PaperlayerView exact revision on arXiv

Source-linked AI summary

TICKing All the Boxes: Generated Checklists Improve LLM Evaluation and Generation

Jonathan Cook, Tim Rocktäschel, Jakob Foerster, Dennis Aumiller, Alex Wang

arXiv:2410.03608v1cs.AIcs.CLcs.HCcs.LG

TL;DR

Instruction-following evaluations need to be reliable and interpretable, but preference labels and direct scores can obscure reasoning and yield disagreements. TICK uses generated instruction-specific checklists, and STICK uses checklist-based self-evaluation for refinement and selection. TICK raises exact agreement with human preferences, while STICK improves generation across several benchmarks.

  • Problem

    Common instruction-following evaluations, including preference judgments and direct scores, can obscure evaluation reasoning and produce disagreement among annotators and between models and humans.

    Source: §1 · p. 1

  • Method

    TICK uses LLM-generated instruction-specific YES/NO checklists to evaluate responses, and STICK uses TICK self-evaluations for iterative refinement and response selection.

    Source: §1 · pp. 1–2

  • Results

    TICK increases exact agreement between an LLM judge and human preferences by 5.8%, while STICK improves performance across several benchmarks through self-refinement and Best-of-N selection.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Fully automated TICK is cheaper and faster than existing checklist-based evaluations and can be applied to arbitrary instruction-following datasets.

    Source: §1 · p. 2

Abstract

from arXiv · show

Given the widespread adoption and usage of Large Language Models (LLMs), it is crucial to have flexible and interpretable evaluations of their instruction-following ability. Preference judgments between model outputs have become the de facto evaluation standard, despite distilling complex, multi-faceted preferences into a single ranking. Furthermore, as human annotation is slow and costly, LLMs are increasingly used to make these judgments, at the expense of reliability and interpretability. In this work, we propose TICK (Targeted Instruct-evaluation with ChecKlists), a fully automated, interpretable evaluation protocol that structures evaluations with LLM-generated, instruction-specific checklists. We first show that, given an instruction, LLMs can reliably produce high-quality, tailored evaluation checklists that decompose the instruction into a series of YES/NO questions. Each question asks whether a candidate response meets a specific requirement of the instruction. We demonstrate that using TICK leads to a significant increase (46.4% $\to$ 52.2%) in the frequency of exact agreements between LLM judgements and human preferences, as compared to having an LLM directly score an output. We then show that STICK (Self-TICK) can be used to improve generation quality across multiple benchmarks via self-refinement and Best-of-N selection. STICK self-refinement on LiveBench reasoning tasks leads to an absolute gain of $+$7.8%, whilst Best-of-N selection with STICK attains $+$6.3% absolute improvement on the real-world instruction dataset, WildBench. In light of this, structured, multi-faceted self-improvement is shown to be a promising way to further advance LLM capabilities. Finally, by providing LLM-generated checklists to human evaluators tasked with directly scoring LLM responses to WildBench instructions, we notably increase inter-annotator agreement (0.194 $\to$ 0.256).

1 INTRODUCTION

The introduction presents TICK as an automated checklist-based approach to more interpretable LLM evaluation, then extends it as STICK for self-improvement and human evaluation support.

1 INTRODUCTION

  • TICK increases exact agreement between an LLM judge and human preferences by 5.8%.The checklist evaluation protocol decomposes instructions into YES/NO questions about specific response requirements.Source: §1 · pp. 1–2
  • STICK improves self-refinement across benchmarks and outperforms vanilla Self-Refine, while Best-of-N selection beats greedy decoding and a general-purpose reward model.Command-R+ gains 6.5% on InFoBench and 7.1% on WildBench; Best-of-N improves on greedy decoding by 5.1% on InFoBench and 5.3% on WildBench.Source: §1 · p. 2
  • LLMs generate checklists similar in quality to those written by trained human annotators across diverse instruction-following datasets.Source: §1 · p. 2
  • TICK is fully automated, cheaper and faster than existing checklist-based evaluations, and applicable to arbitrary instruction-following datasets.Source: §1 · p. 2
  • Providing LLM-generated checklists to human evaluators tasked with directly scoring responses improves inter-annotator agreement.Source: §1 · p. 2

2 RELATED WORK

Prior work evaluates instruction-following with broad benchmarks, judge LLMs, curated checklists, and language critiques. The paper contrasts these approaches with dynamic checklists intended for on-the-fly evaluation and structured refinement.

