints, and ZENOTRAVEL-NUM couples passenger routing with fuel constraints. Each TRACS variant and its native counterpart were evaluated on the same held-out instances. We have included the PDDL instance generators in the supplementary material.

Q5. Evaluation on official IPC instances and additional numeric domains remains open. On newly generated DEPOTS-NUM instances at scales 45 and 50, the full TRACS–Lazy solver improved both coverage and plan quality over Count Downward without additional proposer training (Table 3(c)).

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

This paper proposes TRACS, a framework that combines classical heuristic search with LLM-generated plan trajectories for numeric planning. The main idea is the LLM is queried when the search appears to stagnate on the residual planning problem. The proposed trajectory is replayed using the symbolic model, only the valid prefix is retained, and this is then incorporated into the native eager or lazy search. The LLM proposer is trained using solver-generated residual problems, first through SFT and then via policy optimisation. The overall idea seems sensible, particularly the decision not to rely on the LLM for correctness. The actions are symbolically validated and the underlying planner remains responsible for the final solution. The results for the lazy-search variant are also quite strong. My main concerns are around the novelty relative to previous LLM-guided planning, the relatively limited experimental evaluation, and whether the paper establishes that an LLM is specifically required for the observed improvements.

Critical strengths
  • I like the decomposition between the LLM and the classical planner. The LLM potentially provides a useful long-range guide, while symbolic validation prevents hallucinated or invalid actions from compromising correctness.
  • The results empirical for TRACS-Lazy are strong, particularly on the harder instances. They are in fact sufficiently large improvements that I think the underlying idea is worth investigating.
  • The paper is generally careful in its evaluations and includes useful ablations on proposer training, adaptive triggering, and the number of generated trajectories. I also appreciate that the authors explicitly distinguish reductions in native search expansions
Critical weaknesses
  1. I am not fully convinced about the level of methodological novelty relative to previous LLM-guided planning.

In particular, Silver et al. (2022), and the co-authors, already proposed using valid prefixes of imperfect LLM-generated plans to seed heuristic search. Conceptually, this seems quite close to the central idea here, in the sense that an LLM trajectory can move search to a useful region even when the LLM cannot generate a complete valid plan. I agree that TRACS extends this: queries occur adaptively during search, are asynchronous, operate on residual problems, and are applied to numeric planning. However, I think the paper needs to establish more clearly how much of the improvement comes from these new components rather than from the previously established idea of LLM plan guidance. The most useful baseline for me would be relatively simple, to use use the same trained proposer, generate trajectories only from the initial state as in Silver et al., validate their prefixes and insert them into the same search backbone, under the same generation budget. I think this would make the novelty clearer.

  1. I find the evaluation too narrow to establish generality, particularly because a separate proposer is trained for each domain.

The experiments cover only Depots-NUM and ZenoTravel-NUM. At the same time, the proposer is trained independently for each domain using thousands of solver-generated residual problems. This is quite a different setting from a domain-independent classical planner. I think the current experiments convincingly show that domain-specific learned trajectory guidance can substantially improve this particular lazy-search planner on these two domains. I am not however convinced that they establish a more general improvement in numeric planning. This is also related to the train/test construction. The test set is obtained through a random split of generated instances from the same problem pool. As discussed recently by Katz et al. (2025) [“Make Planning Research Rigorous Again!”], this can be problematic for planning benchmarks because separately generated instances can be structurally very similar. I would find results across considerably more IPC numeric domains, or at least genuinely distribution-shifted/independently generated test sets, much more convincing. The scale-extrapolation experiment is interesting, but does not completely address this issue, for example, at scales 45–50 the model no longer solves complete residual problems, and the paper evaluates whether useful prefixes remain rather than whether the method actually improves the final planner at those scales.

  1. It is unclear whether the benefit is specifically coming from an LLM rather than from learned macro-action/trajectory guidance. This is probably my main methodological question, related to my previous question. The LLM receives a lot of domain-specific training from solver-generated trajectories. Therefore, it is difficult to know whether pretrained language-model knowledge is providing something important, or whether almost any sufficiently capable sequence policy trained on these data would learn useful continuations. Given that the symbolic transition model is known, I think the classical comparisons such as heuristic search, macro-actions/preferred operators, learned policies and learned/generated heuristics would be relevant, and of course recent work in LLM-generated planning heuristics.
Questions for authors
  1. Could the authors compare TRACS against a Silver et al.-style baseline using exactly the same trained proposer and approximately the same total generation budget, but generating guidance only from the initial state?
  2. Could the authors clarify whether they have tested a substantially simpler non-LLM trajectory generator trained on the same solver data, or an obfuscated-domain version where semantic names are removed?
  3. Can the authors provide evidence on additional numeric-planning domains or stronger distribution shift? In particular, could full TRACS performance be reported at the extrapolation scales rather than only the quality of generated prefixes?
1 comment

Thank you for the feedback. We added experiments in Section 5.3 to clarify the value of adaptive querying, the contribution of partial continuations, and how far the solver’s gains generalize.

