PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-Consistency Falls Short! The Adverse Effects of Positional Bias on Long-Context Problems

Adam Byerly, Daniel Khashabi

arXiv:2411.01101v3cs.CL

TL;DR

The paper asks whether self-consistency’s benefits on short tasks extend to long-context problems, where position bias is a challenge. It tests varied models, tasks, and SC formulations, finding that SC fails to improve and often degrades performance. The authors identify positional bias as a key issue and note that extremely long and multimodal contexts remain open questions.

  • Problem

    Whether self-consistency’s benefits on short tasks extend to long-context problems remains underexplored, despite the growing need to process lengthy inputs.

    Source: §1 · p. 1

  • Method

    The study evaluates multiple SC implementations across state-of-the-art models, smaller variants, and nine diverse long-context tasks.

    Source: §1 · pp. 1–2

  • Results

    Across 651 experiments and eight models, SC consistently failed to improve performance and often degraded it, with positional biases identified as a key factor.

    Source: §7 · p. 12

  • Takeaways & Limitations

    The findings challenge the generalization of inference-time self-consistency to long contexts and point to positional bias as an issue for future work.

    Source: §7 · p. 12

    Evaluation covers contexts up to 30 documents (5–10K tokens), leaving behavior at 100K+ tokens and in multimodal long-context scenarios open.

    Source: Limitations · p. 12

Abstract

from arXiv · show

Self-consistency (SC) improves the performance of large language models (LLMs) across various tasks and domains that involve short content. However, does this support its effectiveness for long-context problems? We challenge the assumption that SC's benefits generalize to long-context settings, where LLMs often struggle with position bias, the systematic over-reliance on specific context regions-which hinders their ability to utilize information effectively from all parts of their context. Through comprehensive experimentation with varying state-of-the-art models, tasks, and SC formulations, we find that SC not only fails to improve but actively degrades performance on long-context tasks. This degradation is driven by persistent position bias, which worsens with longer context lengths and smaller model sizes but remains invariant to prompt format or task type. Unlike short-context tasks, where SC diversifies reasoning paths, long-context SC amplifies positional errors. These comprehensive results provide valuable insight into the limitations of current LLMs in long-context understanding and highlight the need for more sophisticated approaches.

1 Introduction

The paper asks whether self-consistency’s benefits extend from short to long-context tasks, where position bias can make sampled errors correlated. Across varied models, tasks, and SC configurations, it finds that SC often degrades performance rather than correcting those errors.

1 Introduction

  • The study evaluates SC across state-of-the-art models and smaller variants, nine long-context tasks, and multiple aggregation strategies.The tasks span summarization, question answering, multihop reasoning, and controlled position experiments.Source: §1 · p. 1
  • SC consistently degrades performance across models and tasks, including synthetic and real-world settings.This challenges the assumption that SC universally improves LLM reliability.Source: §1 · p. 2
  • SC’s long-context limitations persist across multiple metrics, aggregation methods, and parameterizations.The study frames this as a robustness finding rather than an artifact of one SC implementation.Source: §1 · p. 2
  • Controlled experiments find that SC amplifies position bias in retrieval and reasoning through correlated positional errors.Varying prompt structure and SC configuration revealed the correlated-error pattern.Source: §1 · p. 2
  • The findings challenge SC’s presumed universality and motivate more robust approaches to long-context tasks.The paper points to position-aware voting and debiased sampling as possible directions for future research.Source: §1 · p. 2

2 Related Work

Related work establishes SC’s short-task successes and long-context position bias as separate concerns. The paper connects them by examining how aggregation may interact with biased model outputs, and evaluates across controlled position probes and naturally long-text tasks.

2 Related Work

  • Prior SC work largely studies short tasks, while long-context SC impact remains insufficiently addressed beyond a summarization study using ROUGE.The paper evaluates multiple SC implementations with automated metrics and LLM judges.Source: §2 · p. 2
  • Position bias includes primacy, recency, and U-shaped neglect of middle-context information.Prior work reports that these biases persist across architectures, instruction tuning, and context scaling.Source: §2 · pp. 2–3
  • Existing mitigation approaches target individual model outputs, leaving how aggregation methods such as SC might amplify positional errors unexamined.The cited approaches include context compression, attention calibration, and finetuning.Source: §2 · p. 3
  • Long-context evaluation spans constrained synthetic benchmarks and more representative naturally long-text tasks; this study combines both with controlled position probing.The passage notes concerns that synthetic benchmark performance may not predict downstream performance.Source: §2 · p. 3

3 Methodology

The methodology combines broad long-context benchmarking with controlled position-bias experiments, using multiple tasks, models, and self-consistency strategies. It compares SC with greedy-decoding baselines using task-specific and judge-based metrics, with bootstrapped confidence intervals.