Instruction-Following Evaluation

  • Instruction-following benchmarks use diverse instruction sets and judge LLMs, while some checklist-based approaches rely on human curation that limits use beyond their benchmarks.Source: §2 · p. 3
  • The paper positions dynamic checklists as usable in custom workflows and as structured feedback for generation, without human checklist curation.Source: §2 · p. 3

Language Critiques

  • Language critiques can support evaluation, human annotation, and in-context refinement, but naive self-correction can degrade performance.Source: §2 · p. 3

3 TICK: TARGETED INSTRUCT-EVALUATION WITH CHECKLISTS

TICK evaluates instruction-following by decomposing instructions into targeted YES/NO questions and evaluating responses against each requirement. The section validates checklist quality and question answering, then compares TICK’s agreement with human preferences against other LLM-as-judge protocols.

3.2.1 GENERATING CHECKLISTS

  • GPT-4o and Llama3.1-70B checklists more closely match human ground-truth checklists than alternate human-written checklists, while GPT-4o has the strongest pass-rate correlation with human checklists.Command-R+ has the lowest string-level similarity but is close in question count.Source: Table 2 · p. 4, Table 3 · p. 5, §3.2.1 · p. 5

3.1.1 GENERATING CHECKLISTS

  • For each instruction, an LLM generates YES/NO questions targeting distinct explicit and implicit requirements, with YES indicating that the response meets the requirement.The questions are tailored to the instruction, and different LLMs may generate and answer the checklist.Source: §3.1.1 · p. 4

3.1.2 USING CHECKLISTS

  • TICK scores a response using checklist Pass Rate and aggregates checklist answers into an instruction-following measure across a dataset.Pass Rate is the fraction of checklist questions answered YES; DRFR measures the percentage of total checklist questions correctly answered across examples.Source: §3.1.2 · p. 4

3.2.2 USING CHECKLISTS

  • GPT-4o achieves the strongest checklist-answer accuracy among the evaluated LLMs, and CoT followed by majority voting further improves accuracy.The majority-vote setting samples k CoT evaluations; higher k yields a more substantial increase.Source: §3.2.2 · p. 5, Figure 2 · p. 6
  • TICK achieves the lowest overall WPLD against human preferences, while Check-then-Score outperforms direct scoring but trails TICK.TICK is also the only LLM-as-judge evaluation to achieve PLD 0 more often than not.Source: Table 4 · p. 6, §3.2.2 · p. 6

4 IN-CONTEXT SELF-IMPROVEMENT WITH STICK (SELF-TICK)

STICK uses checklist-based self-evaluations to guide response refinement and select among candidate responses. The section evaluates both strategies against baselines on instruction-following benchmarks.

4.1 SELF-REFINEMENT WITH CHECKLISTS AS FEEDBACK

  • Self-refinement generates a response, evaluates it against an LLM-generated checklist, then uses failed checks as actionable feedback for optional refinement iterations.The validation compares STICK with unstructured detailed critiques and evaluates Command-R+ over four iterations on InFoBench and WildBench, using GPT-4 as judge.Source: §4.1 · p. 7, §4 · p. 7
  • STICK improves self-refined responses over Self-Refine by 6.5% on InFoBench and 7.1% on WildBench, but quality plateaus or regresses by the fourth iteration.InFoBench uses a human-written checklist not seen during refinement, while WildBench uses a holistic response score.Source: §4.1 · p. 7
  • On LiveBench, one STICK refinement improves responses, but further refinement degrades them; Command-R+ benefits particularly on reasoning tasks.Both Command-R+ and GPT-4o benefit considerably on explicit instruction-following, while unstructured self-critiques immediately degrade performance in most categories.Source: §4.1 · p. 8

