PaperlayerView exact revision on arXiv

Source-linked AI summary

KARL: Knowledge Agents via Reinforcement Learning

Jonathan D. Chang, Andrew Drozdov, Shubham Toshniwal, Owen Oertell, Alexander Trott, Jacob Portes, Abhay Gupta, Pallavi Koppol, Ashutosh Baheti, Sean Kulinski, Ivan Zhou, Irene Dea, Krista Opsahl-Ong, Simon Favreau-Lessard, Sean Owen, Jose Javier Gonzalez Ortiz, Arnav Singhvi, Xabi Andrade, Cindy Wang, Kartik Sreenivasan, Sam Havens, Jialu Liu, Peyton DeNiro, Wen Sun, Michael Bendersky, Jonathan Frankle

arXiv:2603.05218v1cs.AIcs.LG

TL;DR

Knowledge agents need to retrieve and reason over proprietary evidence across diverse, hard-to-verify search tasks, but existing evaluations cover only limited grounded-reasoning behaviors. KARL addresses this with KARLBench, agentic synthetic-data generation, and iterative large-batch off-policy RL. It reports Pareto-optimal cost-quality and latency-quality trade-offs against Claude 4.6 and GPT 5.2, with sufficient test-time compute surpassing the strongest closed models.

  • Problem

    Existing grounded-reasoning work provides limited evidence across diverse enterprise search behaviors and often relies on public knowledge or narrow benchmarks.

  • Method

    KARL combines the six-regime KARLBench, vector-search-based agentic synthesis with iterative bootstrapping, and iterative large-batch off-policy RL for multi-task training.

  • Results

    KARL is Pareto-optimal against Claude 4.6 and GPT 5.2 across KARLBench cost-quality and latency-quality trade-offs, and sufficient test-time compute exceeds the strongest closed models.

  • Takeaways & Limitations

    Training across heterogeneous search behaviors yields better generalization than single-benchmark optimization, supporting cost-efficient grounded reasoning across diverse tasks.

  • Takeaways & Limitations

    KARL improves retrieval strategy but remains limited in post-retrieval numerical computation, with some trajectories ending prematurely under arithmetic difficulty.

Abstract

from arXiv · show

We present a system for training enterprise search agents via reinforcement learning that achieves state-of-the-art performance across a diverse suite of hard-to-verify agentic search tasks. Our work makes four core contributions. First, we introduce KARLBench, a multi-capability evaluation suite spanning six distinct search regimes, including constraint-driven entity search, cross-document report synthesis, tabular numerical reasoning, exhaustive entity retrieval, procedural reasoning over technical documentation, and fact aggregation over internal enterprise notes. Second, we show that models trained across heterogeneous search behaviors generalize substantially better than those optimized for any single benchmark. Third, we develop an agentic synthesis pipeline that employs long-horizon reasoning and tool use to generate diverse, grounded, and high-quality training data, with iterative bootstrapping from increasingly capable models. Fourth, we propose a new post-training paradigm based on iterative large-batch off-policy RL that is sample efficient, robust to train-inference engine discrepancies, and naturally extends to multi-task training with out-of-distribution generalization. Compared to Claude 4.6 and GPT 5.2, KARL is Pareto-optimal on KARLBench across cost-quality and latency-quality trade-offs, including tasks that were out-of-distribution during training. With sufficient test-time compute, it surpasses the strongest closed models. These results show that tailored synthetic data in combination with multi-task reinforcement learning enables cost-efficient and high-performing knowledge agents for grounded reasoning.

1 Introduction

KARL targets grounded reasoning over proprietary enterprise data, where existing work provides limited evidence across diverse search behaviors. It combines heterogeneous benchmark evaluation, agentic data synthesis, and iterative off-policy reinforcement learning to improve generalization and efficiency.