3 Methodology

  • The study combines real-world long-context benchmarking with controlled experiments investigating position bias.Source: §3 · p. 3

3.1 Tasks and Datasets

  • Seven datasets cover summarization, question answering, and multi-hop reasoning, while NQ-Open supports paired QA and retrieval tests that separate information localization from answer generation.Source: §3.1 · p. 3

3.2 Self-Consistency Implementations

  • USC aggregates open-ended tasks with an LLM judge, QuALITY uses majority voting, and Soft-SC selects the response with the highest mean token likelihood.Source: §3.2 · pp. 3–4

3.3 Model Selection and Configuration

  • The evaluation tests proprietary and open-source model families ranging from 3B to 72B parameters, deploying all models in a standardized environment to eliminate implementation-specific confounders.Source: §3.3 · p. 4

3.4 Evaluation Framework

  • Evaluation uses task-specific metrics and GPT-4o judging for summaries, compares SC with greedy decoding, and applies bootstrapped 95% confidence intervals to assess performance differences.Source: §3.4 · p. 5

4 How Effective is Self-Consistency for Long-Context Problems?

Across long-context tasks and models, self-consistency yields almost no significant gains and often degrades performance. Larger models show milder degradation, but alternative aggregation with Soft-SC does not resolve the pattern.

4.1 Performance Degradation Across Tasks

  • 53 of 56 task-model pairs show no statistically significant improvement from SC or USC; summarization ROUGE declines significantly in 19 of 24 model-dataset pairs.Source: §4.1 · p. 5, Table 2 · p. 6

4.2 Model-Scale Dependency

  • Larger models show weaker degradation, but scale rarely produces gains; ROUGE and LLM-Judge correlations with model size are monotonic but weak-to-moderate and nonlinear.Source: §4.2 · p. 7, Figure 2 · p. 7

4.3 Robustness to SC Aggregation Method

  • Soft-SC does not improve over baseline and significantly degrades GovReport ROUGE, consistent with standard SC and USC limitations across aggregation strategies.Source: §4.3 · pp. 7–8, Table 3 · p. 8

5 Understanding the Phenomenon: Position Bias and Self-Consistency

Controlled QA and retrieval experiments test how document position and context length affect self-consistency, while robustness checks vary prompt format and sampling parameters. The results link SC’s degradation to correlated errors associated with position bias.

5.1 Controlled Exploration of Position Bias

  • The controlled NQ-Open experiments vary context size, gold-document position, sample count, temperature, and prompt format across QA and retrieval tasks.Source: §5.1 · p. 8

5.2 Evidence of Correlated Errors

  • QA accuracy follows a U-shaped position pattern across models and context lengths, with Qwen-2.5-7B dropping 15%±2% at middle positions in 30-document contexts.Source: §5.2 · p. 8, Figure 3 · p. 9
  • Retrieval accuracy declines as the gold document moves later; in 30-document contexts it drops 40%±3% in the final third, while SC reduces performance by 20–25% in larger tested models.Source: Figure 3 · p. 9, §5.2 · p. 10
  • Before aggregation, many samples already share positional errors: in 30-document retrieval cases with the answer at the end, nearly 60% had no correct sample among eight generations.Source: §5.2 · p. 10, Figure 4 · p. 10

5.3 Robustness Analysis

  • Prompt and parameter changes affect absolute performance but do not remove positional degradation; doubling QA generations from 8 to 16 improves accuracy by less than 1% while doubling computational overhead.Source: §5.3 · p. 11

6 Discussion

The discussion attributes SC’s long-context failures to positional bias and correlated errors across samples. It identifies position-aware aggregation and other bias-countering approaches as directions for future work.