4.2 BEST-OF-N SELECTION WITH CHECKLIST-BASED SCORES

  • For Best-of-N, the generating LLM scores N candidate responses with STICK and selects the highest-scoring response; the comparison includes direct self-scoring and an external reward model.The study uses Command-R+ and the benchmarks’ standard evaluation metrics.Source: §4.2 · p. 8
  • With N = 8, STICK Best-of-N gains +5.1% on InFoBench and +5.3% on WildBench, the largest gains among the compared methods.STICK is also the most precise scoring function against each benchmark’s ground-truth evaluation.Source: §4.2 · p. 8, Table 5 · p. 8

5 ASSISTING HUMAN EVALUATORS WITH GENERATED CHECKLISTS

The study tests whether instruction-specific checklists can make human scoring more cognitively manageable and help evaluators attend to relevant requirements. In a WildBench case study, checklist-first scoring increases agreement while leaving average scores consistent.

5.1 MOTIVATION

  • The study investigates whether LLM-generated checklists improve scoring consistency; checklists partially decompose annotation, making it more cognitively feasible and helping ensure instruction-specific considerations are not missed.Across two evaluation rounds, annotators scored Command-R+ responses to WildBench instructions from 1–5; in one round they answered GPT-4o-generated checklist questions first and used them to inform, not limit, holistic scores.Source: §5.1 · p. 9, §5.2 · p. 9

5.2 CASE STUDY: SCORING RESPONSES TO WILDBENCH INSTRUCTIONS

  • Checklist-first scoring yields stronger inter-annotator agreement than direct scoring, while average Command-R+ scores remain consistent; agreement nevertheless remains low.Responses were triply annotated and agreement was measured with Krippendorff’s alpha.Source: §5.2 · p. 9

6 CONCLUSION

The paper concludes that structured checklists support automated instruction-following evaluation and can be used for self-improvement. It also identifies limits to checklist evaluation and directions for future work.

6 CONCLUSION

  • TICK and STICK support evaluation and self-improvement, respectively, while checklist-based evaluation is only one heuristic for structuring evaluation.The conclusion reports gains from STICK refinement and Best-of-N selection across diverse datasets, as well as improved agreement when human evaluators receive generated checklists.Source: §6 · p. 9

LIMITATIONS AND FUTURE WORK

  • Checklist evaluations add inference cost, may not help in every setting, and relying on LLMs throughout evaluation may propagate or exacerbate their biases.The paper names learned or discovered evaluation structures, fine-tuning on synthetic STICK-selected data, and reward models conditioned on checklist evaluations as future directions.Source: LIMITATIONS AND FUTURE WORK · p. 10

A USING TICK FOR CATEGORICAL EVALUATIONS

The appendix explores grouping checklist questions by capability to use TICK for categorical evaluations. GPT-4o-generated categories identify Length as Command-R+’s weakest capability in the Internal evaluation.

A USING TICK FOR CATEGORICAL EVALUATIONS

B.1 COMPARISON OF HUMAN-WRITTEN AND GENERATED CHECKLIST LENGTHS

Human-written and GPT-4o-generated checklist lengths have similar distributions for Internal instructions, although generated lengths are more concentrated near the mode.

B.1 COMPARISON OF HUMAN-WRITTEN AND GENERATED CHECKLIST LENGTHS

  • Generated checklist lengths have a distribution similar to human-written lengths, but are more peaked near the mode for Internal instructions.Figure 5 shows histograms for both checklist types.Source: §B.1 · p. 16, Figure 5 · p. 16

B.2 RELATIONSHIP BETWEEN CHECKLIST LENGTH AND TASK DIFFICULTY

Checklist length shows no clear relationship with task difficulty on InFoBench or WildBench, based on correlations with Command-R+ scores.

