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

Grounded-reasoning tasks require varied search skills, but existing benchmarks cover only a limited slice of knowledge-agent behaviors. KARL combines a six-regime benchmark with synthesized training data and multi-task reinforcement learning, and is Pareto-optimal against Claude 4.6 and GPT 5.2 across cost-quality and latency-quality trade-offs.

  • Problem

    Existing benchmarks capture only a limited slice of knowledge-agent behaviors, despite practical grounded-reasoning tasks requiring varied search and reasoning skills.

    Source: §1 · p. 2

  • Method

    KARL combines a six-regime evaluation suite with agentic training-data synthesis and iterative large-batch off-policy reinforcement learning for multi-task training.

    Source: §1 · p. 2, Agentic synthesis · p. 2, §9 · p. 29

  • Results

    Across KARLBench, heterogeneous training generalizes better than single-benchmark training, and KARL is Pareto-optimal against Claude 4.6 and GPT 5.2 across cost-quality and latency-quality trade-offs.

    Source: §1 · p. 2, Agentic synthesis · p. 2

  • Takeaways & Limitations

    The paper concludes that cost-efficient knowledge agents for grounded reasoning across diverse tasks can emerge from specialized synthetic data, multi-task reinforcement learning, and test-time compute scaling.

    Source: Agentic synthesis · p. 2

    KARL can prematurely terminate on difficult arithmetic tasks and has not improved post-retrieval numerical computation, motivating explicit arithmetic and tabular-reasoning rewards.

    Source: §8.2.2 · p. 28

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 enterprise reasoning with a diverse benchmark, agentic data synthesis, and iterative off-policy RL. It reports Pareto-optimal cost and latency trade-offs against closed models on KARLBench, with higher quality at sufficient test-time compute.

1 Introduction

  • Grounded reasoning requires multi-step evidence gathering over knowledge beyond model parameters, while public-web research and task-specific benchmarks do not establish broad competence across diverse enterprise skills.Enterprises rely on proprietary data, and practical applications include constraint-based search, report synthesis, numerical reasoning, and procedural reasoning.Source: §1 · p. 2
  • KARLBench spans six search regimes, includes the proprietary PMBench, and shows that heterogeneous search training generalizes better than training optimized for a single benchmark.The regimes cover constraint-driven entity search, report synthesis, tabular numerical reasoning, exhaustive retrieval, procedural technical reasoning, and enterprise-note fact aggregation.Source: §1 · p. 2

Agentic synthesis.

  • The agentic synthesis pipeline explores corpora with vector search to generate evidence-grounded question-answer pairs, then bootstraps improved agents to synthesize further data.The recipe generalizes across TREC-Biogen and BrowseComp-Plus, which require different search behaviors.Source: Agentic synthesis · p. 2
  • OAPL uses iterative large-batch off-policy RL designed to tolerate trainer–inference discrepancies without clipping, data deletion, or router replay; combined task losses improve both tasks and generalize to four held-out KARLBench tasks.The multi-task training combines BrowseComp-Plus and TREC-Biogen losses.Source: Agentic synthesis · p. 2
  • KARL is Pareto-optimal against Claude 4.6 and GPT 5.2 on KARLBench, matching quality at lower cost and latency and exceeding strongest closed-model quality with sufficient test-time compute.The reported results use a GLM 4.5 Air starting point with varying levels of test-time scaling.Source: Agentic synthesis · p. 2

2 KARLBench

KARLBench evaluates knowledge agents across six distinct search tasks, each isolating a capability involved in acquiring information, retrieving evidence, or reasoning over it. It uses controlled, closed-corpus evaluation with minimal dataset-specific preprocessing.

2.1 Tasks Overview

  • KARLBench’s six tasks span constraint-driven entity search, cross-document report synthesis, tabular numerical reasoning, exhaustive entity retrieval, procedural reasoning, and fact aggregation over internal notes.BrowseComp-Plus requires filtering candidates against interacting attributes; TREC-Biogen integrates biomedical sources; FinanceBench requires extracting and calculating dispersed values.Source: §2.1 · p. 4
  • QAMPARI tests exhaustive retrieval of all matching entities, FreshStack tests step-by-step solutions from technical documentation and code, and PMBench tests aggregation across noisy internal notes.PMBench was developed in-house to measure search robustness under realistic enterprise conditions.Source: §2.1 · p. 4