  • Motivation: Grounded reasoning combines multi-step information gathering with complex reasoning over evidence outside model parameters, including proprietary enterprise data.The paper highlights applications in finance, law, medicine, and manufacturing.
  • Motivation: Existing deep-research results rely on public knowledge and black-box web search, leaving cross-task generalization to other grounded reasoning settings unclear.The paper identifies a relative lack of study at the grounded reasoning frontier.
  • Contributions: KARLBench spans six search regimes, and models trained across heterogeneous behaviors generalize better than models optimized for a single benchmark.The regimes include entity search, report synthesis, tabular reasoning, exhaustive retrieval, procedural reasoning, and enterprise-note fact aggregation.
  • Contributions: The agentic synthesis pipeline dynamically explores corpora with vector search to create grounded training pairs and bootstraps improved agents for further data generation.The synthesis recipe is reported to generalize across TREC-Biogen and BrowseComp-Plus.
  • Contributions: Iterative large-batch off-policy RL is designed to tolerate trainer–inference-engine discrepancies and extends to multi-task training by combining task losses.The approach avoids several stabilization heuristics associated with online GRPO and supports out-of-distribution evaluation.
  • Results: KARL is Pareto-optimal against Claude 4.6 and GPT 5.2 across KARLBench cost-quality and latency-quality trade-offs, and sufficient test-time compute exceeds the strongest closed models.The result is reported across grounded reasoning tasks, including out-of-distribution tasks.

2 KARLBench

KARLBench evaluates knowledge agents across six structurally distinct grounded-search capabilities using controlled, minimally transformed corpora. Its evaluation unifies answer quality while preserving task-specific document and answer structures.

  • Suite design: KARLBench evaluates six distinct search regimes, each isolating a structural challenge in acquiring information, retrieving evidence, or reasoning over retrieved content.The suite includes constraint-driven entity search, report synthesis, tabular reasoning, exhaustive retrieval, procedural reasoning, and enterprise-note fact aggregation.
  • Evaluation design: Agents are restricted to vector search to isolate knowledge acquisition and evidence integration from broader multi-tool orchestration effects.This differs from benchmarks requiring coordinated use of multiple tools.
  • Task capabilities: BrowseComp-Plus progressively filters candidates to identify one entity satisfying multiple interacting attributes distributed across documents.TREC-Biogen instead requires integrating dispersed biomedical findings into a structured report, while FinanceBench requires numerical calculation from financial reports.
  • Task capabilities: QAMPARI requires exhaustive retrieval of every entity satisfying a condition across many passages, whereas FreshStack requires procedural answers assembled from technical documentation and source code.PMBench aggregates distributed facts from heterogeneous internal company notes under realistic enterprise conditions.
  • Corpus construction: The benchmark preserves original document structure and segmentation with minimal indexing transformations rather than tuning preprocessing for individual datasets.This design prioritizes generalization across corpora with differing structural properties and uses closed-corpus settings for controlled comparison.
  • Evaluation limitation: The BrowseComp-Plus protocol indexes only the first 512 tokens of each document, covering 86.5% of gold evidence and imposing an upper bound under pure vector retrieval.The restriction is retained for comparability with published results.
  • Evaluation: The evaluation reports questions, indexed chunks, average token counts, relevant chunks, and answer nuggets, using nugget-based completion across tasks.BrowseComp-Plus and FinanceBench use a single predicted nugget, while other tasks use task-specific or reference-derived nuggets.

3 Agent Harness

The agent harness uses iterative vector-search interaction with automatic context compression for long rollouts. The synthesis pipeline uses few-shot examples and corpus exploration to produce grounded, deduplicated question-answer data.

  • Agent interaction: The agent has one external tool, Vector Search, and iteratively issues queries before producing a final answer.This setup focuses the harness on agentic retrieval rather than broader tool orchestration.
  • Retrieval configuration: Retrieved-chunk count is chosen to maintain a consistent token budget across datasets, with k capped at 20.BrowseComp-Plus uses Qwen3-8B embeddings with k = 20, while PMBench uses GTE-large.
  • Context management: When interaction history exceeds a token threshold, the model compresses it into a shorter summary within a predefined token count.Compression is intended for long rollouts and is performed by the model itself.
  • Question-answer synthesis: Stage I takes few-shot examples and a task corpus, then uses vector search to propose question-answer pairs grounded in retrieved documents.The synthesis prompt guides the agent toward diverse and difficult questions in the expected task format.
  • Question-answer synthesis: A Deduplication Agent removes exact or near-duplicate synthetic pairs to reduce test-data leakage and overlap with provided examples.The deduplication step follows question-answer proposal in the synthesis pipeline.