B.2 RELATIONSHIP BETWEEN CHECKLIST LENGTH AND TASK DIFFICULTY

  • −0.001 and 0.028 are the Pearson correlations between generated checklist length and Command-R+ scores on InFoBench and WildBench, respectively.The authors interpret these values as indicating no clear relationship between task difficulty and generated checklist length.Source: §B.2 · p. 16

C ADDITONAL STICK BEST-OF-N RESULTS

The appendix includes examples of instructions, generated checklists, and STICK self-refinement, alongside a Best-of-8 comparison using Llama-3.1-70B.

C ADDITONAL STICK BEST-OF-N RESULTS

D INTERNAL INSTRUCTION SET EXAMPLES

  • Internal instruction examples include requests for a cowboy-style relationship history, a brief correction and summary, and a genre-sorted movie list in JSON.Another example asks for post-2000 activities from a biographical passage.Source: D INTERNAL INSTRUCTION SET EXAMPLES · p. 18

E GENERATED CHECKLIST EXAMPLES

  • A generated checklist example for a circuits theme checks whether projects form a relevant, creative list and can be built at home.The example instruction asks for creative at-home projects related to circuits.Source: INFOBENCH · p. 19

INFOBENCH

  • The appendix shows an Associated Press-style article instruction and a checklist that evaluates style, factual accuracy, attribution, and neutrality.The checklist includes criteria about retaining key details and avoiding editorializing.Source: INFOBENCH · p. 20

LIVEBENCH: CODING

  • The coding example asks for a Python solution continuation that maximizes the frequency of an array element using at most k changes.Its generated checklist checks continuation format, correctness, syntax, and efficiency under the stated constraints.Source: LIVEBENCH: CODING · p. 21

LIVEBENCH: MATH

  • The math example pairs a differentiation instruction with checklist questions about product-rule application, intermediate steps, and final-answer formatting.The displayed checklist also asks about differentiating the individual components.Source: LIVEBENCH: MATH · p. 22

LIVEBENCH: REASONING

  • The reasoning example asks for truth-status judgments about three people and a bold list of yes-or-no answers; its checklist checks reasoning and each answer.The checklist also checks that every question is answered, including an educated guess if uncertain.Source: LIVEBENCH: REASONING · p. 22

F STICK SELF-REFINEMENT EXAMPLES

F.1 WILDBENCH

This section presents checklist-based STICK refinement examples on WildBench and LiveBench and documents the prompts and procedures for checklist generation and evaluation, preference ranking, direct scoring, self-refinement, and human annotation. It also describes annotator feedback on whether checklists made scoring easier.

F.1 WILDBENCH

  • The assessed EA was marked Yes on five of six listed checks, covering complete MetaTrader 5 code, recognized strategies, $10-to-$10,000 growth potential, margin management, and documentation; syntax correctness was marked No.Source: §F.1 · p. 25

F.2 LIVEBENCH

  • LiveBench’s example score rose from 0 for the initial response to 1 for the refined response; the refined response received Yes on all four checklist criteria, covering reasoning, piece count, formatting, and answering.Source: §F.2 · pp. 26–27

G.1 CHECKLIST GENERATION

  • Checklist generation asks for concise, precise, instruction-specific yes/no questions covering clearly relevant criteria, after analyzing the instruction.The requested checklist contains at least two and no more than eight questions.Source: §G.1 · p. 28

G.2 CHECKLIST EVALUATION

  • Checklist evaluation assigns Yes only when the response entirely fulfills a question’s condition; failure or insufficient information receives No.Even a minor inaccuracy disqualifies a Yes judgment.Source: §G.2 · p. 29

G.3 PREFERENCE

  • Preference ranking selects A over B, near-identical, or B over A, reserving near-identical for semantically and syntactically insignificant differences.Responses with substantially different content must be ranked even when their quality is similar.Source: §G.3 · pp. 29–30