2.2 Corpus Construction

  • The benchmark preserves original document structures and segmentation where possible, avoids performance-tuned preprocessing, and uses closed corpora to support controlled comparisons across datasets.For BrowseComp-Plus, indexing only the first 512 tokens covers 86.5% of gold evidence, imposing an upper bound under pure vector retrieval; the restriction preserves comparability with published results.Source: §2.2 · p. 4, §2.3 · p. 5

2.3 Evaluation

  • Evaluation uses nugget-based completion: individual entities count as QAMPARI nuggets, while FreshStack and PMBench answers are converted into fixed task-specific nuggets and TREC-Biogen references are separately converted and consolidated.BrowseComp-Plus and FinanceBench each require predicting a single nugget correctly.Source: §2.3 · pp. 5–6

3 Agent Harness

KARL’s harness uses vector search as its sole external tool, with the agent iteratively querying and answering from a bounded interaction history. It automatically compresses long histories, and retrieval settings are chosen to keep token budgets consistent across datasets.

Vector search as the sole external tool

  • The agent uses vector search as its sole external tool, iteratively querying retrieved documents before producing a final answer.Its trajectory includes prior tool calls and outputs; earlier steps are selectively summarized when context exceeds a predefined limit.Source: §3 · p. 6

Retrieval configuration

  • Retrieval chunk counts scale inversely with average chunk length to maintain a consistent token budget, capped at k = 20.BrowseComp-Plus uses Qwen3-8B embeddings (Zhang et al., 2025b), PMBench uses GTE-large (Li et al., 2023), and other datasets use Qwen3-0.6B with k = 20 or k = 10 for FreshStack.Source: §3 · p. 6

Context management via compression

  • Compression triggers when history exceeds a token threshold and is trained end-to-end with query and answer generation using task outcome rewards.The agent itself compresses the history without separate pretraining on summarization data, learning what to preserve to maximize task rewards.Source: §3 · p. 6

4 Training a Knowledge Agent via Reinforcement Learning (KARL)

KARL trains knowledge agents through grounded two-stage data synthesis, large-batch iterative off-policy RL, and multi-task training intended to improve out-of-distribution generalization. The method filters for useful, reliable learning signals and adapts RL to long, multi-step search trajectories.

4.1 Agentic Synthesis

  • Synthetic training data comes from retrieved-document question-answer generation, then solver pass-rate and quality filters remove trivial, unsolvable, ambiguous, or inaccurate examples.A deduplication agent removes exact and near-duplicates; retained question-answer pairs and solution trajectories feed RL training.Source: §4.1 · pp. 7–8, Figure 3 · p. 8

4.2 Post-training via Off-Policy RL

  • KARL uses OAPL, a large-batch iterative off-policy RL recipe with separate parameters for value estimation smoothness and KL regularization.It can regenerate a large offline dataset from the latest policy and repeat optimization for up to three iterations.Source: §4.2 · pp. 8–9
  • For multi-step agent rollouts, KARL masks non-policy tokens and segments long histories at compression steps, training summaries end-to-end with rollout outcome rewards.Assigning the rollout reward to each segment avoids training on extremely long rollouts that require large GPU memory.Source: §4.2 · p. 9
  • Large-batch off-policy training amortizes data-generation costs across policy updates and hyperparameter sweeps, and trains large MoE models stably without prior GRPO heuristics.The authors describe it as a computationally more efficient and less complex alternative to online RL infrastructure.Source: §4.2 · p. 9

4.3 Multi-task RL Post Training

  • Multi-task RL combines BrowseComp-Plus deep search and TREC-Biogen wide search with roughly balanced training tokens, improving both tasks and OOD generalization over multi-expert distillation.The two approaches show similar in-distribution performance, while multi-task RL generalizes better out of distribution.Source: §4.3 · p. 9

5 Scaling KARL via Test-time Compute

Test-time compute boosts KARL while accounting for cost and latency: parallel thinking improves performance broadly, including on out-of-domain tasks, while value-guided search offers task-dependent gains. The section presents both approaches and explains how parallel rollout aggregation can improve on every individual rollout.