4 Training a Knowledge Agent via Reinforcement Learning (KARL)

KARL trains knowledge agents with agentic synthetic-data generation, iterative large-batch off-policy RL, and multi-task training for out-of-distribution generalization.

  • Agentic Synthesis: Agentic synthesis uses vector search and compression to create grounded question-answer pairs, then bootstraps improved agents for further data generation.The pipeline comprises question-answer synthesis and solution synthesis.
  • Agentic Synthesis: Multiple solver attempts estimate question difficulty; all-easy and all-hard examples are filtered before a quality agent removes ambiguity and incorrect references.Passing question-answer pairs and solution trajectories become RL training inputs.
  • Off-Policy RL: OAPL is a post-training recipe based on large-batch iterative off-policy reinforcement learning.The reference policy generates offline data, and the optimization can be iterated with the latest policy.
  • Off-Policy RL: The objective uses rewards relative to an estimated optimal value while applying KL regularization against a reference policy.Two parameters separately control value-estimate smoothness and KL-regularization strength.
  • Multi-task RL: For multi-step agentic rollouts, optimization masks prompt and tool-output tokens and can split long trajectories at compression steps.Each segment receives the rollout reward, and compression actions are included in RL optimization.
  • Multi-task RL: Multi-task training combines BrowseComp-Plus and TREC-Biogen, whose deep-search and wide-search demands support out-of-distribution generalization.The approach combines both losses and balances the datasets.

5 Scaling KARL via Test-time Compute

KARL scales test-time compute with parallel thinking and value-guided search. Parallel thinking improves performance broadly, while value-guided search provides more task-dependent gains.

  • Parallel Thinking: Parallel thinking boosts KARL performance across KARLBench, including tasks outside the model’s training distribution.Parallel computation is used to remain conscious of latency.
  • Parallel Thinking: Parallel thinking generates N independent rollouts in parallel and feeds their extracted answers back to the same model for aggregation.Tools are available during both generation and aggregation.
  • Parallel Thinking: 23.7% of PMBench cases with 5 parallel rollouts produce an aggregated answer better than every individual rollout.Tool-using aggregation is therefore more expressive than Best-of-N or Majority Vote.
  • Value-Guided Search: The value model predicts future success probability from partial rollouts and is trained token-by-token from policy-generated responses and binary rewards.The model’s sigmoid output estimates the probability of a correct answer from the current partial rollout.
  • Value-Guided Search: Value-guided search uses a value model to score candidate continuations, selects the highest-valued branch, and repeats the search N times before aggregation.The implementation performs k independent candidate steps at each assistant step.
  • Value-Guided Search: KARL uses N parallel breadth-first searches and aggregates the resulting rollouts with strategies such as Best-of-N or weighted majority vote.The number of candidate steps is fixed to k=2 while test-time compute scales with N.

6 Agent Infrastructure

KARL’s aroll infrastructure orchestrates high-throughput agent-environment rollouts with composable rewards and lifecycle plugins for search-agent training and evaluation.

  • Search Infrastructure: The vector-search infrastructure uses an embedded columnar database with offline chunking, embedding, indexing, and shared-storage caching to support high query throughput.The design addresses large volumes of retrieval queries during synthesis and evaluation.
  • Infrastructure Requirements: The rollout framework targets hundreds of thousands of long-horizon rollouts, composable task-specific rewards, and identical harness behavior across stages.These requirements motivate purpose-built infrastructure rather than general-purpose frameworks.
  • Harness Architecture: The dispatcher feeds prompts to exploration strategies that instantiate environment-agent pairs and produce finished rollouts concurrently.The environment executes tool calls, evaluates rewards, and manages context.
  • Lifecycle Plugins: Lifecycle plugins handle cross-cutting operations such as context compression, step budgeting, and tool gating without changing core execution code.Plugins can reshape history, gate or rewrite tool calls, or override termination.
  • Composable Strategies: Parallel Thinking configures concurrent rollouts followed by aggregation, while Value-Guided Search adds per-step candidate selection through a dedicated value model.This separates evaluation from generation compared with using the language model itself as evaluator.

