acceptodds
Under review as a conference paper at ICLR 2027

TUCH: Temporal Unfolding via Causal Hypotheses for LLM-Based Forecasting

Abstract

Forecasting requires estimating the probability of a future outcome from evidence that becomes available over time. Existing LLM-based forecasting methods remain largely temporally flat, organizing evidence without explicitly reasoning across successive stages of real-world time. We introduce temporal unfolding and instantiate it in TUCH, where causal hypotheses represent the model’s current estimate of the event’s causal state, are updated with new evidence, and guide subsequent retrieval. To provide step-level supervision, we construct reference trajectories from resolved forecasting questions and train the model with supervised fine-tuning and process-outcome reinforcement learning. Experiments on PROPHET and ForecastBench show that TUCH consistently outperforms same-backbone temporally flat baselines, while strong LLMs under the same temporal-unfolding scaffold show mixed performance relative to the trained TUCH models.

Primary area:
causal reasoning
Keywords:
Forecasting, temporal unfolding, causal hypotheses, causal reasoning, large language models

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_GkNjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TUCH, a temporal unfolding framework for LLM-based forecasting. Instead of aggregating retrieved evidence in a temporally flat manner, TUCH divides historical evidence into consecutive time intervals, maintains evolving causal hypotheses, and uses them to guide subsequent retrieval. The model is trained on hindsight-constructed reference trajectories through supervised fine-tuning and process-outcome reinforcement learning. Experiments on PROPHET and ForecastBench demonstrate improvements over several same-backbone forecasting baselines.

Critical strengths

TUCH provides a structured approach to forecasting by explicitly tracking evolving causal hypotheses across temporal stages. The authors make reasonable efforts to prevent information leakage, including separating training and evaluation periods, constructing evaluation evidence without access to outcomes, and manually validating synthesized trajectories. The experimental results are promising, with TUCH-RL achieving Brier scores of 21.70 and 17.19 on PROPHET and ForecastBench, respectively. The component-wise ablations also suggest that hypothesis memory, hypothesis-guided retrieval, and process rewards contribute to the reported improvements.

Critical weaknesses

1. The novelty and effectiveness of the central temporal unfolding mechanism are not sufficiently established.

Prior studies have already explored maintaining and updating linguistic belief states during sequential reasoning [1, 2]. While TUCH introduces hindsight-supervised causal-hypothesis trajectories, its contribution relative to these existing approaches could be more clearly distinguished.

More importantly, the experiments do not directly isolate the benefit of temporal partitioning. All ablations in Table 2 retain the same monthly temporal scaffold (T = 12), examining only hypothesis memory, retrieval, and process rewards. Without a temporally flat variant using identical evidence, supervision, and training configurations, it remains unclear whether the improvements arise from temporal unfolding itself or from additional supervision and structured retrieval.

2. Several experimental claims are not sufficiently supported by the reported comparisons.

In Table 1, the untrained Gemini 2.0 Flash achieves a Brier score of 21.13 on PROPHET, outperforming TUCH-RL at 21.70. This result complicates the claim that the temporal scaffold alone is insufficient, yet the discussion focuses primarily on models that TUCH outperforms.

Furthermore, the claimed best overall performance on ForecastBench is evaluated on a self-constructed subset of 274 questions rather than through the official benchmark protocol. The comparison omits important reference points, including simple base-rate forecasts, crowd or expert predictions, and prediction-market consensus. Given the importance of human forecasting baselines in ForecastBench [3], these omissions limit the interpretation of the reported performance.

3. The comparison with existing methods does not adequately control computational and retrieval budgets.

TUCH retrieves up to 60 facts across each of 11 temporal perception steps, resulting in approximately 660 evidence slots per question. However, the corresponding retrieval budgets, search rounds, and training costs of the baselines are not clearly reported. Since Search-R1 already employs multi-round retrieval and reinforcement learning, differences in computational resources may contribute to the observed performance gap.

Additionally, a contemporaneous study [4] reports a Brier score of approximately 0.18 using retrieval-augmented 7–8B models on a different set of 500 resolved ForecastBench questions. Although these results are not directly comparable, they further motivate stronger and resource-matched baselines.

References

[1] ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction.

[2] Agentic Forecasting Using Sequential Bayesian Updating of Linguistic Beliefs.

[3] ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities.

[4] The Corroboration Illusion: When More News Makes LLM Forecasts Less True.