5 Scaling KARL via Test-time Compute

  • Parallel thinking boosts KARL across KARLBench, including out-of-domain tasks, while task-specific value-guided search offers more task-dependent gains.The reported comparison frames parallel computation as a latency-conscious alternative to sequential computation.Source: §5 · p. 10, §5.1 · p. 10

5.1 Parallel Thinking TTC

  • Parallel thinking generates independent rollouts in parallel, then uses the model—with tools—to aggregate their short answers into a final answer.The aggregator processes short contexts, making the approach efficient while allowing it to synthesize an answer rather than simply select a rollout.Source: §5.1 · p. 10
  • 23.7% of the time on PMBench with 5 parallel rollouts, the aggregator generates a better answer than any rollout, making the approach more expressive than Best-of-N or Majority Vote.This finding shows that aggregation can produce answers beyond those present in the parallel rollouts.Source: §5.1 · p. 10

5.2 Reward-based TTC via Value-Guided Search

  • VGS uses a value model to predict future success from partial rollouts, performs N parallel breadth-first searches yielding N rollouts, and aggregates them with strategies such as Best-of-N or weighted majority vote.In experiments, a small Qwen3-4B-Thinking-2507 value model was sufficient to steer the policy toward higher-reward generations.Source: §5.2 · pp. 10–11

6 Agent Infrastructure

Section 6 describes infrastructure built to support high-throughput search and consistent agent rollouts across data collection, training, evaluation, and serving. Its embedded vector search and modular harness support scalable retrieval and configurable task behavior.

6.1 Scaling Vector Search

  • Embedded columnar vector search loads cached indexes into each worker’s in-process database, exceeding 500 queries per second per host without client-server network I/O.The knowledge corpus is chunked, embedded, and indexed offline; retrieval is exposed to the agent as a query-and-context environment function.Source: §6.1 · p. 11

6.2 Agent Harness Implementation

  • The aroll framework was built for high-throughput long-horizon rollouts, composable KARLBench rewards, and identical harness behavior from collection through training, evaluation, and test-time compute.The authors report that no existing open-source framework met these requirements.Source: §6.2 · p. 11
  • The harness separates prompt orchestration, environment interaction and reward evaluation, and per-step agent generation; configured reward functions let new tasks join multi-task training without code changes.The exploration strategy runs environment-agent pairs concurrently, while the environment executes tool calls and evaluates task-specific rewards when a rollout ends.Source: §6.2 · p. 12
  • Lifecycle plugins handle context compression, step budgets, and tool gating by intercepting the interaction loop, and the same configured plugins run in collection, training evaluation, and serving.Plugins can reshape history, gate or rewrite tool calls, or override termination; compression activates at a token threshold and prompts the agent to summarize its history.Source: §6.2 · p. 12
  • The layered interfaces allow parallel rollouts, value-guided candidate selection, compression, and multi-task reward composition to be configured while reducing distributional shift between data collection and serving.Parallel Thinking aggregates concurrent rollouts, while Value-Guided Search selects among candidate continuations using a trained value model.Source: §6.2 · p. 12

7 Experiments

The experiments evaluate how multi-task RL, training iterations, and test-time compute affect KARL’s efficiency and generalization across search tasks, including the contributions of learned context compression, search strategies, parallel thinking, and value-guided search.

7.1 Evaluation and Training Set

  • Evaluation trains on TREC-Biogen and BrowseComp-Plus while holding other tasks out, and uses a calibrated 230-question BrowseComp-Plus subset for reported results.Synthesis uses the evolving RL model; task-specific pipelines generate candidate questions and rollouts, then filter duplicates, ambiguous items, incorrect answers, and all-correct or all-incorrect cases. BrowseComp-Plus document seeding increased document coverage by 25%.Source: §7.1 · p. 13, §7.2 · p. 13, §7.2.1 · p. 13

7.3.1 Main Results

  • KARL matches Claude Sonnet 4.5 without test-time compute and Claude Opus 4.6 with 10 parallel thinking traces, while generalizing to out-of-distribution grounded reasoning tasks.Single-task KARL-TREC scores 85.0 on TREC-Biogen and KARL-BCP scores 59.6 on BrowseComp-Plus, rising to 70.4 with value-guided search; neither transfers to the other training task.Source: §7.3.1 · p. 15, Table 4 · p. 16

7.3.2 Cost and Latency

  • KARL defines the 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, the lowest-cost model above 55 points; with parallel sampling it matches Claude Opus 4.6 at roughly 33% lower cost and 47% lower latency, and without parallel thinking is the fastest model above 55 points.Source: §7.3.2 · pp. 15–16, Table 4 · p. 16

7.3.3 Multi-Expert Distillation vs. Multi-Task RL

  • SFT distillation gains little out-of-distribution from parallel sampling, whereas KARL benefits from test-time compute in both in- and out-of-distribution tasks.SFT in-distribution performance rises 69.1 →75.3, but out-of-distribution performance only changes 59.4 →59.6; KARL’s broader scaling keeps it on the Pareto frontier as compute increases.Source: §7.3.3 · p. 16, Figure 8 · p. 17

