PaperlayerView exact revision on arXiv

Source-linked AI summary

Memento No More: Coaching AI Agents to Master Multiple Tasks via Hints Internalization

Minttu Alakuijala, Ya Gao, Georgy Ananov, Samuel Kaski, Pekka Marttinen, Alexander Ilin, Harri Valpola

arXiv:2502.01562v3cs.LG

TL;DR

Prompt-based agents can depend on expanding prompts to handle multiple tasks, with performance degrading as prompts grow. This paper iteratively distills human-provided hints into agent weights, and its three-round Llama-3.1-70B agent achieves higher reported success rates than GPT-4o and DeepSeek-V3 on ToolQA and OfficeBench.

  • Problem

    Prompt-based agents can require long, exhaustive prompts for diverse tasks, and performance can degrade as prompts grow.

    Source: §1 · p. 2

  • Method

    The method iteratively distills initial and corrective human-provided hints into model weights so agents can improve over successive training rounds.

    Source: §1 · p. 2, §3.1 · pp. 3–4

  • Results

    After three rounds, Llama-3.1-70B achieves 97.9% success on ToolQA and 90.3% on OfficeBench, versus GPT-4o’s 92.8% and 89.9%, respectively.

    Source: §6 · p. 10

  • Takeaways & Limitations

    The experiments show that the trained agents can perform without task-specific guidance in their prompts while achieving the reported benchmark success rates.

    Source: §1 · p. 2, §6 · p. 10

    Training was evaluated within single domains, cross-domain continued training was not considered, and the protocol still depends on human hints that may be hard to scale.

    Source: §5.5 · p. 10

Abstract

from arXiv · show

As the general capabilities of artificial intelligence (AI) agents continue to evolve, their ability to learn to master multiple complex tasks through experience remains a key challenge. Current LLM agents, particularly those based on proprietary language models, typically rely on prompts to incorporate knowledge about the target tasks. This approach does not allow the agent to internalize this information and instead relies on ever-expanding prompts to sustain its functionality in diverse scenarios. This resembles a system of notes used by a person affected by anterograde amnesia, the inability to form new memories. In this paper, we propose a novel method to train AI agents to incorporate knowledge and skills for multiple tasks without the need for either cumbersome note systems or prior high-quality demonstration data. Our approach employs an iterative process where the agent collects new experiences, receives corrective feedback from humans in the form of hints, and integrates this feedback into its weights via a context distillation training procedure. We demonstrate the efficacy of our approach by implementing it in a Llama-3-based agent that, after only a few rounds of feedback, outperforms advanced models GPT-4o and DeepSeek-V3 in tasksets requiring correct sequencing of information retrieval, tool use, and question answering.

1 Introduction

Prompt-based agents can require long, exhaustive prompts to handle diverse tasks, and performance can degrade as those prompts grow. The paper proposes iteratively internalizing human-provided hints into model weights, reducing reliance on task-specific prompt guidance.

1 Introduction

  • As prompts accumulate hints for diverse tasks, agents must analyze them exhaustively at each action, and performance can degrade as prompts grow.The passage also notes that Transformer computational cost typically scales quadratically with prompt length.Source: §1 · p. 2
  • Iterative coaching internalizes human-provided hints into model weights, and the agent progressively improves over three rounds without task-specific prompt guidance.After each training round, a human observes the agent, provides hints relevant to particular execution steps, and the agent further internalizes them.Source: §1 · p. 2
  • The introduction frames agent coaching as a response to the need for agents to master diverse tasks and learn from mistakes.The proposed iterative process is illustrated in Figure 1.Source: Figure 1 · p. 1, §1 · p. 2

2 Agent architecture

The agent uses a ReAct-style loop in which reasoning traces and executable code build a trajectory, with execution output returned as the next observation. It invokes tools through Python functions and submits its final report and answer through a required function call.

2 Agent architecture

  • The agent’s actions combine a reasoning trace and executable code, while execution output becomes the next observation.The current trajectory forms the prompt at each step.Source: §2 · pp. 2–3
  • Figure 3 contrasts initial hint distillation with later refinement using human feedback and mistake-specific corrective hints.The figure describes the student learning teacher outputs while not seeing the hints itself.Source: Figure 3 · p. 3
  • Tasks invoke tools as Python functions, and every execution ends with complete_task(report, answer) to submit the report and requested answer.After Round 1, tool descriptions are removed from the hints as the agent internalizes that knowledge.Source: §2 · p. 3

3 Agent training

The training procedure first distills initial task hints into the agent’s weights, then uses observed mistakes and corrective hints to refine behavior in later rounds. Its iterative, hint-based supervision resembles DAgger while avoiding the need for explicit expert demonstrations.