Questions for authors
  1. Can the authors provide a temporally flat TUCH variant with identical evidence, supervision, and training configurations, along with an untrained Qwen2.5-7B baseline under the same temporal scaffold? These controls would help isolate the contributions of temporal partitioning and process learning.

  2. Can the authors report the retrieval budgets, search rounds, and training/inference costs for all baselines, and provide resource-matched comparisons with Search-R1?

  3. How do the authors interpret Gemini 2.0 Flash outperforming TUCH-RL on PROPHET? Can they also include base-rate, human, and prediction-market reference forecasts on the same ForecastBench subset to better contextualize the claimed improvements?

3 comments
AuthorsOpenReview ↗

Rebuttal for Weakness 1 and Question 1

Thank you for highlighting these highly relevant studies. We appreciate the opportunity to clarify TUCH’s distinctions from prior linguistic belief-state methods and present additional experimental evidence for its temporal-unfolding process.

Differences from prior linguistic belief-state methods.

  1. Temporal unfolding follows real-world event development.

ABBEL [1] updates beliefs across environment interactions, while BLF [2] updates a structured belief state through iterative tool use under a forecast-date cutoff for leakage control. In TUCH, perception steps correspond to consecutive real-world intervals (Section 3.2). The current causal hypotheses guide retrieval from the next temporal interval, and the evidence and hypotheses are retained in an ordered core memory. Thus, state evolution and evidence acquisition progress jointly along the event’s real-world timeline, rather than only across interaction or search rounds.

  1. Supervision targets chronological causal-hypothesis trajectories.

ABBEL trains belief updates to preserve task-relevant information while improving memory efficiency. TUCH instead constructs forecasting-specific reference trajectories from resolved questions, using candidate upstream mechanisms and temporally organized evidence to synthesize step-level causal hypotheses (Section 3.3). These trajectories supervise perception-step hypothesis generation during SFT, followed by process-outcome RL for the unfolding process (Section 3.4). During RL and inference, hypothesis generation and retrieval are fully self-conditioned.

The contribution is therefore the joint formulation and learning of causal-state evolution and evidence acquisition along real-world time, extending linguistic belief updating to an explicitly supervised temporal-unfolding process for forecasting.

Isolating the benefit of temporal partitioning.

We have conducted a chronological-order ablation of TUCH-SFT to examine the contribution of chronological organization under matched training and inference conditions.

Experimental setup.

We retrain TUCH-SFT with the chronological order of its perception steps shuffled, starting from the same Qwen2.5-7B-Instruct backbone and using the same SFT configuration and number of optimization steps. We reuse the same step-level evidence and reference causal hypotheses and retain the same teacher and inference budgets.

We keep the first perception step fixed, preserving its question-based retrieval and initial hypothesis generation. We randomly shuffle the remaining perception steps using a fixed permutation for each example, keeping the retrieved evidence and reference causal hypotheses within each step together. The final judgment step remains last, conditions on the resulting shuffled core memory, and retains the same reference judgment target. This preserves the overall information content, the number of perception steps, and the evidence–hypothesis pairing within each step.

Results.

Brier scores ×100, where lower is better:

MethodPROPHET ↓ForecastBench ↓
TUCH-SFT24.8019.37
Shuffled chronological order29.3424.54

Shuffling increases the Brier score by 4.54 points on PROPHET and 5.17 points on ForecastBench. Under these matched conditions, chronological processing outperforms the shuffled-order sequential variant on both datasets. This within-TUCH comparison provides direct evidence for the contribution of chronological organization beyond performing the same number of sequential reasoning steps.

Untrained Qwen2.5-7B-Instruct under the same temporal scaffold.

We provide the results of the untrained Qwen2.5-7B-Instruct evaluated under the exact same temporal scaffold.

The results are as follows (Brier ×100; lower is better):

MethodPROPHET ↓ForecastBench ↓
Qwen2.5-7B-Instruct (under the temporal-unfolding scaffold)41.3739.31
TUCH-SFT24.8019.37

This comparison supports the contribution of learning the temporal-unfolding process beyond applying the scaffold alone.

We will add results to Table 1 and Table 2 in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal for Weakness 2 and Question 3

Thank you for raising these concerns and giving us the opportunity to clarify our results. We address three points: Gemini’s performance, the scope of our ForecastBench evaluation, and comparisons with reference forecasts.

Gemini 2.0 Flash on PROPHET.