7.3.4 Multi-Iteration Training

  • KARL-TREC improves from 66.0 at the base model to 85.0 after three iterations, while FreshStack reaches 56.7 by iteration 2 and QAMPARI rises from 45.9 to 50.8 across three iterations.The TREC-Biogen expert surpasses Claude Sonnet 4.5 by iteration 1 and Claude Opus 4.5 by iteration 2; its out-of-distribution gains suggest transfer to structurally different search tasks.Source: §7.3.4 · pp. 17–18, Figure 9 · p. 17

7.3.5 RL Generalizes beyond Sharpening

  • KARL-TREC’s max@k improves at every k, and after three iterations its max@1 matches base GLM 4.5 Air’s max@8 while its max@2 exceeds base max@16.For BrowseComp-Plus prompts filtered from solved and unsolved categories before training, 33.3% of partial prompts become solved and 37.2% of unsolved prompts become partial, while degradation is minimal.Source: §7.3.5 · p. 18, Figure 11 · p. 19

7.3.6 Training Ablations: Search Environment Generalization

  • KARL-BCP performance scales with search horizon through 200 steps and plateaus through 400, remains stable at 10–20 retrieved documents, and drops sharply at 40.Removing compression drops accuracy 0.570 →0.389. KARL-BCP compression improves scores from 0.44 →0.54 with GLM 4.5 Air search and 0.46 →0.57 with KARL-BCP search; replacing its compressor with GLM 4.5 Air drops performance 0.57 →0.46, while a comparable retriever swap has nearly identical performance.Source: §7.3.6 · pp. 19–20, Table 5 · p. 19, Table 6 · p. 20

7.4.1 Parallel Thinking

  • With 5 parallel rollouts, the PMBench aggregator produces a better answer than every individual rollout 23.7% of the time, with diminishing returns beyond N=15.Across KARLBench, KARL consistently outperforms GLM 4.5 Air as parallel thinking scales from N=5 to 20, with gains at N=20 ranging from +1.9 on FinanceBench to +5.9 on TREC-Biogen; a generative aggregator combines candidate answers for open-ended tasks.Source: §7.4.1 · pp. 21–22, Figure 13 · p. 21
  • KARL maintains its advantage over GLM 4.5 Air across benchmarks and rollout counts as both benefit from more test-time compute.The results suggest RL-training gains complement test-time compute.Source: §7.4.1 · p. 22
  • Parallel-thinking aggregation adds few steps to the overall rollout compared with trajectory length.The passage describes Table 7 as measuring the additional steps for parallel thinking.Source: §7.4.1 · p. 22

7.4.2 Value-Guided Search (VGS)

  • 70.4 accuracy: value-guided search with weighted majority voting scales best on BrowseComp-Plus, exceeding parallel thinking.Weighted voting outperforms majority voting, indicating that the value model adds a useful reward signal; best-of-N performs worse, while scaling search improves document recall despite no recall training objective.Source: Figure 14 · p. 22, §7.4.2 · p. 22

8 Understanding the Impact of RL

RL training changed search trajectories: it increased diversity and efficiency, improved accuracy, and altered how KARL persists, verifies, and commits under uncertainty, while revealing numerical-reasoning failures.

8.1.1 Quantitative Analysis on Synthetic Data

  • Training shortened trajectories across Pass@16 categories, most sharply for Solved questions, despite no training on them.For questions with 150–200-step base-model trajectories, more than 20% of unsolved questions became partially solved, and about 10% of partially solved questions became solved.Source: §8.1.1 · p. 23

8.1.2 Quantitative Analysis on Evaluation Sets

  • Across cumulative unique documents retrieved over successive vector-search queries, KARL Iter. 2 retrieved 37% more for BrowseComp-Plus and 8% more for TREC-Biogen than GLM 4.5 Air.Source: Figure 17 · p. 24, §8.1.2 · p. 25
  • RL improved BrowseComp-Plus answer accuracy across all, some, and none of the ground-truth-document retrieval conditions; Iter. 2 gained almost 20% absolute for full and partial retrieval.KARL also sometimes answered correctly without retrieving annotated ground-truth documents, suggesting those annotations were not exhaustive.Source: §8.1.2 · p. 25, Figure 18 · p. 25
  • On 87 BrowseComp-Plus queries with perfect ground-truth recall, accuracy rose from 53% to 64% to 71% as searches fell from 91 to 52 to 32.The largest reductions were in unnecessary post-retrieval searches on incorrect instances: 134.0 →95.6 →56.5.Source: §8.1.2 · p. 25, Figure 19 · p. 26