7 Experiments

Across KARLBench, multi-task RL improves grounded-reasoning performance, generalization, and cost-latency efficiency, while iterative training develops broader search capabilities and environment-robust context management.

  • Main Results: Single-task variants reach 85.0 on TREC-Biogen and 59.6 on BrowseComp-Plus, improving to 70.4 with value-guided search, but neither transfers across the two tasks.The transfer failure indicates that the two in-distribution benchmarks test fundamentally different search capabilities.
  • Main Results: Multi-task KARL matches Claude Opus 4.6 with 10 parallel thinking traces while generalizing from TREC-Biogen and BrowseComp-Plus to held-out grounded-reasoning tasks.Without test-time compute, KARL matches Claude Sonnet 4.5 at high reasoning effort and outperforms similarly sized open-source models.
  • Main Results: KARL defines both cost-quality and latency-quality Pareto frontiers, delivering frontier-quality search at lower cost and latency than alternatives.Single-call KARL costs under $0.10 per query above 55 KARLBench points; with parallel sampling, it matches Claude Opus 4.6 at roughly 33% lower cost and approximately 47% lower latency.
  • Iterative Training: RL gains continue across iterations: KARL-TREC rises from 66.0 to 85.0, while QAMPARI increases from 45.9 to 50.8 across three iterations.FreshStack dips at iteration 1 before recovering to 56.7 at iteration 2, approaching the Claude Opus 4.5 baseline.
  • RL Generalizes beyond Sharpening: KARL-TREC improves max@k at every k, with three iterations reaching the base model’s max@8 at max@1 and exceeding its max@16 at max@2.Parallel Thinking consistently exceeds the base model’s performance ceiling across iterations, indicating that RL gains compound with test-time compute.
  • Training Ablations: Search Environment Generalization: KARL-BCP performance scales through 200 search steps and plateaus through 400, remains stable from 10 to 20 retrieved documents, and drops sharply at 40.Removing compression reduces accuracy from 0.570 to 0.389, whereas replacing the embedding model produces nearly identical performance.
  • Training Ablations: Search Environment Generalization: Replacing GLM 4.5 Air with KARL-BCP as the compression model improves performance from 0.44 to 0.54 and from 0.46 to 0.57 across search-model settings.Using GLM 4.5 Air for compression with KARL-BCP search instead degrades performance from 0.57 to 0.46.

8 Understanding the Impact of RL

RL training reshapes search behavior by shortening trajectories, increasing retrieval diversity, improving answer accuracy, and changing termination decisions. These gains coexist with failure modes involving premature stopping and unresolved post-retrieval reasoning.

  • Trajectory Length Shortening: RL training shortens trajectories across Unsolved, Partial, and Solved categories, with the largest decrease for Solved questions despite no direct training on them.This indicates improved efficiency even on problems the model already knew how to solve.
  • Search Diversity: 37% more documents retrieved for BrowseComp-Plus and 8% more for TREC-Biogen accompany increased search diversity across training iterations.The authors suggest this may contribute to improved KARL performance, while lower diversity in GLM 4.5 Air may reflect repetitive searches.
  • Answer Accuracy: Almost 20% absolute accuracy gains occur for KARL Iter. 2 over GLM 4.5 Air when all or some ground-truth documents are retrieved.RL improves final answer accuracy across full, partial, and absent retrieval conditions.
  • Search Efficiency: Across 87 full-recall queries, iterative RL reduces searches after all necessary evidence has been retrieved, improving search efficiency.The analysis isolates cases where all three models achieve full recall across both rollouts.
  • Termination Behavior: Adaptive stopping balances persistence against efficiency, but some shorter traces reflect premature termination during difficult reasoning rather than genuine efficiency.KARL can commit with partial evidence or stop when further searches seem ineffective, while arithmetic cases reveal unresolved post-retrieval reasoning failures.
  • Behavioral Profiles: KARL more reliably transitions from exploration to commitment, although its increased Giving Up Early behavior may reflect a spurious link between short traces and correctness.Its profile is dominated by Explore then Commit trajectories, unlike GLM 4.5 Air’s greater Exhaustive Search, No Convergence, and context truncation.