Gemini 2.0 Flash achieves a lower Brier score than TUCH-RL on PROPHET (21.13 vs. 21.70). TUCH-RL nevertheless outperforms most evaluated models, including the compared open-source models. Separately, our same-backbone comparison shows that TUCH-SFT reduces the Brier score of Qwen2.5-7B-Instruct from 41.37 to 24.80 under the same temporal scaffold, supporting the benefit of TUCH-specific training for this backbone. We will revise the abstract and Section 5.1 to explicitly acknowledge the Gemini result and qualify the corresponding claims.

Scope of the ForecastBench evaluation.

Our evaluation uses a subset of 274 resolved binary questions and a separately constructed evidence corpus. The reported results therefore apply to this evaluation setting. We will label it “ForecastBench (274-question subset)” and clarify that the results are obtained under our evaluation setup, rather than the official ForecastBench protocol.

Reference forecasts contextualize the reported improvements.

We have computed the following reference scores on our 274-question subset (Brier ×100; lower is better):

MethodForecastBench (274-question subset) ↓
Base rate21.39
Human crowd10.88
Prediction-market consensus7.47
TUCH-SFT19.37
TUCH-RL17.19

We will include this table in the appendix.

AuthorsOpenReview ↗

Rebuttal for Weakness 3 and Question 2

We appreciate the reviewer's careful attention to the fairness of the computational comparison. In response, we detail the baseline budgets and costs, provide a resource-matched chronological-order ablation, and contextualize the results from [4] to address these concerns.

Baseline configurations and computational costs are explicitly reported below.

We reproduce Search-R1 and OpenForecaster following their original methods, with the same Qwen2.5-7B-Instruct backbone and training questions as TUCH. Their retrieval settings follow the original papers: Search-R1 retrieves the top 3 passages per search call with a maximum action budget of 4; OpenForecaster uses a single retrieval of 5 chunks, each containing 512 tokens, during evaluation.

Training, where applicable, and inference are performed on 8 NVIDIA H200 GPUs.

MethodTraining epochsTraining timeH200 GPU-hoursInference GPU·s/question
Search-R155 h 56 min47.470.22
OpenForecaster145 min6.000.55
Vanilla RAG——0.000.46
TUCH-SFT263 min8.402.30

Vanilla RAG requires no task-specific training. TUCH-SFT outperforms Search-R1 on both datasets while requiring substantially less model-training compute (8.40 vs. 47.47 GPU-hours). Its advantage therefore cannot be attributed to a larger model-training budget. TUCH-SFT incurs higher inference costs, reflecting a trade-off between forecasting performance and inference cost.

Chronological organization contributes under fixed evidence and resource budgets.

Our chronological-order control preserves the backbone, step-level evidence and reference hypotheses, SFT configuration, and teacher and inference budgets, while shuffling the perception steps after the first. This increases Brier scores from 24.80 to 29.34 on PROPHET and from 19.37 to 24.54 on ForecastBench.

The performance difference in this controlled comparison cannot be explained by different resource budgets. It provides evidence that how evidence is organized along real-world time contributes to forecasting performance, beyond the amount of evidence and computation alone. Together with the lower training cost relative to Search-R1, this supports the framework’s contribution while making its inference-cost trade-off explicit. We will include the resource settings and cost table in the appendix.

The result in [4] requires its own baseline context.

Table II in [4] reports a Brier score of 0.184 for calibrated RAG, compared with 0.127 for its constant base-rate reference. On our subset, TUCH-SFT and TUCH-RL achieve 0.1937 and 0.1719, respectively, compared with a base-rate reference of 0.2139. These differences illustrate why raw Brier scores across different subsets do not establish a ranking between methods.

Reviewer_WM4YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework in which the model reads news in their actual chronological order, continuously updates hypotheses about how events might unfold, and uses these hypotheses to guide subsequent retrieval.

Critical strengths
  1. The proposed method goes beyond simply combining a series of prompts.
  2. The overall framework forms a reasonably complete closed loop.
Critical weaknesses
  1. The underlying intuition behind the method appears questionable in some respects.
  2. The overall framework does not demonstrate rigorous causal reasoning.
  3. The baselines appear to be relatively weak.
Questions for authors

1. Why do the authors set the evidence cutoff time to τ=d−30\tau = d - 30 days? Is it reasonable to assume that evidence updates for all events occur on a monthly timescale?

2. I have a major concern regarding the underlying intuition of the proposed framework. In principle, more recent evidence should generally carry greater weight. However, the proposed reasoning chain seems questionable in this regard, as older evidence introduced at earlier stages continues to influence subsequent reasoning. What happens if conclusions drawn from earlier evidence are contradicted by newly available evidence? Would the model not ultimately have to rely on the newer evidence anyway? If so, how is this fundamentally different from directly considering the latest evidence in a temporally flat manner? It seems that the proposed design could introduce substantial reasoning redundancy, and its underlying intuition is not entirely convincing.