8.2.1 Comparison with Baselines

  • In case studies, KARL persisted to answer correctly at step 155 after Sonnet stopped at 25 and GLM used 200 steps unsuccessfully, and another time identified a book and genre in 7 steps.In the latter case, Sonnet made an unsupported genre assumption and GLM committed to the wrong book despite narrowing candidates to two.Source: §8.2.1 · pp. 26–27

8.2.2 Behavioral Impact on Efficiency

  • KARL reduced redundant verification but showed flexible stopping, sometimes committing with partial evidence when further queries offered diminishing returns.One example answered after 57 steps or 7 steps, while GLM exhausted its 200-step budget; another stopped after 13 queries despite having the evidence needed for arithmetic, revealing a post-retrieval reasoning failure.Source: §8.2.2 · pp. 27–28

8.2.3 Behavioral Profiles

  • A six-class behavior taxonomy found KARL’s profile resembled Sonnet’s Explore then Commit pattern, while GLM more often searched exhaustively without converging or ran out of context.KARL also showed more Giving Up Early behavior, possibly reflecting a spurious association between shorter traces and correctness.Source: §8.2.3 · p. 28

9 Conclusion · Appendix Contents

KARL combines heterogeneous search training, agentic synthetic-data generation, and off-policy reinforcement learning to build a grounded knowledge agent that generalizes across diverse tasks. Its current single-tool and prompt-based memory setup leaves clear directions for extending the agent and improving context management.

9 Conclusion

  • KARL is Pareto-optimal against the latest Claude and GPT models on KARLBench’s cost/quality and latency/quality trade-offs, generalizes to out-of-distribution tasks, and benefits from test-time compute scaling.Together, these results support a cost-efficient path to high-performing knowledge agents for grounded reasoning and other hard-to-verify tasks.Source: §9 · p. 29
  • KARLBench spans six search regimes, and training across heterogeneous search behaviors generalizes substantially better than optimizing for a single task.The regimes include constraint-driven entity search, cross-document synthesis, tabular reasoning, exhaustive retrieval, procedural reasoning, and fact aggregation.Source: §9 · p. 29
  • An agentic synthesis pipeline uses long-horizon reasoning, tool use, and iterative bootstrapping from increasingly capable models to improve diverse, grounded training data.Source: §9 · p. 29
  • OAPL is an iterative large-batch off-policy RL method designed for sample efficiency, robustness to trainer/inference discrepancies, and multi-task training without typical online-RL stabilization heuristics.Source: §9 · p. 29
  • KARL currently operates with vector search as its single tool, with structured retrieval, code execution, and callable compositional sub-agents proposed as future additions.Source: §9 · p. 29
  • KARL currently uses simple prompt-based context compression; hierarchical memory management is proposed to improve its handling of a finite context window.Source: §9 · p. 29

B Cost and Latency Experiment Details

This section describes how the highest-scoring model configurations were selected and how cost and latency were measured across KARLBench. Latency measurement used agentic rollouts, a defined first-answer-token metric, and averaging across benchmarks and splits.

Configuration selection

  • Models were evaluated on the agentic harness, with the highest-scoring reasoning-effort and compression configuration used for cost and latency measurements.Medium effort worked best for GPT 5 and GPT 5.2, while high effort worked best for Claude models; compression did not benefit GPT or Claude but improved open-source model scores.Source: B Cost and Latency Experiment Details · p. 35

Cost measurement

  • Per-query cost was calculated from average input and output token counts across KARLBench and token prices from artificialanalysis.ai.Counts were recorded across the benchmark using four generations per prompt.Source: B Cost and Latency Experiment Details · p. 35

Latency measurement

  • Latency tests used an 8-H200 vLLM setup for GLM 4.5 Air and KARL, a shared vector index, concurrency 1, and 30 measured trajectories per prompt.For each benchmark, five prompts were sampled; the first three trajectories per prompt were discarded for warm-up, for 33 total, and tests were staggered by several hours.Source: B Cost and Latency Experiment Details · p. 35
  • Latency was wall-clock time to the first answer token, with full end-to-end time reported when rollouts exhausted the step budget or context window.Latency was averaged across trajectories within each benchmark, across benchmarks within each split, and across three dataset splits; the reported latency results cover GLM 4.5 Air, KARL, Sonnet 4.6, Opus 4.6, and GPT 5.2.Source: B Cost and Latency Experiment Details · p. 35

C Dataset Details and Examples … C.6 PMBench (Internal Benchmark)

This appendix gives dataset details and example KARL outputs across search regimes, alongside the scope, curation, and evaluation of the internal PMBench benchmark.