3.1 Round 1: Internalizing initial hints on tools and other instructions for novel tasks

  • Initial hints are developed by observing mistakes under minimal guidelines and adding only the guidance needed for successful task performance.Hints can include tool-use examples, step-by-step strategies, and inner-monologue recommendations; the paper denotes the initial hints h1.Source: §3.1 · p. 3
  • Round 1 collects state-action-hint trajectories, then trains a student to mimic the hint-conditioned teacher without seeing those hints in its prompt.The student is trained by minimizing KL divergence between token output distributions; initial hints are dropped from its prompt with probability 0.9 to help preserve general skills.Source: §3.1 · p. 4

3.2 Round 2, . . . , I: Internalizing corrective hints

  • Later rounds inspect the current agent’s trajectories, identify mistakes, and supply corrective hints for problematic states before further distillation.The hints are targeted to states with mistakes, and datasets are balanced to represent different task and mistake types.Source: §3.2 · pp. 4–5

3.3 Connection to imitation learning

  • Unlike DAgger’s expert demonstrations, the method uses an LLM empowered by human-written hints to produce target signals and token-level output distributions.The paper describes this as an iterative process that observes learner mistakes and addresses them with corrective hints.Source: §3.3 · pp. 5–6

4 Related work

Related work improves LLM reasoning through stepwise decomposition, code, and reflection, and uses human feedback and interactive prompts to train or guide models. The paper distinguishes its approach from these methods by extending context distillation to interactive agents.

Reasoning with language models

  • Prior work elicits reasoning through unit-step decomposition, structured code, and self-critique, but self-correction is bounded by the model’s ability to recognize its mistakes.The paper instead uses human feedback for more accurate corrections.Source: §4 · p. 6

Human feedback in LLM training

  • Human-feedback frameworks have trained LLMs using ranked answers or lists of rules the model should follow.Source: §4 · p. 6

Interactive LLM frameworks

  • Interactive LLM agents commonly place tools and interaction history in prompts without changing model weights, while prompt-only knowledge can cause information overload.Other approaches optimize the agent framework or use retrieval-augmented generation.Source: §4 · p. 6

Supervised fine-tuning for interactive settings

  • Supervised fine-tuning for interactive settings often depends on high-quality human-generated trajectories, which are difficult to scale and cap performance at annotator skill.The paper also compares against filtered synthetic trajectories; related corrective-feedback work for VLM agents uses abstractions rather than trajectories.Source: §4 · p. 6

Knowledge distillation

  • Unlike knowledge distillation from a more capable teacher, context distillation uses the same model as teacher and student with different inputs; this work extends it to interactive LLM agents.Knowledge distillation requires a more capable model, while context distillation gives the teacher additional context unavailable to the student.Source: §4 · p. 6

5 Experiments with multi-task agents

The experiments test whether agents can handle diverse ToolQA and OfficeBench tasks without task-specific information at test time. After iterative hint internalization, the agents reach strong success rates with lower inference costs, while the study also examines generalization and scope limitations.

5.1 Evaluation benchmarks

  • The benchmarks test tool use, information retrieval, and question answering across 1,076 ToolQA tasks and 286 OfficeBench workflows.ToolQA has 100 question templates; OfficeBench tasks are unique and span six applications.Source: §5.1 · p. 7

5.2 Baseline agents

  • Combining task-specific hints reduced success from 96.9% to 88.4% on ToolQA and from 89.1% to 81.3% on OfficeBench, a pattern termed the Memento effect.GPT-4o reached 92.8% on ToolQA and 89.9% on OfficeBench, while remaining below task-filtered Llama on ToolQA.Source: §5.2 · p. 7

5.3 Training a multi-task agent

  • After two rounds, hint-free agents reached 95.7% on ToolQA and 87.3% on OfficeBench; Round 3 reached 97.9% and 90.3%, respectively.Corrective hints target recurring errors detected in trajectories; Round 3 used 18 ToolQA and 12 OfficeBench filters.Source: §5.3 · p. 9
  • 97.9% average success on ToolQA and 90.3% on OfficeBench; Round 3 agents also run 3–4x faster and use 7–10% of other models’ input tokens.Training costs are amortized after 255–285 task executions in either domain.Source: Table 2 · p. 8, Table 4 · p. 8, §5.3 · p. 9