3. How exactly is the proposed framework related to causality? The use of causal terminology appears to be largely superficial, as the method seems to capture correlations or reasoning dependencies rather than genuine causal relationships.

4. The current ablation experiments do not seem to address an important question: If all other components are kept unchanged and only the actual chronological ordering of evidence is removed, would the performance still deteriorate?

5. For binary prediction, consistently outputting p=0.5p = 0.5 yields a score of 25 regardless of the ground-truth outcome. This seems to suggest that the vast majority of temporally flat baselines fail to outperform even a constant 50% prediction baseline on both datasets. If this is indeed the case, part of the reported substantial improvement may simply be attributable to the weakness of the comparison baselines.

6 comments
AuthorsOpenReview ↗

Rebuttal for Question 1

Thank you for asking us to clarify these points.

Evidence cutoff τ=d−30\tau=d-30 days.

We set the evidence cutoff to τ=d−30\tau=d-30 days to prevent outcome leakage. This is consistent with OpenForecaster (Chandak et al., 2026), which restricts retrieval to news published no later than one month before each question’s resolution date to avoid leakage. Following this precedent, we instantiate the cutoff as a fixed 30-day gap for our retrospective forecasting evaluation.

Monthly timescale assumption.

We do not assume that evidence updates for all events occur on a monthly timescale. Monthly granularity specifies how we partition the evidence history and update the model’s hypotheses; it does not imply that evidence arrives only once per month. As stated in Section 3.3, we use T=12T=12 at monthly granularity, which covers more than 99% of the forecasting horizons in our training and evaluation data. This yields T−1=11T-1=11 perception steps followed by a final judgment step. The choice provides a practical balance between temporal resolution and computational cost. All methods using the temporal-unfolding scaffold share the same temporal partitions and evidence budget, keeping these aspects of the comparison controlled.

We will add this justification to Section 3.1.

AuthorsOpenReview ↗

Rebuttal for Question 2

Thank you for raising this concern. We address two aspects: why earlier evidence remains useful, and how the framework handles contradictions introduced by newer evidence.

Why retain earlier evidence?

We agree that more recent evidence should generally carry greater weight when assessing the current state of an event. However, this does not mean that earlier evidence becomes unimportant. Earlier evidence may remain relevant to persistent conditions and longer-term developments, and provide the context needed to interpret newer evidence. The importance of retaining this context varies across forecasting tasks, which in our datasets span diverse categories including politics, geopolitics, cryptocurrency, sports, and entertainment.

This is particularly important because our data spans highly heterogeneous categories. In structural domains (e.g., crypto, macro), early commitments often remain binding for months. In episodic domains (e.g., sports), the latest update (e.g., an injury) is only interpretable against the earlier baseline of team strength. For example, assessing whether inflation is following a sustained downward trajectory can require examining a sequence of releases over several months; the latest reading alone may not distinguish a persistent trend from a temporary fluctuation. In cryptocurrency forecasting, a previously announced token-unlock schedule may remain useful to future supply for months, even when the latest news concerns short-term trading activity. A recent report may summarize some historical context, but it does not necessarily contain the full set of earlier observations or commitments useful to the forecast.

What happens when newer evidence contradicts an earlier conclusion?

At each perception step, TUCH generates its current causal hypotheses from the newly retrieved evidence and the accumulated core memory. The model can therefore revise or reject an earlier causal hypothesis when contradictory evidence becomes available.

Retaining an earlier causal hypothesis does not impose a requirement that later reasoning remain consistent with it. The final judgment conditions on the complete core memory, including the contradictory evidence and subsequent causal hypotheses. This core memory provides the context for understanding what changed and which earlier assumptions remain applicable.

Unlike temporally flat approach that sees only the latest evidence pool, TUCH maintains an evolving causal state represented by causal hypotheses. These hypotheses are updated using newly retrieved evidence and accumulated core memory, allowing earlier assessments to be revised and guiding subsequent retrieval.

AuthorsOpenReview ↗

Rebuttal for Question 3

Thank you for the question. We address the concerns about correlations, reasoning dependencies, and the role of causal terminology separately.

Causal hypotheses about upstream mechanisms, rather than mere correlations.

