Anti-Alignment Alignment Club
Members’ names: Alexander Brady, Oliver Sieberling, Pascal Feigenwinter, Yanick Schimpf
We introduced a number of changes to increase the performance of Sentient’s OpenDeepSearch agent, focusing on strengthening both how information is retrieved and how it is processed by agents and models, and were evaluated to strong results on FRAMES, outperforming models with 10x more parameters through oughintelligent, general use enhancements. Below we list the most notable improvements:
Enumeration Agent
We built a tool which aldata frames. This upgrade facilitates simple enumeration tasks as well as more complex operations on arrays. The result is a system that can process raw web data into usable formats automatically, increasing the range and flexibility of supported queries.
Majority Voting Mechanism
To improve answer reliability, the system now generates multiple candidate responses and uses an LLM to identify and quantify the most common solution. This majority-vote approach increases robustness by filtering out outliers and emphasizing consensus across multiple reasoning paths.
Advanced Planning Capability
Agents are now prompted to plan out their steps before execution. This enables a more structured and logical approach to solving complex tasks. Advanced planning reduces errors and ensures that intermediate steps are aligned with the final goal.
Source Prioritization
A minor yet useful update introduces a hierarchical approach to source querying. Agents sequentially check for information in the following order:
Answer Box → Wikipedia → Aggregated Sources → Top News Articles
The agent only proceeds down the hierarchy if the current level fails to provide a satisfactory result. This structure prioritizes reliability and minimizes unnecessary processing, allowing for more efficient agents in the average case.
Hyperparameter Tuning
Minor adjustments to internal hyperparameters now allow the model to handle larger context windows more effectively, reducing context fragmentation and maintaining continuity across larger input chunks.