PRInTS Framework for Long-Horizon Reward Modeling

• PRInTS is a generative reward model that provides dense, multi-dimensional scoring to assess each reasoning step’s information gain in complex tasks.
• It employs recursive trajectory summarization to compress evolving histories into focused representations for improved context management.
• Empirical results show significant performance gains on benchmarks like FRAMES, GAIA, and WebWalkerQA using efficient, smaller backbone agents.

PRInTS (Process Rewarding Information-seeking Trajectory Summarization) is a generative process reward model (PRM) architecture for long-horizon information-seeking tasks, designed to guide agents that interact with complex toolchains—such as search engines, browsers, and code interpreters—across multi-step reasoning trajectories. PRInTS addresses limitations of conventional PRMs by providing (1) dense, multi-dimensional scoring of tool-augmented reasoning steps and (2) dynamic trajectory summarization that compresses evolving histories. These capabilities enable competitive or superior performance to larger specialized agents and existing reward modeling baselines on benchmarks such as FRAMES, GAIA (Levels 1–3), and WebWalkerQA, using much smaller backbone agents (Lee et al., 24 Nov 2025).

1. Challenges in Long-Horizon Information-Seeking

Traditional PRMs are optimized for isolated or short reasoning chains, such as brief mathematical or logical inferences, typically producing binary outputs that mark steps as “valid” or “invalid.” This binary approach is insufficient for complex, long-horizon tasks, where:

• Interaction with Tools: Reasoning must interleave with diverse tool calls. Step quality depends on multifaceted considerations including informativeness of the tool call, accuracy and depth of tool output interpretation, and the foresight in subsequent planning.
• Accumulation of Context: The full trajectory $H_t = (s_1, a_1, o_1, \ldots, s_t, a_t, o_t)$ rapidly grows as each action yields potentially lengthy outputs. LLM-based PRMs are unable to effectively attend over lengthy or noisy histories, resulting in reduced evaluation granularity and unnecessary computational overhead.

PRInTS is developed to overcome these granular and contextual limitations.

2. Core Framework and Model Architecture

PRInTS unifies two principal capabilities via a generative PRM $f_\theta$:

• Dense Step-Level Scoring: Each reasoning step (state-action pair $(s_t, a_t)$) is assessed as an information-gain event. The model estimates to what degree the step increases the probability of reaching a correct final answer, across multiple quality dimensions.
• Trajectory Summarization: PRInTS recursively summarizes the growing history $H_t$ into a compact representation $h_t$, ensuring context is bounded and focused on salient information.

2.1 Information-Gain-Based Scoring

For each step, $g_t$ is computed as the marginal increase in success probability, estimated through Monte Carlo rollouts from the current state $(q, H_{t-1}, s_t, a_t)$. If $M$ rollouts yield $m_t$ as the fraction that completes with a correct answer $a^*$, the marginal gain is

$f_\theta$0

PRInTS then generates both a chain-of-thought (CoT) analysis and a scalar score prediction:

$f_\theta$1

In practice, $f_\theta$2 emits a vector $f_\theta$3 corresponding to per-dimension scores (e.g., informativeness, interpretation correctness, planning quality), which are linearly aggregated:

$f_\theta$4

with learnable parameters $f_\theta$5 and $f_\theta$6.

2.2 Recursive Summarization

To prevent excessive context growth, PRInTS maintains a summary $f_\theta$7 at each step:

$f_\theta$8

$f_\theta$9 is trained to filter irrelevant context and retain key facts and planning steps, emitting $(s_t, a_t)$0 as a generative sequence.

3. Training Procedures and Optimization

PRInTS is trained with two alternating objectives:

• Summarization Loss: Given annotated “gold” summaries $(s_t, a_t)$1, the summarizer $(s_t, a_t)$2 optimizes standard sequence cross-entropy loss

$(s_t, a_t)$3

• Scoring Rewards and RL Loss: The scalar scorer $(s_t, a_t)$4 is trained with a hybrid reward comprising:
  • Score Reward $(s_t, a_t)$5: Measures absolute deviation from ground-truth gain.
  • Comparison Reward $(s_t, a_t)$6: Evaluates pairwise ranking correctness versus other candidate steps.
  • Adaptive Weighting $(s_t, a_t)$7: Down-weights noisy or low-margin score pairs.