The reference hypotheses are constructed around proposed mechanisms relevant to the forecasting outcome. In Section 3.3.1, the generator identifies candidate upstream mechanisms, links them into causal chains, and converts them into retrieval queries. These queries guide both news retrieval and hypothesis synthesis, favoring upstream and state-tracking evidence over direct outcome confirmation. Thus, the construction process explicitly targets hypotheses about how developments may influence an outcome, rather than merely collecting co-occurring information.

The hypotheses concern real-world event development, rather than only computational reasoning steps.

As described in Section 1, TUCH accounts for how earlier conditions develop into later events. Section 3.2 organizes this process into consecutive real-world intervals. Each hypothesis contains a prediction statement and an evidence-grounded explanation, and the current hypothesis set explicitly represents the model’s causal state. These hypotheses are updated with newly available evidence and persist across temporal steps, tracking the model’s evolving assessment of the event and its possible future developments.

Causal hypotheses play an active role in training and evidence acquisition rather than serving as superficial labels.

Reference hypotheses provide step-level supervision during SFT. During RL and inference, the model generates its own hypotheses in a fully self-conditioned process, with their prediction statements guiding retrieval from the next temporal interval (Eq. 3). The evidence and hypotheses are retained in core memory for subsequent updates and the final judgment. These hypotheses therefore influence what evidence the model acquires and how its forecast evolves, rather than serving as post-hoc labels attached to a completed prediction.

The framework’s connection to causality lies in constructing evidence-grounded hypotheses about candidate upstream mechanisms, updating them as evidence accumulates, and using them to guide subsequent evidence acquisition and forecasting.

Reviewer_3wMTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TUCH processes evidence chronologically, maintains natural-language causal hypotheses, and uses them to guide later retrieval. It learns this process through supervised training on hindsight-generated trajectories and reinforcement learning with outcome and process rewards. The evaluation reports forecasting gains and includes retrained memory/retrieval ablations and an outcome-only RL comparison.

Critical strengths
  • The connection between temporal memory and hypothesis-guided retrieval is well motivated. The paper also clearly distinguishes teacher-conditioned training from self-conditioned inference, which helps explain how the method is intended to work (pp. 4–6).
  • Retrained memory/retrieval ablations and the outcome-only RL comparison provide useful evidence for these components, alongside substantial reported forecasting gains (Tables 1–2).
Critical weaknesses
  • Every component ablation retains temporal partitioning, while external baselines use different supervision and objectives. The results therefore do not isolate the benefit of chronological organization, which is central to the paper’s contribution (p. 7; Table 2).
  • Important details needed to interpret the temporal reasoning are unresolved. The process reward’s matching and score-combination rules are not fully specified. Figure 2 also includes December evidence despite a November 20 cutoff under Section 3.1. The latter may be an illustration error, but needs clarification; it does not by itself establish leakage in the experiments (p. 6; Appendix B).
Questions for authors
  1. Does chronological processing outperform a non-temporal sequential control with the same backbone, available information, teacher budget, and inference budget? Both versions should be trained on targets appropriate to the evidence they receive.
  2. What are the exact matching and score-combination rules for the process reward? How does it handle a hypothesis that is supported by the evidence but differs from the teacher’s reference?
  3. Is Figure 2 a schematic illustration or an actual trace? How does the December information fit the stated cutoff?
2 comments
AuthorsOpenReview ↗

Rebuttal for Weakness 1 and Question 1

Thank you for highlighting the need to distinguish the benefit of chronological organization from differences in supervision and objectives across external baselines. We have added a chronological-order ablation of TUCH-SFT.

We retrain TUCH-SFT with the chronological order of its subsequent perception steps shuffled. Training starts from the same Qwen2.5-7B-Instruct backbone and uses the same SFT configuration, training data, hyperparameters, and number of optimization steps as the original TUCH-SFT model. The control retains the same teacher and inference budgets. We keep the first perception step fixed, preserving its question-based retrieval and initial hypothesis generation, and randomly shuffle the remaining perception steps, t=2,…,T−1t=2,\ldots,T-1, using a fixed random permutation per example. The retrieved evidence and causal hypotheses within each step are kept together, so each step retains its own evidence and hypothesis target and thus receives supervision appropriate to its evidence. The final judgment step remains at t=Tt=T conditions on the resulting shuffled history, and uses the same final-outcome target. This changes only the temporal order of the subsequent perception steps while preserving the overall information content, the number of steps, and the evidence–hypothesis pairing within each step.