1. Initial-state guidance versus adaptive querying. We added a Silver et al. (2022)-inspired control that queries the original initial state at the start of each search phase. It uses the same trained proposer, search backbone, and per-phase sampling cap as adaptive TRACS. Initial-state guidance modestly improves coverage over Count Downward, while adaptive querying yields a substantially larger gain (Table 3(b)). This supports using states reached during search to generate continuations tailored to the remaining problem.

2. What does the LLM contribute? We agree that a simpler trajectory generator trained on the same solver data and an evaluation with semantic names removed would help clarify the role of the pretrained LLM. Our current evaluation leaves that question open. The new partial-only control addresses which outputs account for the solver’s gains. After discarding every proposal that reaches the goal, TRACS retains most of its coverage gain over Count Downward (Table 3(b)). This shows that executable continuations can improve search even when they stop short of a solution.

Regarding concerns about the dataset, we have uploaded our PDDL problem generation algorithm in the supplementary material. In short, the process involves randomly distributing base objects, followed by the addition of corresponding auxiliary objects and predicate relationships. You may examine the source code to verify the specific algorithm. Given this approach, it is virtually impossible to generate problems with identical structures. Our test set was derived from the original pool of problems; our independent verification confirms that none of the problems in the test set are isomorphic to those in the training set.

3. Full solver performance and generalization. Table 3(c) now reports full TRACS–Lazy runs on newly generated DEPOTS-NUM instances at scales 45 and 50. With the same proposer and no further training, TRACS–Lazy improves both coverage and plan quality over Count Downward, even though the proposer solved none of the larger instances in the separate offline evaluation. This shows that trajectory guidance remains useful to the full planner beyond the proposer’s training scales.

Reviewer_mc1iOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose TRACS, a framework that combines LLM-generated trajectories with classical search for numeric planning. Instead of relying on an LLM to generate a complete plan, TRACS uses stagnation signals to identify when the classical planner is not making enough progress and requests an LLM-generated continuation from the current search state. The proposed actions are checked using the symbolic planner, and the valid portion of the trajectory is incorporated into the search. The authors consider both eager and lazy search, with LLM requests handled asynchronously so that the planner can continue searching while waiting for responses.

The authors also introduce a training pipeline where a 9B LLM is first fine-tuned on solver-verified plan continuations, followed by reinforcement learning that rewards goal completion and useful partial progress while penalizing repeated states and invalid actions. The method is evaluated on two numeric planning domains, Depots-Num and ZenoTravel-Num. The results show improvements in solution coverage and plan quality over the classical planner baselines, particularly for the lazy-search variant. Additional ablations examine the effects of adaptive triggering, proposer training, and generalization to larger problem sizes.

Critical strengths

Combining LLMs with classical search can address some of the issues with end-to-end LLM planners, while also providing useful starting points when classical search gets stuck. The proposed method provides a way to use partially valid LLM-generated trajectories to guide classical search.

The authors show improvements over the classical planner baselines, particularly for lazy search in several settings. The authors also provide ablations to study the effect of different components of the proposed method. However, the evaluation is limited to only two planning domains.

Critical weaknesses

The main concern is the lack of comparisons against existing methods that already combine LLMs with classical search. Prior work, such as Silver et al. (2022), has explored using LLM-generated plan prefixes to guide search, while methods like Marvin have used action sequences to escape heuristic plateaus. The authors compare against classical planners and end-to-end LLM baselines, but not against these existing hybrid approaches. This makes it difficult to tell how much of the improvement comes from the proposed stagnation detection and trajectory integration methods, versus simply adding LLM-generated trajectories to classical search.

Another concern is the limited evaluation. The experiments cover only two planning domains, with a separate LLM trained for each domain, so it is unclear how well the approach would generalize to other planning problems. The authors also test larger problem sizes, but only evaluate the LLM's ability to generate partial trajectories rather than the performance of the complete planner. Additionally, Table 3(b) shows that the SFT-only model already achieves full coverage and nearly the same final plan quality as the RL-trained model. This raises questions about how much the proposed RL training actually contributes to the overall improvement.

Questions for authors
  1. The authors compare TRACS against classical planners and end-to-end LLMs, but not against existing methods that already combine LLMs with classical search. For instance, Silver et al. (2022) already explored using LLM-generated plan prefixes to guide classical search. Why is there no comparison against such an approach? Even a simpler baseline that uses the same LLM and inference budget to generate partial trajectories and add them to the search frontier, without adaptive stagnation detection, would help show what TRACS actually adds. Other recent approaches, such as Inspire or Predict? (2025) and Classical Planning with LLM-Generated Heuristics (NeurIPS 2025), also seem relevant and would be worth discussing, even if direct comparisons are difficult due to differences in planning settings..

