1 of 15

BrainBench Milestone 3

Orion Powers and Daniella Seum

2 of 15

Milestone 3 Progress Matrix

Task

Daniella

Orion

Completion

To-do

Compile math problem dataset for advanced Probability and Statistics.

50%

50%

100%

N/A

Finalize free or local LLMs we are using for initial testing.

50%

50%

100%

N/A

Finish developing testing pipeline, complete with robust output parsing to verify correctness of responses. Implement, test, demo.

50%

50%

100%

N/A

Test LLMs on problem dataset.

50%

50%

25%

Finish running tests, just need time

3 of 15

Stats Problems

01

Problem Count

Completed a dataset of 1,000 probability and statistics problems.

02

Topics Covered

  • Expected value
  • Variance
  • Conditional probability
  • Binomial coefficients
  • Permutations
  • Combinations
  • Distributions

03

XML Export

  • Exported to XML file
  • Same format as 8th Grade and Calc 1 dataset

4 of 15

Local LLMs

Model Selection

  • gpt-oss:20b
  • qwen3:8b
  • gemma3:12b

Local GPU Feasibility

Models are open-license and optimized for local GPU inference without cloud costs through Ollama.

5 of 15

Prompt Forces Keyword

01

Keyword-Based Extraction

The system prompt includes 'Print FINAL_ANSWER: <value> on its own line.' This keyword is captured using regex, ensuring high extraction success rates with a confidence score of 1.0.

02

Fallback Mechanism

If the keyword is missing, the system falls back to other extraction methods, maintaining an overall extraction success rate of over 99% across all models.

6 of 15

Fallback Chain

LaTeX Boxed Notation

If the keyword is not found, the system scans for LaTeX boxed notation (\boxed{...}) with a confidence score of 0.8, capturing answers in mathematical format.

Natural Language Patterns

If LaTeX notation is absent, the system searches for natural language patterns like 'Answer is' with a confidence score of 0.6, handling descriptive responses.

Last Numeric Token

As a last resort, the system extracts the last numeric token with a confidence score of 0.4, ensuring answers are captured even in less structured responses.

7 of 15

Auto-Type Detection

Regex-Based Type Detection

The system uses regex patterns to detect answer types: fractions, decimals, integers, coordinates, ranges, expressions, and text. Each type is routed to a specific normalization function.

NormalizedAnswer Object

Detected types are stored in a NormalizedAnswer object, ensuring structured and consistent data for accurate comparison and verification.

8 of 15

Per-Type Cleanup

01

Fraction Normalization

Fractions are kept in their raw tuple form (numerator, denominator) without reduction, preserving the exact values provided by the LLM.

02

Decimal Normalization

Decimals are converted to float values, with precision bits tracked to ensure accurate tolerance calculations during comparison.

03

Expression Normalization

Expressions are stripped of spaces, sorted by exponent in descending order, and simplified (e.g., x^1 becomes x) to ensure consistent string comparison.

04

Text Normalization

Text answers are converted to lowercase to ensure case-insensitive comparison, handling qualitative responses like 'rational' or 'irrational'.

9 of 15

Per-Type Rules

01

Fraction Comparison

Fractions are reduced to their lowest terms using the greatest common divisor (GCD) and compared as tuples to ensure mathematical equivalence.

02

Decimal Comparison

Decimals are compared with a tolerance based on their precision, ensuring accurate matching within acceptable error margins.

03

Integer Comparison

Integers are compared exactly, requiring an exact match without any tolerance, ensuring precise verification.

04

Expression Comparison

Expressions are compared as whitespace-free strings, ensuring exact symbolic equivalence without evaluating the expressions.

10 of 15

Demo

11 of 15

Milestone 4 Task Matrix

Task

Daniella

Orion

Finish running local LLMs on problem datasets.

33%

66%

Begin development on the website and add data from local LLMs.

66%

33%

Explore possibilities of expanding to larger LLMs and their APIs, or expanding on problem types.

50%

50%

12 of 15

Finish Runs

01

Batch Completion

Allow ongoing batch runs to complete without interruption.

02

Data Harvesting

Harvest accuracy and latency data into CSV format for analysis.

03

Log Backup

Backup logs and checksums to ensure data integrity.

04

Report Generation

Generate detailed per-model reports summarizing results.

13 of 15

Build Dashboard

01

Angular Skeleton

Develop the Angular skeleton for the evaluation dashboard.

02

Chart.js Integration

Integrate Chart.js for dynamic charts and tables.

03

Responsive Design

Ensure the dashboard is mobile-responsive for accessibility.

14 of 15

Expand

01

API Quotes

Research and quote costs for GPT-4/5, Claude, Gemini, and DeepSeek-R1.

02

Other Problem Types

Coding, logic, image generation, etc.

15 of 15

Thank you.