C Dataset Details and Examples

TREC-Biogen

  • KARL synthesizes findings from 15+ source documents into a cited report spanning seven thematic categories of gene-therapy effects.The example covers therapeutic, molecular and cellular, immune, safety, long-term and developmental, and physiological effects.Source: TREC-Biogen · p. 41, Figure 26 · p. 42

BrowseComp-Plus

  • KARL resolves a BrowseComp-Plus query by matching four independent constraints to identify an obscure 1983 television episode.The constraints concern the show’s decade, episode plot, an actor’s medical history, and childhood biography.Source: Figure 27 · p. 42

FinanceBench

  • KARL locates capital expenditure in a lengthy 10-K cash-flow statement and cross-validates the figure against a second table.Source: Figure 28 · p. 43

Qampari

  • KARL exhaustively enumerates 17 distinct designs by James B. Longacre across two categories, including coins and other items [20].The example includes the Indian Head cent, Flying Eagle cent, and Longacre’s addition of “In God We Trust” to coins in 1866 [20].Source: Qampari · p. 43, Figure 29 · p. 44

Freshstack

  • KARL diagnoses why LangChain’s JSONLoader returns 13 documents instead of three and identifies syntax errors in the JSON structure.Without JSON pointers, the loader extracts all string values into separate documents; the example then provides four progressively detailed solutions, including a complete QA pipeline.Source: Freshstack · pp. 44–45, Figure 30 · p. 46

C.6 PMBench (Internal Benchmark)

  • PMBench comprises 57 questions over roughly 3,000 heterogeneous customer-conversation documents and reflects the messy realities of internal enterprise data.Questions are manually verified and combine manual and synthetic curation; two approaches start from real product-manager queries or monthly summaries verified against customer notes. Answers receive partial credit for correct information nuggets using an LLM judge.Source: §C.6 · p. 46

D Prompts · E Qualitative Case Studies

Appendix D documents prompts used for evaluation, agentic synthesis, and test-time compute; Appendix E describes behavioral analyses of KARL’s search and reasoning.

D Prompts

  • The appendix presents prompts for evaluation, agentic synthesis, and test-time compute experiments.Source: D Prompts · p. 47

D.1 Prompts for Evaluation Judges

  • The nugget-completeness judge assesses whether an answer supports, partially supports, or does not support each decompositional fact.It is instructed to read the question, answer, and each fact carefully.Source: §D.1 · p. 47, Figure 31 · p. 47

D.2 Prompts for Agentic Synthesis

  • The synthesis pipeline deduplicates questions, generates eight solver rollouts per question, classifies outcomes, and uses partial-correct rollouts as RL data.Deduplication judges identify pairs expressing the same underlying fact, including inverse questions; BrowseComp-Plus retrieves similar validation answers before judging, while quality filters check factual correctness and ambiguity or nugget definition.Source: Figure 33 · p. 49, §D.2 · p. 49, Figure 35 · p. 50

D.4 Deduplication Filter Examples

  • Deduplication examples show the filter recognizing semantic duplicates despite changed wording or answer forms.For TREC-Biogen, a near-duplicate received cosine similarity 0.855 and was identified as equivalent; for BrowseComp-Plus, “Motor Mouth” and “The Lip” were recognized as the same person’s nicknames.Source: Figure 38 · p. 53, §D.4 · p. 53, Figure 39 · p. 54

D.3 Data Synthesis Statistics

  • BrowseComp-Plus synthesis yield was 8.8% in Iter. 1 and 16.2% in Iter. 2, but Iter. 2 skipped quality filtering and is not directly comparable.In Iter. 1, 1,218 questions with eight solutions each passed the filters from 13,882 starting pairs. Deduplication filtered BrowseComp-Plus pairs by 40% (13,882 → 8,358) in Iter. 1 and 15% (8,244 → 7,050) in Iter. 2; for TREC-Biogen, it filtered 6% (13,728 → 12,946) and 19% (29,288 → 23,694), respectively. The quality filter flagged 5.8% of BrowseComp-Plus pairs in Iter. 1; its impact was not measured by ablation.Source: §D.3 · p. 53