5.4 Additional experiments

  • On held-out ToolQA templates, Round 3 outperformed DeepSeek-V3 and narrowly trailed GPT-4o; standard coding and math benchmarks showed no signs of degradation.The SFT-loss alternative performed slightly worse than the proposed training procedure.Source: §5.4 · pp. 9–10

5.5 Discussion and limitations

  • The experiments train within one domain at a time and rely on human hints, although 32 ToolQA and 24 OfficeBench corrective hints sufficed.The authors also caution that real-world deployment may have unintended effects and recommend context-appropriate guardrails.Source: §5.5 · p. 10

6 Conclusions

After three rounds of human mistake correction, the Llama-3.1-70B agent surpasses GPT-4o and DeepSeek-V3 on ToolQA and OfficeBench. The broader-impact discussion identifies possible benefits and risks of autonomous agents and recommends oversight, logging, and deployment guardrails.

6 Conclusions

  • 97.9% ToolQA and 90.3% OfficeBench after three rounds, exceeding GPT-4o (92.8% and 89.9%) and DeepSeek-V3 (87.5% and 86.9%).The method internalizes human-provided hints into model weights through iterative training.Source: §6 · p. 10
  • Weight-level hint internalization reduces reliance on increasingly complex prompts while training agents to master multiple tasks.The approach uses iterative training with human-provided hints.Source: §6 · p. 10

Broader Impact Statement

  • AI agents could accelerate knowledge-work automation, affect employment, raise transparency and accountability questions, and enable harmful uses such as cyberattacks.The paper anticipates some routine knowledge work may shift toward guiding how problems are solved, while potentially affecting job numbers.Source: Broader Impact Statement · p. 11
  • Internalized knowledge is not directly interpretable or externally auditable, making it harder to verify which hints or constraints an agent follows at a given step.The paper recommends monitoring interfaces, intervention mechanisms, and transparent action logs.Source: Broader Impact Statement · p. 11
  • The authors recommend setting bounds for AI agency and tailoring deployment guardrails to context, including restricting unnecessary access to the internet or file system.They also recommend human review protocols for critical decisions such as purchases.Source: Broader Impact Statement · p. 11
  • Comprehensive trajectory logs—including reasoning traces, actions, and observations—can support diagnosis of unintended behaviors that may have been internalized during training.The paper also cites generated reasoning traces as providing a degree of built-in interpretability.Source: Broader Impact Statement · p. 11

A Limitations

The limitations section identifies assumptions about task overlap and failure detection, and notes that hint-based supervision can be difficult to scale. It also reports annotation effort, discusses possible oversight reductions, and summarizes results on coding and reasoning benchmarks and a cross-entropy variant.

A Limitations

  • The method’s transfer is most effective when tasks share tools; training also benefits from knowing which tools are unnecessary for a task.The authors note that these assumptions are not requirements for the method.Source: A Limitations · p. 15
  • The experiments assume training-task outputs can be checked against ground-truth answers; without them, failures may be identified through execution signals and manual labeling.The suggested signals are nonterminating trajectories and code-action errors.Source: A Limitations · p. 15
  • Human supervision through hints can be difficult to scale, although annotation effort for Rounds 2 and 3 was approximately 8 hours for ToolQA and 10 hours for OfficeBench.The pipeline reuses hints for similar failures and lets humans review subsets of detected failures.Source: A Limitations · p. 15
  • Further oversight reduction is left for future work; proposed automation depends on access to a capable external model or informative environment feedback.The method is designed to remain applicable when neither condition is guaranteed.Source: A Limitations · p. 15

B.1 Coding and reasoning benchmarks

  • The authors report no performance degradation on HumanEval and GSM8K after training the agents.The cited table reports backbone LLM performance on coding and reasoning benchmarks.Source: Table 6 · p. 16

B.2 Cross-entropy loss

  • The cross-entropy comparison evaluates two distillation variants, including KL training on all trajectories.The supplied table caption identifies the comparison but does not report its outcome.Source: Table 7 · p. 16

B.3 Ablations

The ablations favor a 0.9 hint-dropout rate on both benchmarks and find statistically significant gains from data balancing. The authors attribute potential benefits of dropout to continued attention to new information, data augmentation, and reduced pressure to internalize all guidance at once.