The results are as follows (Brier ×100; lower is better):

MethodPROPHET ↓ForecastBench ↓
TUCH-SFT24.8019.37
Shuffled chronological order29.3424.54

Shuffling increases the Brier score by 4.54 points on PROPHET and 5.17 points on ForecastBench. With the backbone, available information, supervision targets, teacher budget, inference budget, and objectives held fixed, the chronological variant achieves lower Brier scores than the non-temporal sequential control on both datasets. We believe this within-TUCH controlled comparison provides the evidence for the benefit of chronological organization.

AuthorsOpenReview ↗

Rebuttal for Weakness 2 and Question 2&Question 3

Thank you for the opportunity to clarify the details of the process reward and the temporal information presented in Figure 2.

Process reward: matching and score combination Process reward.

At perception step tt, let xt,ii=1Mt\\{x_{t,i}\\}_{i=1}^{M_t} and yt,jj=1Nt\\{y_{t,j}\\}_{j=1}^{N_t} denote the textual representations used to match the generated and reference causal hypotheses, respectively. Using the frozen, ℓ2\ell_2-normalized retrieval encoder ϕ\phi, we compute the cosine-similarity matrix:

St,ij=ϕ(xt,i)⊤ϕ(yt,j).S_{t,ij}=\phi(x_{t,i})^\top\phi(y_{t,j}).

Given a similarity threshold δ∈[0,1]\delta\in[0,1], we first construct the set of admissible matches:

Et=(i,j):St,ij>δ.E_t=\\{(i,j):S_{t,ij} > \delta\\}.

We then select a one-to-one matching Vt⊆EtV_t\subseteq E_t that first maximizes the number of valid matches ∣Vt∣|V_t| and, among all maximum-cardinality matchings, maximizes the total similarity ∑(i,j)∈VtSt,ij\sum_{(i,j)\in V_t}S_{t,ij}. Each generated hypothesis and each reference hypothesis can participate in at most one match.

If Vt=∅V_t=\emptyset, including cases where either hypothesis set is empty, we set the step reward rt=0r_t=0. Otherwise, matching precision, recall, and quality are

\qquad \mathrm{Rec}_t=\frac{|V_t|}{N_t}, \qquad Q_t=\frac{1}{|V_t|}\sum_{(i,j)\in V_t}S_{t,ij}.$$ The step reward is the product of F1-based coverage and matching quality: $$F_{1,t}=\frac{2\\,\mathrm{Prec}_t\\,\mathrm{Rec}_t}{\mathrm{Prec}_t+\mathrm{Rec}_t}, \qquad r_t=F_{1,t}Q_t.$$ The F1 term balances reference coverage against the number of unmatched generated hypotheses, while the quality term rewards stronger semantic alignment. Let $\mathcal{T}$ denote the fixed set of scheduled perception steps, with $|\mathcal{T}|=11$ in our experiments. Missing or truncated steps receive $r_t=0$ and remain included in $\mathcal{T}$. The process reward is $$R_{\mathrm{pro}}=\frac{1}{|\mathcal{T}|}\sum_{t\in\mathcal{T}}r_t.$$ **Hypotheses that differ from the teacher’s reference** The process reward uses semantic matching against the reference trajectory rather than exact string matching. A differently phrased hypothesis can contribute to the process reward if it is matched to a reference hypothesis through the optimal one-to-one assignment. A semantically distant hypothesis may receive a lower similarity score or remain unmatched. The process reward itself does not independently verify evidence support. It measures alignment with the reference trajectory. An evidence-supported hypothesis that is semantically distant from the reference may therefore receive low or no process credit. However, we also optimize the outcome reward, which evaluates the final prediction using the negative Brier score. A rollout containing such a hypothesis can still receive a higher outcome reward if its final prediction achieves a lower Brier score, even when the hypothesis does not closely match the reference. **Clarification of Figure 2** Figure 2 presents an illustrative example of the TUCH framework. The “December 2024” phrase appears in the Causal Hypotheses box and expresses a prediction about future developments, rather than describing a retrieved December evidence. Causal hypotheses refer to future periods, whereas the evidence cutoff applies to the publication dates of retrieved evidence. We will clarify this distinction in the figure caption.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Acknowledgments and Rebuttal Summary

We sincerely thank all reviewers for their time and constructive feedback. Your comments helped us identify where our explanations and experimental evidence needed to be clearer.

In response, we have conducted a chronological-order ablation that directly tests the contribution of chronological organization while holding evidence, supervision, and computational budgets fixed. We also provide additional results and clarify methodological and evaluation details in our responses below. We will incorporate these additions and clarifications into the revised manuscript and appendix. We hope our responses address the concerns raised and welcome any further questions or suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.