D.5 Quality Filter Examples

  • Quality-filter examples show removal of synthetic tasks with incorrect nuggets, ambiguous references, or incorrect ground truth.In the TREC-Biogen example, nine of ten nuggets were accurate but one was factually wrong. In a BrowseComp-Plus ambiguity example, four of eight attempts gave the literal answer 0 instead of the intended 3; another example had seven of eight attempts identify Apple despite a Tesla ground truth.Source: §D.5 · p. 55, Table 8 · p. 56, Table 9 · p. 57

E Qualitative Case Studies

  • KARL’s qualitative analyses report improved search and reasoning over the base GLM 4.5 Air and include a failure example.The analyses are presented as evidence about behavioral changes associated with improved performance and areas for further improvement.Source: E Qualitative Case Studies · p. 59

E.1 Case Study: KARL Outperforms GLM 4.5 Air and Sonnet 4.5 on Search … E.7 Case Study: Parallel Thinking Identifies and Merges the Best of Candidate Solutions

These case studies examine KARL’s search and answer behavior, including recovered nuggets, missed or uncertain answers, numerical-reasoning failures, and verification. They also show how aggregating complementary candidate answers can produce a response better than an individual rollout.

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

  • KARL identifies Sol Campbell as the away team’s third substitute after verifying the match constraints, while Sonnet 4.5 stops early and GLM 4.5 Air fails to find the key detail.The comparison concerns BrowseComp-Plus Query ID 280.Source: §E.1 · pp. 59–60, Table 11 · p. 60

E.2 Case Study: KARL Outperforms GLM 4.5 Air and Sonnet 4.5 on Reasoning

  • KARL correctly identifies Wallah bin Wallah’s first book genre as Swahili poetry, while Sonnet 4.5 assumes a genre and GLM 4.5 Air confuses the candidate.The book is identified as “Malenga wa Ziwa Kuu,” whose subject matter indicates poetry.Source: §E.2 · pp. 61–62, Table 12 · p. 62

E.3 Case Study: Search Efficiency of KARL

  • KARL’s searches often verify a candidate before committing; in the illustrated case, it confirms Saturna appears in the third paragraph of the specific review.The query analysis was conducted with Claude Opus 4.5 assigning intent to KARL’s queries.Source: §E.3 · p. 63, Table 13 · p. 64

E.4 Case Study: Near Misses from GLM 4.5 Air Fixed by KARL

  • On its second attempt, KARL resolved the publisher question in 4 search queries and 7 steps; GLM 4.5 Air found Sasquatch Books by query 4 but never committed, searching 161 times.Source: §E.4 · pp. 65–66, Table 14 · p. 66

E.5 Case Study: KARL Learns to Give Up

  • KARL exhibits a numerical-reasoning failure in BrowseComp-Plus Query ID 1185 and does not find the correct answer.The supplied passage provides no further case details.Source: Table 15 · p. 67

E.6 Case Study: KARL Identifies the Nuggets Missed by GLM 4.5 Air

  • 9 nuggets in its best attempt versus 5 for GLM 4.5 Air: KARL covers all ground-truth causes, including non-obvious ones the base model misses.With one fewer search, KARL’s first attempt found genetic disorders and renal disease but missed endocrine causes and malignancy.Source: §E.6 · p. 68
  • KARL used 7 searches to target non-obvious fracture causes, then included seizure-related mechanisms and drug-related bone loss in a final-response section absent from the base model’s answer.After finding seizure-related fractures in initial results, it followed up with a dedicated search and also searched explicitly for medication-induced bone loss.Source: §E.6 · p. 69
  • KARL’s first attempt covered all 9 nuggets, whereas GLM 4.5 Air never searched by name for seizure-related fractures, medication-induced bone loss, or genetic bone disorders.A second KARL attempt covered 7/9 nuggets, missing only genetic disorders and malignancy.Source: §E.6 · pp. 68–69

E.7 Case Study: Parallel Thinking Identifies and Merges the Best of Candidate Solutions

  • The aggregation rollout produced a response better than the best candidate by combining complementary information for nugget 5 and surfacing nugget 8, found by only 40% of rollouts.Rollouts 1 and 5 retrieved complementary information for nugget 5, which aggregation combined into a supported nugget.Source: §E.7 · p. 70, Figure 40 · p. 70

F Categorizing Search Behavior … G.2 Context Compression Case Studies

The section describes a rule-based classifier for search behavior and reports how often KARL compresses context on BrowseComp-Plus. It also introduces two trajectories showing compression’s effect on downstream performance.