B.3 Ablations

  • A 0.9 hint-dropout rate yields the highest average success rates on ToolQA and OfficeBench among tested rates 0.5, 0.7, and 1.0.The ablations compare dropout rates on both benchmarks.Source: §B.3 · p. 16, Table 8a · p. 17, Table 8b · p. 17
  • The authors suggest hint dropout may aid attention to new prompt information, vary training samples across epochs, and reduce spurious-correlation memorization.They attribute the latter benefit to focusing on a subset of hints at a time.Source: §B.3 · p. 16
  • Data balancing significantly improves success rates, by 1.4 percentage points on ToolQA and 1.5 on OfficeBench.The gains are reported in the data-balancing ablations.Source: §B.3 · p. 16, Table 8c · p. 17, Table 8d · p. 17

B.4 Cross-domain transfer

In a held-out ToolQA experiment, hints distilled on two tabular domains transfer positively to Airbnb and Coffee, while transfer to Flight is more challenging. A separate experiment reports nearly identical average ToolQA success rates with semantically rephrased hints.

B.4 Cross-domain transfer

  • Hints distilled from one pair of tabular domains transfer positively to held-out Airbnb and Coffee, with a significant +26.4% to Airbnb.Table 9 compares agents trained on two source domains against zero-shot performance on held-out domains.Source: §B.4 · pp. 17–18, Table 9 · p. 18
  • Transfer to Flight is more challenging, likely because flight-specific reasoning such as parsing identifiers and calculating delays needs targeted experience.The authors say the results support generalizability within a shared tool ecosystem.Source: §B.4 · pp. 17–18

B.5 Sensitivity to hint phrasing

  • Average success rates across ToolQA domains are nearly identical when Round 1 uses semantically rephrased rather than original hints.The comparison is reported for ToolQA Round 1.Source: §B.5 · p. 18, Table 10 · p. 18

C Inference efficiency

Round 3 MNM agents run faster than the base Llama-3.1-70B agent with combined hints, with estimated sequential break-even after 285 ToolQA or 255 OfficeBench tasks. The appendix describes the training adapters, round-specific datasets, and computational costs.

C Inference efficiency

  • MNM Round 3 runs ToolQA tasks in approximately 90 seconds versus 340 seconds for Llama-3.1-70B with combined hints, and OfficeBench tasks in 140 versus 600 seconds.The authors calculate a 3–4x end-to-end inference speedup on four AMD MI250X GPUs.Source: C Inference efficiency · p. 19
  • Sequential inference savings offset training after 285 ToolQA task executions or 255 OfficeBench executions, under the stated hardware and time assumptions.The authors describe these break-even estimates as conservative because batched inference can reduce wall-clock break-even time.Source: C Inference efficiency · p. 19

D Training details

  • Each training round adds a full-precision rank-128 LoRA adapter across linear layers, merges it into the base weights, and adds another adapter for the next round.The rank was the largest trainable on the authors’ single-node setup.Source: D Training details · p. 19

D.1 ToolQA

  • ToolQA Round 3 combines 120 corrected transitions with 69 new full trajectories to address small-state and task-coverage imbalance, yielding 359 training samples.The agent is then trained for six epochs.Source: §D.1 · p. 20

D.3 Computational cost

  • Training takes 16, 1.25, and 2.5 hours across ToolQA rounds, and 27, 3.5, and 2 hours across OfficeBench rounds.All rounds use four AMD MI250X GPUs, totaling eight GPU chips.Source: §D.3 · p. 20

D.4 Human annotation cost

The annotation-cost estimate is about 19 minutes per finalized hint, assuming prior task-domain familiarity. Estimated totals are 7.6 hours for OfficeBench and 10.1 hours for ToolQA under an asynchronous-inference assumption.

D.4 Human annotation cost

  • 19 minutes is the approximate time to finalize each hint, assuming prior familiarity with the task domain.The estimate includes reviewer-prompt and hint refinement.Source: §D.4 · p. 20
  • On average, finalizing each identified mistake type required 0.5 prompt-refinement and 0.5 hint-refinement iterations.Outputs were refined until filtering accuracy and automatically evaluated mistake-correction rates were deemed sufficient.Source: §D.4 · p. 20
  • 7.6 hours for 24 OfficeBench hints and 10.1 hours for 32 ToolQA hints are the estimated annotation totals.The estimate assumes LLM inference can run asynchronously from human annotation time.Source: §D.4 · p. 20

E Reviewer evaluation

The reviewer evaluation reports false-positive and false-negative rates for automated mistake filters across hint types and corrective-training rounds. Its tables cover ToolQA and OfficeBench, with script-based filters applicable only in some settings.