The full reward is $(s_t, a_t)$8. PRInTS utilizes Generative Reward Policy Optimization (GRPO), treating the scorer as a sequence generating policy $(s_t, a_t)$9 and optimizing expected reward using policy gradients:

$H_t$0

where $H_t$1 is the concatenated CoT analysis and scalar score.

4. Best-of-n Sampling Policy

At test time, PRInTS employs a best-of-$H_t$2 selection mechanism to improve agent performance:

1. For each of $H_t$3 samples at step $H_t$4, the agent proposes $H_t$5 and executes the corresponding tool call, yielding $H_t$6.
2. PRInTS computes a score for each candidate step.
3. The step with the highest score is selected to advance the trajectory.
4. The trajectory summary is updated.
5. This process repeats until a terminal state (final answer) is reached.

This approach empirically demonstrates that increasing $H_t$7 improves performance up to an optimal range, after which gains plateau or over-exploration degrades results.

5. Empirical Evaluation and Ablations

PRInTS is evaluated on FRAMES (300 factual and multi-step retrieval questions), GAIA Levels 1–3 (103 per level), and WebWalkerQA (easy, medium, hard subsets; 247 total), using a variety of base agents including Qwen3-32B, Tongyi DeepResearch-30B-A3B, and Gemini-2.5-Flash. Performance is measured via Avg@3 accuracy using GPT-5 as the LLM-as-Judge.

Method	FR	GAIA L1	GAIA L2	GAIA L3	Web E	Web M	Web H	Avg (Qwen3-32B)
Base agent	49.3	35.1	23.7	11.1	30.1	26.9	30.3	29.5
Confidence	55.7	36.8	24.4	16.7	31.7	31.3	32.9	32.8
Relevance	56.3	34.2	20.5	8.3	33.3	29.5	32.5	30.7
GenPRM-7B	50.0	32.5	25.7	16.7	33.3	32.8	34.6	32.2
Web-Shepherd-8B	49.0	38.5	23.7	5.5	28.5	31.8	33.3	30.0
StepWiser	51.3	37.6	22.4	8.3	31.7	31.8	33.8	31.0
PRInTS	58.7	49.6	33.3	19.4	39.8	33.3	37.3	38.8

These results show consistent and substantial improvements, with PRInTS providing gains of +3–14 percentage points over base agents and +2–6 over the strongest existing PRMs.

Additional ablations reveal that:

• Summarized contexts outperform raw histories or recent step windows (47.2% vs. 44.1% and 39.5% average accuracy).
• Reward components $H_t$8 and $H_t$9 are complementary; jointly weighted training yields higher accuracy (up to 47.2%).
• Increasing $h_t$0 in test-time sampling improves performance up to $h_t$1 (+8.9pp), but larger $h_t$2 causes over-exploration.
• Using only 50% of annotated pairs preserves over 90% of full-data gain, indicating sample efficiency.

6. Limitations and Open Directions

PRInTS exhibits several open challenges and areas for future research:

• Over-exploration at large $h_t$3 can cause agents to indefinitely defer providing final answers. Integration of cost-of-action regularization is a suggested avenue.
• Summaries may lose rare but critical details; integration of richer memory architectures or contrastive objectives may alleviate this.
• Current scalar scoring could be extended to per-dimension attention maps or multi-modal tool outputs (e.g., tables, images).
• Robustness to adversarial preference annotations and joint modeling of both partial-step and end-to-end outcome rewards are identified as promising directions.

A plausible implication is that hybrid reward modeling approaches incorporating both process supervision and end-state evaluation may further enhance long-horizon performance in tool-augmented LLM agents.

7. Significance and Comparative Advantages

By producing dense, dimension-aware information-gain scores and compact trajectory summaries, PRInTS addresses critical granularity and context-management bottlenecks in long-horizon information seeking. This enables consistent, state-of-the-art performance—matching or surpassing specialized or larger-scale agents—while maintaining efficient model size and sampling efficiency. These results demonstrate the impact of unified, generative reward modeling in advancing the tool-augmented capabilities of LLM-backed agents (Lee et al., 24 Nov 2025).