acceptodds
Under review as a conference paper at ICLR 2027

StratTR: One Strategy Does Not Fit All in Temporal Reasoning

Abstract

Temporal reasoning involves distinct operations, including event-graph reasoning for relation questions, date arithmetic for duration questions, and direct extrac- tion for factual questions. Existing systems, however, typically use one prompting scaffold or policy across these question types, making the reasoning procedure does not match the operation need for the question. We introduce StratTR, a trainable multistrategy agentic framework that support four strategies (graph con- struction, code generation, direct extraction, and hybrid reasoning). StratTR dis- tills routed, multi-call trajectories from a proprietary teacher into a single-call 4B policy. To train the strategy-conditioned policy, StratTR leverages the key deter- ministic temporal signals across heterogeneous artifacts with reinforcement learn- ing. Across seven benchmarks comprising 22,728 questions, StratTR with a 4B- scale backbone outperforms the strongest same-backbone baseline with macro- averaged exact match gains of 16.23%, top-performing API baseline 11.46%, and 3.04% for it’s teacher model with reduction of external model calls from 6.0 to 1.0. Further analysis suggests that learned verifiers and LLM judges can support candidate selection when deterministic checks are unavailable, although their ef- fectiveness varies across datasets.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Temporal Reasoning, agentic RL, test-time compute

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BYUhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Addressing the "strategy mismatch" problem, where different subtypes of temporal question answering (QA) require distinct reasoning procedures, this paper proposes a framework that incorporates graph-based construction, code generation, direct extraction, and hybrid reasoning. It trains a student policy through multi-step trajectory distillation (from a teacher model) and policy-conditioned reinforcement learning (RL).

Critical strengths

The paper identifies a clear, well-defined contribution: it highlights "strategy mismatch" as a core challenge, demonstrating through factorial experiments that different temporal sub-types favor specific reasoning procedures.

It unifies diverse reasoning procedures into a set of policy labels and distillable trajectories. By compressing multi-step teacher execution into direct generation by the student, the approach balances comprehensive reasoning capabilities with deployment efficiency.

Critical weaknesses

The approach relies heavily on predefined subtype mappings and teacher-based routing. Such fixed mappings underperform compared to a question-level "oracle," and the fact that strategy labels remain largely static after reinforcement learning suggests the student has failed to fully master dynamic, question-level strategy selection.

The comparison between the teacher-scaffolded approach and the distilled student model conflates changes in the model family with changes in the execution structure, making it impossible to isolate the causal effect of the distillation process. Furthermore, the failure of the "cold-start" control experiment leaves the value of structured strategy trajectories insufficiently demonstrated.

Key components, such as route-related weights, normalization factors, and reward subsets, are largely hand-tuned. The study lacks comprehensive ablation studies or sensitivity analyses for these reward components, and the evaluation of generated artifacts is limited to representations that can be reliably parsed.

Questions for authors

Please refer to the weaknesses.

Reviewer_4rqDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes StratTR, a multi-strategy framework for temporal reasoning. It assigns temporal question subtypes to four reasoning strategies, uses a teacher to generate multi-step reasoning trajectories, and distills these trajectories into a single-call 4B model through supervised fine-tuning. A GRPO stage is then employed based on strategy-conditioned rewards. Experiments across seven temporal reasoning benchmarks report improvements over the selected prompting baselines.

Critical strengths

The observation that different temporal reasoning tasks benefit from different reasoning strategies is reasonable, and the factorial experiment provides some empirical motivation for this direction.

Critical weaknesses
  1. The experiments do not convincingly validate the central methodological contributions. GRPO improves macro EM only from 74.12% to 74.92%, without adequate ablations of the individual reward components or comparisons with outcome-only RL.

  2. There are substantial inconsistencies and presentation problems that undermine confidence in the experimental findings. In Table 1, direct extraction outperforms graph construction for Explicit/Relative Reasoning, whereas Table 2 assigns this subtype to graph construction. Furthermore, the reported accuracy of the designed fixed mapping in Discussion (57.41%, line 460) does not appear consistent with the strategy assignments and per-subtype results in Tables 1 and 2.

Questions for authors
  1. Can the authors provide separate ablations of the strategy-conditioned reward components, including a comparison with answer-outcome-only GRPO?

  2. Could the authors explain the strategy assignments in Table 2 with the factorial results in Table 1 and the fixed-mapping analysis in Discussion more clearly?

Reviewer_Qw9zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors evaluate different LLM generation strategies (graph generation, code generation, etc.) on different temporal question types. They then design a simple routing system to send each question to a large LLM (Gemini-3.1-flash-lite) prompted with a question-type-specific generation strategy. This LLM-routing system is distilled via SFT and GRPO (with strategy-based rewards) into a smaller LLM (Qwen3-4B). The approach is evaluated on 5 temporal question answering datasets and compared to one prior work (TISER) and various ablations of the proposed approach.

Critical strengths
  • Demonstrates cleanly that different strategies (graph generation, code generation, etc.) are differentially effective for different types of temporal questions, and leverages this fact to design a simple but appropriate LLM-routing system.
  • The distilled one-call 4B LLM yields better performance than prior work (and its teacher LLM) on 4 of 5 temporal question answering datasets, and there is analysis of where/why RL leads to improvements over SFT alone in this distilled model.
Critical weaknesses
  • The process for going from a question to a question type, which the entire proposed approach depends on since question types define which strategy should be applied, is not well defined. We are told only that they "unified evaluation schemata so that each question is associated with one of nine subtypes" but not by what process. Was it manual? Automatic? If automatic, using what technology? This is critical for understanding whether the approach can only be applied to datasets where question types are known, or whether the approach will also work on new questions that have not been pre-annotated for a question type.
  • The end of the paper, especially from line 413 to line 472, is hard to follow. This is especially true of the "Exploratory reranking" section and its Tables 6 and 7. The judge's prompt is not defined. PRM, PAV, BoN, and "structural score" are not defined. Neither Table 6 nor Table 7 defines the units of the numbers (exact match scores? something else?) nor do they define which model/recipe is being run (broken reference in "the same-recipe SFT re-run in Table ??"). See below for many other questions about this unclear part of the paper.
Questions for authors
  • Where is "Appendix defines the remaining answer-attempt and date-precision terms and gives the complete route-specific objectives and weights.", "Detailed hyperparameters are reported in Appendix", and "Appendix lists the hyperparameters taken directly from the released training configurations"? I don't see any of these in the appendix.
  • Where do you "report descriptive macro averages for the six structural and three extraction subtypes, grouped by the majority strategy in the teacher trajectories."?
  • Where/why does the distilled model outperform its teacher?
  • For both Case 2 and Case 3 in the case study, you report that SFT and RL select the same strategies but end up with different labels. Why? What is different in their traces that causes this?
  • What do you mean when you say "Deterministic checks", "artifacts with encoded constraints", "outcome feedback", "Learned process verifiers", "deterministic verifier", "outcome-only routes" in the discussion? I don't see these defined, nor do I see how they relate to the proposed approach since, e.g., there's no constraints or verifiers in it.
  • How do you define the "oracle" on lines 461 and 473?

Minor writing suggestions (not questions):

  • 6.2: "Distillation reduces external model calls" (and its corresponding figure) is just repeating 5: "Distillation reduces multi-call execution to one call". This space could be better used for something else.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.