  2. The evaluation is limited to two planning domains, with a separate LLM trained for each. Does this mean the LLM needs to be retrained for every new domain? How well would the approach generalize to domains it has not been trained on?

  3. In Table 3(c), the authors test larger problem sizes but only evaluate whether the LLM can generate useful partial trajectories. Do these trajectories actually help the complete planner? Can the authors show results for the full TRACS framework on these larger problems?

  4. In Table 3(b), the SFT-only model already achieves full coverage and nearly the same final plan quality as the RL-trained model. How much is the proposed RL training actually contributing? Do these improvements hold across different seeds and both planning domains?

  5. The authors use three stagnation signals: global stagnation, heuristic plateaus, and branch stagnation. However, the ablation only compares adaptive triggering against random intervention. Which of these three signals is actually responsible for the improvements? An ablation removing each signal individually would help clarify this.

2 comments

Thank you for the feedback. We added four experiments in Section 5.3 to clarify what drives TRACS’s gains and how far they generalize.

1. Adaptive trajectory guidance. Inspired by Silver et al. (2022), we added a control that queries the original initial state at the start of each search phase, using the same proposer and per-phase sampling cap as TRACS. Initial-state guidance improves coverage over Count Downward, and adaptive querying during search yields a substantially larger gain (Table 3(b)). Together with the window-random comparison, this supports using search progress to decide when and where to query. The partial-only control retains most of the coverage gain, showing that executable continuations can help the planner even when they fall short of the goal.

2. Cross-domain transfer. Applying the DEPOTS-NUM-trained proposer to ZENOTRAVEL-NUM without further training yielded no observed improvement over Count Downward (Table 4). The search integration carried over, while the proposer’s benefit did not transfer in this experiment. Training a shared proposer on a mixture of domains is a direction for future work.

3. Full solver performance beyond the training scales. The proposer did not reach the goal within three attempts on any scale-45 or scale-50 instance in the offline evaluation. In the full solver runs, TRACS–Lazy with the same proposer substantially improved both coverage and plan quality over Count Downward (Table 3(c)). This shows that trajectory guidance can remain useful to the planner at larger scales within the same domain, even when the proposer struggles to generate complete plans.

4. Contribution of RL beyond SFT. We agree that SFT already achieves full coverage and nearly the same final plan quality. DAPO reaches the goal more often in the offline evaluation with only a modest increase in effective prefix length (Table 3(a)). In the paired DEPOTS-NUM runs, DAPO also yields higher average plan quality over the same search expansion intervals (Appendix C.7). Therefore, although both exhibit similar final performance, RL actually enables the solver to reach a solution of the same quality more quickly.

5. Individual stagnation signals. We agree that removing each signal in turn would help identify its contribution. Our current comparison evaluates the three signals together. Their individual contributions remain unmeasured.

Reviewer_mc1iOpenReview ↗

Acknowledging the Response by Authors

Thanks to the authors for the additional experiments. However, I think some of my main concerns remain. In particular, it is still unclear whether the improvements specifically require an LLM, or whether simpler methods such as random valid rollouts or classical macro-action guidance at the same stagnation points could achieve similar results. Given the existing work on plateau-escaping search, I think this is an important comparison.

The cross-domain experiment also shows no improvement when transferring the proposer from Depots-Num to ZenoTravel-Num. This makes the generalization concern more significant, especially since the method requires separately training the proposer for each domain.

Overall, the new experiments address some of my original questions, but I don't think they fully establish the contribution over existing search-guidance approaches. I am maintaining my weak rejection rating.

Comments on the paper

1 comment

Thank you for your feedback. We have uploaded a revised manuscript with four additional experiments in Section 5.3.

  1. Partial versus complete continuations. We added a partial-only ablation that discards every LLM proposal that reaches the goal. Comparing it with the full TRACS configuration shows how much partial continuations alone contribute to search.
  2. Timing of LLM intervention. Inspired by Silver et al., we added a control that queries the LLM only from the initial state. We compare it with adaptive querying under the same per-phase sampling cap to assess the value of intervening as search progresses.
  3. Full solver runs beyond the training scales. Alongside the inference-only evaluation, we ran the full TRACS–Lazy solver on larger DEPOTS-NUM instances to test whether the proposer can still improve search beyond its training scales.
  4. Cross-domain transfer. We used the proposer trained on DEPOTS-NUM to guide search on ZENOTRAVEL-NUM without further training.

We also updated the corresponding experimental protocols and discussion of the results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.