G.4 DIRECT SCORING

  • Direct scoring rates a generated response from 1 to 5, from horrible to excellent, using readability, correctness, usefulness, and fit to prompt as criteria.The rubric ranges from unintelligible or majorly incorrect responses to expert-level, error-free responses with suitable tone and length.Source: §G.4 · p. 30

G.4 DIRECT SCORING; H.3 DIRECT SCORING

  • Direct scoring asks a judge to give the generated response a 1–5 integer score according to guidelines, with ratings spanning unintelligible or seriously incorrect responses to expert-level, accurate responses.Source: §G.4 · p. 31, §H.3 · pp. 35–36

G.5 CHECK-THEN-SCORE

  • Check-then-score asks a judge to assess a response with both the instruction and checklist, alongside a 1–5 quality rubric.The checklist is meant to guide, not limit, the judge’s assessment.Source: §G.5 · p. 31

G.6 SELF-REFINEMENT

  • Self-refinement prompts use checklist answers to address failed requirements while preserving passed ones, or use unstructured critiques to suggest actionable improvements.The checklist-based procedure requests a plan and rewrite, and returns the original response when all checklist answers are yes; both procedures require the rewrite to retain the original formatting.Source: §G.6 · pp. 32–33

H HUMAN ANNOTATION DETAILS

  • The same trained annotators handled all human annotation, receiving both general task-agnostic training and task-specific instructions and examples, with clarification available throughout.Source: H HUMAN ANNOTATION DETAILS · p. 34

H.1 CHECKLIST WRITING

  • Checklist-writing guidance defines facets as distinct prompt elements and calls for one specific, answerable Yes/No question per facet, with minimal overlap.Annotators first assess whether the prompt is safe and makes sense; unless either answer is no, they write questions applicable to responses to that prompt.Source: §H.1 · pp. 34–35

H.2 PREFERENCE LABELLING

  • Preference annotators choose on a five-point scale between two responses, reserving “near-identical” for semantically and syntactically insignificant differences.Responses that differ in content but have similar quality should not be labeled near-identical.Source: §H.2 · p. 35

H.4 CHECK-THEN-SCORE

  • Check-then-score separates prompt-specific criteria into Yes/No checklist questions, then uses the answers to inform—but not determine—the overall response score.A response that mostly fails the checklist should score low, but passing every question can still yield a middle-range score if overall quality is poor.Source: §H.4 · pp. 36–37

ANNOTATOR FEEDBACK ON CHECK-THEN-SCORE

  • 78.5% of annotators said using the checklist made providing an overall score easier; 6.4% said harder and 15% reported no effect.The authors present these responses as further evidence that LLM-generated checklists simplify rating individual responses.Source: ANNOTATOR FEEDBACK ON CHECK-THEN-SCORE · p. 37

DEMONSTRATING WHY HUMAN EVALUATION SHOULD ONLY USE GENERATED CHECKLISTS AS A PARTIAL GUIDE

A WildBench example shows why generated checklists should guide, but not replace, human judgment: checklist questions can assume capabilities the responding model lacks. Human evaluators gave the response low overall scores even when one evaluator answered nearly all checklist questions YES.

DEMONSTRATING WHY HUMAN EVALUATION SHOULD ONLY USE GENERATED CHECKLISTS AS A PARTIAL GUIDE

  • GPT-4o’s checklist assumed the model had accessed the requested materials, making its questions unhelpful for judging a response that should have disclosed it could not fulfill the request.Command-R+ responded as though it had visited the webpages despite using no retrieval-augmented generation.Source: DEMONSTRATING WHY HUMAN EVALUATION… · p. 37
  • All human evaluators gave the response low final scores, reflecting that the desired behavior was to say the request could not be fulfilled.Source: DEMONSTRATING WHY HUMAN EVALUATION… · p. 37
  • Annotator A answered YES to 9/10 checklist questions but still rated the response 2/5, Bad.Source: DEMONSTRATING WHY HUMAN EVALUATION… · p. 38
Loading 2410.03608v1…