6 Discussion

  • Across judge-based aggregation and majority voting, SC either degrades performance or fails to improve, and the discussion attributes this pattern to positional bias inducing correlated errors.Source: §6 · p. 11

6.1 Effects of Position Bias on Long-Context

  • Position-dependent performance indicates biased information access, and SC can amplify shared positional errors, including unanimous incorrect consensus when samples favor the wrong document positions.Source: §6.1 · p. 11

6.2 Practical Implications and Future Directions

  • The authors propose position-aware aggregation, contrastive sampling, attention recalibration, or retrieval-augmented small windows as testable directions for addressing correlated positional errors.Source: §6.2 · p. 12

7 Conclusion

Across 651 experiments and eight models, self-consistency (SC) failed to improve and often degraded long-context performance, including for GPT-4o. The conclusion warns against directly porting short-context methods and points to positional debiasing and context-aware inference as directions for further work.

7 Conclusion

  • Across 651 experiments and eight models, SC consistently failed to improve long-context performance and often degraded it due to positional biases.The degradation persisted across sampling configurations, prompt formats, and evaluation metrics, including for GPT-4o.Source: §7 · p. 12
  • Practitioners should consider the risks of porting short-context methods to long-context applications and investigate hybrid approaches combining SC with positional debiasing.Source: §7 · p. 12
  • The study characterizes SC’s limitations rather than solving them, providing empirical groundwork for targeted context-aware inference methods.The analysis covers 651 experiments across multiple models, tasks, and implementations.Source: §7 · p. 12

Limitations

The evaluation is limited to textual tasks and contexts of up to 30 documents, and the study characterizes SC failures rather than developing solutions.

Limitations

  • The evaluation covers contexts up to 30 documents (5–10K tokens), leaving behavior in extremely long contexts (100K+ tokens) open.The authors note that different dynamics might emerge at those lengths.Source: Limitations · p. 12
  • The study focuses on textual tasks, leaving SC’s utility in multimodal long-context scenarios unexamined.Real-world contexts may also involve more complex positional interactions than the controlled NQ-Open setup captures.Source: Limitations · p. 12
  • The authors characterize SC failures rather than develop solutions, positioning the 651 experiments as groundwork for targeted mitigation research.The stated focus is establishing a mechanistic understanding of why SC fails.Source: Limitations · p. 12

A Robustness Results

Prompt format and SC parameter changes do not remove positional degradation; their effects vary across QA and TR, with TR particularly sensitive.

Robustness Results

  • Up to 20% performance degradation separates prompt formats on TR, while changing formats has minimal impact on mitigating positional bias.Q-Doc-Q slightly improves overall QA performance, but degradation across gold positions indicates position bias persists regardless of query-document ordering.Source: Figure 5 · p. 18
  • Increasing generations slightly boosts QA accuracy but negatively impacts TR, while positional degradation persists as the gold document moves deeper into context.Higher temperatures degrade performance, especially on TR tasks.Source: Figure 6 · p. 19

B Error Analysis of Self-Consistency Failures

Intermediate-generation accuracy reveals different positional error patterns for QA and TR before self-consistency aggregation.

Error Analysis of Self-Consistency Failures

  • Before aggregation, QA accuracy follows a U-shaped curve, with performance degrading in the middle of the context.The analysis presents these intermediate samples as revealing the origin of positional bias.Source: Figure 7 · p. 20
  • Before aggregation, TR accuracy declines monotonically as the gold document is placed later in context.The figure plots accuracy for eight intermediate samples.Source: Figure 7 · p. 20

Illustrative Example: Correlated Errors in Question Answering

The example shows self-consistency amplifying correlated positional errors: seven of eight samples favor an incorrect answer tied to early context documents, and majority voting returns that answer.

Illustrative Example: Correlated Errors in Question Answering

  • Seven of eight samples confidently propose John Peters Humphrey, and majority voting amplifies this correlated error despite the correct answer appearing near the context’s end.The analysis attributes the pattern to primacy bias and says it violates self-consistency’s assumption of independent errors.Source: Illustrative Example: Correlated… · p. 21

C.3 USC Prompt

  • The USC prompt asks an evaluator to select the response most consistent with majority consensus.It specifies a response format beginning with “The most consistent response is Response X.”Source: §C.3 · p. 23

C.5.1 Task Prompts

  • The task prompts distinguish answering a question from identifying which document contains its answer.The QA prompt requests an answer; the retrieval prompt asks which documents are needed.Source: §C.5.1 · p. 25

C.5.2 Prompt Formats

  • Both listed prompt formats use the same Question {QUESTION} Answer template.The supplied formats do not state a difference between the two templates.Source: §C.5.2 · p. 25

Related papers

Loading 2411.01101v3…