9 Conclusion

KARL combines a six-regime benchmark, agentic synthesis, and iterative off-policy multi-task RL to train grounded-reasoning agents. It is Pareto-optimal against recent Claude and GPT models, generalizes to out-of-distribution tasks, and remains limited by its single-tool design and simple context compression.

  • Contributions: KARLBench spans six search regimes, and heterogeneous training yields better generalization than optimizing for a single task.The suite covers diverse grounded-reasoning behaviors, including entity search, report synthesis, tabular reasoning, retrieval, technical reasoning, and enterprise-note aggregation.
  • Contributions: Agentic synthesis uses long-horizon reasoning and tool use with iterative bootstrapping to produce diverse, grounded training data.The data-generation process improves as increasingly capable models are incorporated.
  • Contributions: OAPL is an iterative large-batch off-policy RL method that is sample efficient, robust to trainer-inference discrepancies, and naturally supports multi-task training.The method avoids heuristics typically used to stabilize online RL at scale.
  • Results: KARL is Pareto-optimal relative to recent Claude and GPT models across cost-quality and latency-quality trade-offs, generalizes out of distribution, and benefits from test-time compute.These components produce cost-efficient, high-performing knowledge agents for grounded reasoning.
  • Limitations and Future Work: The current agent uses only vector search and simple prompt-based compression, motivating extensions to structured retrieval, code execution, sub-agents, and hierarchical memory.These are proposed directions for increasing agent capability and improving context management.

B Cost and Latency Experiment Details

The experiments evaluate model cost and latency with standardized agentic-harness procedures, shared retrieval infrastructure, and hierarchical aggregation across rollouts, splits, and benchmarks. The supplied passages also document KARLBench’s diverse task coverage and example report-synthesis outputs.

  • Evaluation Setup: Models are evaluated in an agentic harness using selected reasoning-effort configurations and standardized inference settings.Medium effort performs best for GPT 5 and GPT 5.2, while high effort performs best for Claude models.
  • Cost: Cost per query is computed from average input and output token counts across four generations per prompt using external token prices.The calculation aggregates token usage across the full KARLBench benchmark.
  • Latency: Latency uses wall-clock time to the first answer token, with full end-to-end time reported when rollouts exhaust the step budget or context window.Measurements use vLLM on an 8-H200 node with tensor parallelism 8 and a shared vector-search index.
  • Latency: Latency aggregation averages trajectories within benchmarks, benchmarks within splits, and three dataset splits into the final value.The setup samples 30 measured trajectories per prompt after discarding three warm-up trajectories.
  • Benchmark Coverage: KARLBench includes six evaluation regimes, including cross-document report synthesis, tabular numerical reasoning, procedural technical reasoning, and enterprise-note fact aggregation.The supplied example asks for a comprehensive account of gene-therapy effects across therapeutic, molecular, immune, safety, long-term, and physiological categories.

C.4 QAMPARI

QAMPARI tests exhaustive entity retrieval: the agent must identify the complete set of entities satisfying a condition across dispersed encyclopedic passages. The examples show KARL enumerating 17 designs by James B. Longacre, while FreshStack illustrates procedural reasoning over technical documentation and enterprise notes provide a separate aggregation setting.