E Reviewer evaluation

  • The pipeline’s mistake-localization accuracy is evaluated across all hint types in both rounds of corrective training.The evaluation reports false-positive and false-negative rates for the filters.Source: E Reviewer evaluation · p. 21
  • FPR measures correct trajectories incorrectly flagged as mistaken, while FNR measures mistake-containing trajectories missed by a filter.Source: E Reviewer evaluation · p. 21
  • ToolQA filter-rate tables cover Rounds 2 and 3, while OfficeBench reports average rates; script-based filters apply to only some ToolQA mistake types and none in OfficeBench.For ToolQA Round 3, script-based filters apply to only 2 of 20 mistake types.Source: Table 13 · p. 21, Table 14 · p. 22, Table 15 · p. 22

F Example tasks

The example-task tables show ToolQA question templates and OfficeBench tasks alongside possible solution strategies.

F Example tasks

  • Table 16 presents example ToolQA question templates and possible solution strategies.Source: Table 16 · p. 23
  • Across the two benchmarks, the tables provide examples of tasks and possible approaches to solving them.Source: Table 16 · p. 23, Table 17 · p. 23
  • Table 17 presents example OfficeBench tasks and possible solution strategies.Source: Table 17 · p. 23

G Data balancing strategy

This section describes a three-step strategy for balancing training trajectories across task groups and failure cases, and documents tools for ToolQA and OfficeBench, including OfficeBench functions for converting Excel and Word files to PDF and creating, reading, and writing Word files.

G Data balancing strategy

  • The three-step balancing procedure supplements sparse failure cases, increases representation of underrepresented or repeatedly failing task groups and applications, and samples uncovered tasks to preserve learned skills.For sparse mistakes, it adds a trajectory matching the ToolQA question template or OfficeBench applications; preservation sampling is sized to balance states from failures against trajectories added through balancing.Source: G Data balancing strategy · p. 24

H Tool documentation

  • Tool documentation gives each tool a one-line description, input-argument explanations, output descriptions, and one or two example calls.Source: H Tool documentation · p. 24

H.1 ToolQA tools

  • ToolQA documentation covers graph queries, database loading and filtering, agenda retrieval, and task completion functions.The graph tools expose node and edge information, while database tools filter rows and return column values.Source: §H.1 · pp. 25–26

H.2 OfficeBench tools

  • OfficeBench documentation describes tools for email, OCR, calendar, PDF, and Excel operations, including their inputs, outputs, and example calls.The documented operations include reading and sending email, recognizing image text, managing calendar events, converting or reading files, and editing spreadsheets.Source: §H.2 · pp. 27–29
  • The Excel conversion tool takes source and destination paths and returns whether conversion succeeds.The documented function is excel_convert_to_pdf.Source: §H.2 · p. 29
  • Word tools create a file and read its text content, using a file path as input.Creation returns success or failure; reading returns the document text.Source: §H.2 · p. 29
  • The Word writing tool takes a file path and text and returns whether the text was written successfully.The documented function is word_write_to_file.Source: §H.2 · p. 30
  • The Word-to-PDF conversion tool takes source and destination paths and returns whether conversion succeeds.The documented function is word_convert_to_pdf.Source: §H.2 · p. 30

I.1.1 Task split

This section describes the curation and prompting of ToolQA and OfficeBench tasks, including task-specific corrective hints across multiple rounds. It also presents examples of hints for file handling, tool use, and task requirements, along with an LLM reviewer prompt for judging whether a final step answers the task.