F Categorizing Search Behavior

  • The rule-based search-behavior classifier uses trace features to identify truncated responses and expressed uncertainty.Truncation is detected when a response is cut off; uncertainty is identified by matching a small set of phrases.Source: F Categorizing Search Behavior · p. 71
  • The classifier detects proposed answers in stepwise reasoning and uses a prompt to identify searches intended to verify an already proposed answer.These categories can apply before the model generates its final answer.Source: F Categorizing Search Behavior · p. 71
  • The taxonomy was calibrated on 30 hand-labeled diverse rollouts, then applied to two trajectories per query across 230 BrowseComp-Plus test queries for three models.Borderline traces were manually annotated, and human annotations agreed with the developed rules roughly 75% of the time.Source: F Categorizing Search Behavior · p. 71

G.1 Compression Statistics

  • Median context compression was 6 steps per question and mean was 10.2; 20% of questions required at least 20 steps.The distribution is right-skewed, with a long tail extending to 40 compressions; roughly 80% of questions were resolved with fewer than 20.Source: Figure 41 · p. 72, §G.1 · p. 72
  • KARL compresses accumulated retrieved documents, search results, and reasoning into a condensed context when its context window fills.The concentration of many compression events among a small fraction of hard questions means context-management cost is concentrated in that tail.Source: §G.1 · p. 72

G.2 Context Compression Case Studies

  • Two BrowseComp-Plus trajectories illustrate compression’s effect on downstream performance, with the critical difference being what information the compressor preserves.In both cases, context exceeded 100K characters across four search steps before compression to roughly 1K characters, a ∼100× reduction.Source: §G.2 · p. 72

G.2.1 Successful Compression: Author Identification · G.2.2 Harmful Compression: ICC Hall of Fame Puzzle

The examples show compression succeeding when an author-identification search had converged, but harming an unfinished multi-hop cricket deduction by turning a provisional year into an anchor. The successful answer was “Baby”; the failed answer was “2009” instead of 2010.

G.2.1 Successful Compression: Author Identification · G.2.2 Harmful Compression: ICC Hall of Fame Puzzle

  • Compression returned the correct “Baby” answer after the author search converged, but returned “2009” instead of 2010 after collapsing an unfinished cricket deduction.Both trajectories used 100× compression; the author-identification trajectory was marked correct and the ICC puzzle incorrect.Source: §G.2.1 · pp. 72, 74, §G.2.2 · pp. 74, 76

G.2.1 Successful Compression: Author Identification

  • Successful compression preserved the converged answer and its supporting findings: author, city, renamed title, jobs, and hobbies.The passage attributes success to compression capturing the key findings after pre-compression reasoning had already converged on one answer.Source: §G.2.1 · p. 74

G.2.2 Harmful Compression: ICC Hall of Fame Puzzle

  • The ICC summary retained Kapil Dev’s 2009 induction but discarded the active deduction that the three-year offset pointed to 2012.The search was still pivoting toward 2012 when compression occurred, so 2009 was a clue rather than a settled answer.Source: §G.2.2 · p. 75
  • Post-compression, the model repeated Kapil Dev’s and Clive Lloyd’s 2009 induction years without revisiting the offset logic.It later reinforced “Kapil Dev (2009)” through a second compression and never reconsidered whether the base year was 2009 or 2010.Source: §G.2.2 · pp. 75–76
  • A retrieved document stated that Kapil Dev was inducted in 2010, but the model did not notice the contradiction to its 2009 anchor.The model’s final answer remained 2009, which the passage identifies as incorrect; the correct answer was 2010.Source: §G.2.2 · p. 76

H Evaluation Infrastructure

The evaluation infrastructure combines side-by-side model comparisons on real-world queries with tooling for analyzing model search traces.

H Evaluation Infrastructure

  • The Databricks app compares models and search agents on the same query, showing streaming outputs, tool calls, and retrieval citations in one view.An optional blind mode hides model identities and citations until users record a preference; runs can be shared for discussion. The app uses aroll and PMBench to evaluate single-turn question answering and deep-research over unstructured corpora.Source: H Evaluation Infrastructure · p. 77, Figure 42 · p. 77
  • Lightweight trace-analysis tooling lets researchers compare model behavior per question and inspect time spent grounding, exploring, and verifying.The tooling supports qualitative analysis of search efficiency and differentiated search strategies; models may spend many steps on verification.Source: Figure 43 · p. 77, H Evaluation Infrastructure · p. 77

Related papers

Loading 2603.05218v1…