  • QAMPARI: 17 distinct designs are exhaustively enumerated for James B. Longacre across two categories.The output includes coins, wreaths, medals, banknote engravings, portraits, and the “In God We Trust” motto.
  • QAMPARI: The Longacre example identifies his designs using encyclopedic source text and emphasizes comprehensive entity extraction.The answer includes both major coin designs and additional designs such as medals and engravings.
  • FreshStack: FreshStack’s recommended solution loads each job object, extracts selected fields, builds a vector store, and supports retrieval-based question answering.Alternative JSON-pointer and JSON Lines configurations are also described.
  • FreshStack: FreshStack requires procedural reasoning over technical documentation, including diagnosing JSON-loader behavior and constructing step-by-step solutions.The example explains why 13 documents appear instead of 3 and recommends JSONLoader configurations for question answering.
  • PMBench: PMBench aggregates facts from roughly 3,000 heterogeneous product-manager documents containing customer conversations and summaries.Its questions are difficult because information is diffuse and unorganized across internal notes.

D Prompts

The appendix documents prompts for evaluation, synthesis, deduplication, quality filtering, and multi-iteration data processing. These components judge answer completeness, remove semantic duplicates, and retain difficult, well-defined synthetic tasks.

  • D.1 Prompts for Evaluation Judges: Evaluation prompts judge whether answers support each decompositional fact or nugget.The nugget-completeness judge reads every fact and document pair and assigns support levels.
  • D.2 Prompts for Agentic Synthesis: Deduplication judges identify question-answer pairs encoding the same underlying fact, including inverse question framings.Pairs with the same answer are not duplicates when they ask different facts.
  • D.2 Prompts for Agentic Synthesis: BrowseComp-Plus deduplication compares generated pairs with validation pairs using retrieved similar examples and a judge model.The process handles direct matches and paraphrases, including answer variants such as nicknames for the same person.
  • D.2 Prompts for Agentic Synthesis: Quality filters retain synthetic questions that are factually correct, unambiguous, and empirically neither trivial nor impossible for solvers.BrowseComp-Plus uses binary correctness, while TREC-Biogen uses nugget-completion scores and removes ambiguous or incorrectly specified nuggets.
  • D.2 Prompts for Agentic Synthesis: The two-iteration synthesis pipeline deduplicates questions, generates eight solver rollouts, categorizes outcomes, and optionally quality-filters partial solutions.The pipeline is shown for both BrowseComp-Plus and TREC-Biogen.
  • D.4 Deduplication Filter Examples: Near-duplicate detection catches semantically equivalent tasks despite different wording or answer forms.Examples include a TREC-Biogen pair with cosine similarity 0.855 and BrowseComp-Plus nicknames referring to the same person.
  • D.3 Data Synthesis Statistics: BrowseComp-Plus Iteration 1 yields 1,218 filtered questions from 13,882 synthetic pairs, an 8.8% yield; Iteration 2 reaches 16.2% without comparable quality filtering.Deduplication removes nearly 40% of Iteration 1 pairs and 15% of Iteration 2 pairs.
  • D.5 Quality Filter Examples: A quality-filter example illustrates that a plausible answer can still be rejected when the question has ambiguous or insufficiently supported constraints.The appendix introduces such examples in Table 8.

E.1 Case Study: KARL Outperforms GLM 4.5 Air and Sonnet 4.5 on Search

Case studies show KARL solving difficult multi-constraint search and reasoning problems where competing agents stop early, make unsupported assumptions, or fail to verify all required details. The examples also show improved search commitment and broader nugget coverage.

  • E.1 Search: KARL identifies Sol Campbell as the away team’s third substitute after systematically verifying the question’s constraints.Sonnet 4.5 abstains, while GLM 4.5 Air searches extensively but fails to locate the decisive detail.
  • E.1 Search: The BrowseComp-Plus comparison table presents a case where KARL locates the key details and avoids premature termination.The table concerns Query ID 280.
  • E.2 Reasoning: KARL correctly identifies Swahili poetry by narrowing candidates and checking the first book’s genre against the relevant evidence.Competing systems either select the wrong candidate or infer an unsupported fiction genre.
  • E.4 Near Misses: After RL training, KARL commits after sufficient evidence, whereas GLM 4.5 Air continues verification despite finding the answer early.The base model found the answer by query 4 but searched 161 times, and another attempt continued to step 201.
  • E. Case Studies: KARL covers all 9 ground-truth nuggets in one reasoning example, while another attempt covers 7/9 and misses genetic disorders and malignancy.The reported nugget categories include bone, genetic, metabolic, medication, seizure, systemic, stress, malignancy, and rheumatologic causes.
  • E.7 Parallel Thinking: KARL’s aggregation rollout combines complementary evidence from multiple candidate rollouts into a supported final answer.In the PMBench example, one nugget appears in only 40% of rollouts but is correctly surfaced by aggregation.