I.1.1 Task split

  • ToolQA uses a default task-based split in which each question template appears in every split, and an unseen-template split that holds out templates for evaluation.Table 5 uses the unseen-template split; the other tables use the default split.Source: §I.1.1 · p. 30

I.1.2 Task selection

  • The authors focus on six ToolQA task groups, excluding GSM8K because Python handles arithmetic and SciREX because retrieval is unreliable and overlaps with Agenda.SciREX documents were not consistently retrieved across human-written queries, preventing fair and accurate evaluation.Source: §I.1.2 · p. 30
  • DBLP curation removes mismatched index entries and tasks without reliably retrievable answers, then adds tasks to retain 168 tasks across 15 templates; Agenda retains 140 solvable tasks across 7 templates.DBLP’s additional tasks follow the official dataset generation process.Source: §I.1.2 · pp. 30–31
  • Yelp’s ambiguous or incomplete tasks are removed and distance-calculation errors corrected, while Airbnb’s ambiguous tasks and incorrect ground truths are fixed or removed.The curated sets contain 180 Yelp tasks across 20 templates and 168 Airbnb tasks across 17 templates.Source: §I.1.2 · p. 31
  • Flight descriptions are corrected from a 500-mile to a 300-mile threshold, and Coffee questions and answers are revised to fix ambiguity and missing data.The resulting sets contain 200 Flight tasks across 20 templates and 230 Coffee tasks across 21 templates.Source: §I.1.2 · p. 31

I.2 OfficeBench

  • OfficeBench task review found faulty files, evaluation errors, mismatched descriptions and answers, ambiguity, and tool limitations; after corrections and removals, 286 of 300 tasks remained and were split 67:3:30.Each task requires 1 to 3 applications; uncorrectable file or tool problems led to removal, while fixable descriptions and evaluation conditions were revised.Source: §I.2 · p. 31

J Example agent prompt

  • The example DBLP prompt asks the agent to retrieve authors using provided tools and return a string containing only the answer.Its guidelines direct the agent to use one tool per step, inspect outputs, and check the final answer’s formatting.Source: J Example agent prompt · p. 32

K.1 Round 1 hints

  • Round 1 hints are added after annotators inspect attempts: minor mistakes receive targeted instructions, while failures receive step-by-step guidance.Shared and task-specific hints cover database selection and filtering, query and tool use, answer formatting, and OfficeBench file inspection.Source: §K.1 · pp. 33–35
  • Round 1 OfficeBench hints emphasize correct file and tool workflows, such as listing directory contents before choosing a file and verifying file contents before finishing.Other examples require writing content into newly created Word or Excel files with the designated write tools.Source: §K.1 · pp. 35–36

K.2 Round 2 hints

  • Round 2 and 3 corrective hints appear only in specific states where the agent makes each mistake, localized by the LLM reviewer; examples prescribe FlightData filtering and distance counting, and manual calendar-window reasoning.Source: §K.2 · pp. 36–38
  • Round 2 hints also require manual file inspection and exact task-compliant file creation, using supported tools rather than programmatic parsing or invented functions.For example, Excel content should be read as a human would, and a PDF should be created by writing a Word file and converting it.Source: §K.2 · pp. 38–39

K.3 Round 3 hints

  • Round 3 hints specify task-dependent ToolQA operations, including filtering shared-room listings by location and price, removing missing airtime values, and calculating flight measures.For flight-time questions, the flight identifier is split into its airline code and flight number; the paper’s citation-count attribute is defined as the number of papers that cited it.Source: §K.3 · pp. 40–41
  • Other hints stress exact task requirements: use the stated meeting window, preserve requested file names and content, and set calendar times from task-provided dates and times.When emailing file contents, include available details in the body rather than only saying the file is attached.Source: §K.3 · pp. 41–42
  • Hints require following task-specific procedures, including filtering reviews by NAME, neighbourhood, and id and reading every email individually rather than assuming its order.The email hint says to list email IDs, then call read_email() separately for each one without loops.Source: §K.3 · pp. 41–42
  • Round 3 hints require manually inspecting file contents, printing stored or retrieved values, and choosing the specified reader for each file type.For PDFs and Excel files, the hints specify pdf_read_file() and excel_read_file(); image text should be read with ocr_recognize_text().Source: §K.3 · p. 42
  • Implementation hints specify keeping data_path read-only, using the stated Excel and Word function arguments, and printing retrieved delay values and document contents.The hints say excel_set_cell() takes four arguments and word_write_to_file() takes two.Source: §K.3 · pp. 41–42

L Example prompt for the LLM-based reviewer

  • The example reviewer prompt marks an agent wrong when its final step fails to directly return the requested citation count, and judges the last step using earlier steps only as context.The reviewer uses the trained Llama-3.1-70B teacher model to avoid serving multiple 70B models during training-data generation.Source: L Example prompt for the LLM-based… · p. 43

Related papers

Loading 2502.01562v3…