Skip to content

README.md

Latest commit

 

History

History
142 lines (84 loc) · 4.67 KB

README.md

File metadata and controls

142 lines (84 loc) · 4.67 KB
Raw

Attention Span: The "Needle in a Salient Haystack" Benchmark

Evaluating Selective Attention and Distractor Vulnerability in Frontier LLMs.

Open In Kaggle

AGI Attention: Salient Distractor Benchmark header featuring the Selective Attention metric and Principal Investigator Amey Thakur.

Hello fellow Kagglers!

This notebook presents a robust and reproducible evaluation pipeline for the Measuring Progress Toward AGI competition. The objective is to evaluate a model's cognitive suppression capabilities using a "Needle in a Salient Haystack" methodology.

The solution integrates:

  • procedural generation of 100 high-salience synthetic contexts
  • exploratory data analysis using Seaborn and Matplotlib
  • custom target-vulnerability mapping
  • multi-model benchmarking via the kaggle_benchmarks SDK
  • rigorous validation using cross-model accuracy metrics

Each component is designed to bridge the gap between simple retrieval and true selective attention.

Understanding the Problem Setting

Current AGI benchmarks often focus on "retrieval capacity" but ignore "instruction adherence under noise." Real-world AI must ignore loud, irrelevant distractors to focus on the user's core intent.

This task is modeled as a distractor vulnerability problem:

$$ \text{Model Response} \cap \text{Distractor} = \emptyset $$

Key challenge

  • instruction-needle pairs are low-salience and "boring"
  • injected distractors use high-salience, urgent vocabulary (CRITICAL/EMERGENCY)
  • benchmarking across Scale, Reasoners, and SOTA modalities

1. Data Acquisition and Setup

The environment initializes the kaggle_benchmarks SDK and procedurally generates a 100-row testing suite.

Configuration includes:

  • target/needle token pairs
  • urgency-weighted padding strings
  • randomized positional indexing (injecting distractors at varying depths)

Why this matters

By randomizing the distractor position, we ensure the model isn't just failing due to "lost in the middle" effects, but rather due to the salience of the distractor.

2. Data Inspection & EDA

We perform a thorough analysis of the generated benchmark data.

Histograms and KDE plots in Section 4 show:

  • distribution of context lengths
  • distribution of distractor lengths
  • frequency of urgency triggers

Why this matters

Ensures the benchmark is balanced and that the "haystack" is sufficiently dense to challenge even frontier models.

3. Modeling Strategy (The Task Hook)

We implement the evaluation using the @kbench.task decorator.

Core Functionality

The model is prompted to extract a specific key while various "emergency" overrides are present in the text. We use the updated SDK syntax to trigger batch inference across the entire DataFrame:

results = selective_attention_task.evaluate(llm=[kbench.llm], evaluation_data=df)

4. Performance Analysis (The "AGI Gap")

The results reveal a stark contrast in cognitive control among current frontier models:

  • Gemini 2.5 Flash: Failed (69% vulnerability rate)
  • Claude Opus 4.6: Failed (Attention hijacked by salient noise)
  • DeepSeek-R1: Failed (Reasoning monologue trapped by urgency)
  • Gemini 3.1 Pro Preview: PASSED

Implication

The fact that Gemini 3.1 Pro Preview was the only model to successfully suppress 100% of the salient distractors marks a significant milestone in AGI development. It suggests that newer architectures are moving beyond simple pattern matching and toward true Cognitive Suppression.

Summary

The pipeline demonstrates:

  1. a high-precision method for measuring Selective Attention
  2. the massive failure of current reasoners against high-salience noise
  3. a clear AGI "Tier 1" signal from the newest Gemini architecture

This isolates a specific, actionable metric for the Attention track of the competition.

Closing Remarks

This project highlights that true intelligence requires not just seeing what is there, but strategically ignoring what is loud.

Further research will focus on:

  • adversarial distractor optimization
  • cross-modal (Vision/Text) saliency hijacking
  • measuring "Distraction Latency" in streaming responses

These directions are critical for building reliable AI agents in noisy, critical-infrastructure environments.

Amey Thakur

Kaggle • GitHub