F Categorizing Search Behavior

The paper categorizes search traces by observable behaviors such as truncation, uncertainty, proposed answers, and verification searches. It combines rule-based labeling with hand calibration and analyzes context compression during long trajectories.

  • Search behavior taxonomy: The search taxonomy includes truncated responses, expressed uncertainty, proposed answers, and verification searches.Each category is detected from features of the model’s individual search trace.
  • Classifier calibration: The rule-based classifier was calibrated on 30 hand-labeled diverse rollouts before labeling BrowseComp-Plus trajectories.Borderline traces were manually annotated after automatic labeling.
  • Context compression: BrowseComp-Plus questions require a median of 6 and mean of 10.2 context-compression steps per question.The distribution is strongly right-skewed, with a long tail of harder questions requiring many compressions.
  • Context compression: Compression summarizes accumulated retrieved content and intermediate reasoning when the context window fills during extended search.The mechanism is intended to let the agent continue reasoning within a bounded context.
  • Context compression: In two trajectories, compression reduces more than 100K characters accumulated over four search steps to roughly 1K characters.The examples emphasize that downstream performance depends on which information the compressor preserves.

G.2.1 Successful Compression: Author Identification

The author-identification trajectory compressed a converged solution into a short summary and preserved the key evidence needed to answer the puzzle correctly. In contrast, the ICC Hall of Fame trajectory compressed an unresolved deduction and anchored subsequent reasoning to an incorrect year.

  • Successful Compression: Author Identification: The successful trajectory compressed a 9-step, 100× rollout whose outcome was correct.The pre-compression phase contained four search steps, seven messages, and 112,946 characters.
  • Successful Compression: Author Identification: Retrieved evidence confirmed Manushkin’s Chicago background, 1972 publication, prior occupations, and hobbies of bird watching and cat watching.The trajectory also verified the title change from “Baby” to “Baby, Come Out!”.
  • Successful Compression: Author Identification: The search identified Fran Manushkin as matching the puzzle’s constraints, including Chicago, teaching and tour-guide work, and a 1972 first book.The identified book was later renamed “Baby, Come Out!”.
  • Successful Compression: Author Identification: The compressed summary preserved the author, city, book-title change, occupations, and hobbies, enabling the correct answer “Baby”.The final response matched the expected answer exactly.
  • Harmful Compression: ICC Hall of Fame Puzzle: The harmful trajectory compressed a 15-step, 100× rollout while its outcome was incorrect.The model was still exploring year offsets when compression occurred.
  • Harmful Compression: ICC Hall of Fame Puzzle: Compression anchored the unresolved Hall of Fame reasoning on Kapil Dev’s 2009 induction, despite the correct answer being 2010.A later retrieved document contained the correct year, but the model did not register the contradiction.

H Evaluation Infrastructure

The evaluation infrastructure includes a side-by-side testing application for comparing model outputs and lightweight tooling for analyzing search behavior and efficiency.

  • Testing Suite: Users can compare multiple models and configurations on the same query, including streaming outputs, tool calls, and retrieval citations.An optional blind mode hides model identities and citations until after a preference is recorded.
  • Testing Suite: The application uses PMBench to evaluate single-turn question answering and deep-research behavior over unstructured corpora.It is built on the aroll framework.
  • Testing Suite: The testing application gathers preferences on real-world queries and maintains a running scoreboard of preferred models.Figure 42 presents the testing interface.
  • Qualitative Analysis Tooling: Qualitative-analysis tooling compares candidate models per question and tracks time spent grounding, exploring, and verifying.The tooling supports analysis of differentiated search strategies and verification behavior.
Loading 2603.05218v1…