Create BLOG.MD


BLOG.MD

@@ -0,0 +1,44 @@
+FocusFlow: Training LLM Agents to Protect Student Attention Meta × Scaler OpenEnv Hackathon 2026 — Abdul Hannan
+
+Why I Built This Hey everyone, I'm Abdul Hannan. While looking at other hackathon submissions, I noticed most people were building bots, simple games, or environments that don't connect to real problems. I wanted to build something that actually matters. One problem I've seen destroy people's potential — including my own — is distraction while studying. YouTube, BGMI, Instagram — these apps are scientifically engineered to steal your attention. Students lose hours every day and don't even realize it. So I built FocusFlow — an RL environment that trains an LLM agent to protect student attention.
+
+The Core Idea FocusFlow is built on the Pomodoro technique — 25 minutes of deep focus, 5 minutes of break. But the agent doesn't just follow a timer. It has to:
+
+Read natural language distraction events and judge urgency Manage cognitive load — the mental fatigue of studying Block the right apps at the right time Justify every single decision with graded reasoning
+
+What Makes It LLM-Hard This is the part I'm most proud of. Most RL environments can be solved by a simple rule-based bot. FocusFlow cannot. Every action requires a mandatory reasoning field. The agent must explain:
+
+What signal did it see? Why did it choose this action? How does this protect long-term focus?
+
+Empty reasoning = -0.15 penalty. Good reasoning = +0.10 bonus. A bot that just types "focus" repeatedly scores zero. The agent has to genuinely think.
+
+The 6 Actions
+
+block_app — Block a distracting app proactively BGMI has temptation 0.95, YouTube 0.90. The higher the temptation, the bigger the reward for blocking early.
+take_break — Rest when cognitive load > 0.80 If the agent keeps studying without breaks, cognitive load hits 1.0 and performance collapses. Smart agents rest before that happens.
+defer_event — Reschedule low-urgency distractions Example: "Rohan texted — bhai BGMI chalate hain" → urgency 0.3 → defer it, stay focused.
+respond_to_event — Handle urgent events immediately Example: "Professor posted — deadline moved to TODAY 11:59 PM" → urgency 0.95, can_defer=False → respond immediately or miss the deadline.
+focus — Stay on task when no threats are active Earns small positive reward every step. The foundation of every good session.
+adjust_energy — Restore energy when it drops below 0.35 Multi-day sessions drain energy. The agent must manage this across 3 days in Task 3.
+Anti-Reward-Hacking Design In RL, agents are notorious for finding shortcuts — they maximize reward without actually solving the problem. I designed FocusFlow with 4 independent reward signals that are hard to game simultaneously: Reward = reasoning_quality × 0.20 + action_correctness + cognitive_load_dynamics
++ deadline_pressure_tracking Plus an anti-spam check — if the agent repeats the same words in reasoning, unique_ratio drops below 0.5 and score = 0. You can't just repeat "focus focus focus" and win.
+
+Training Results I trained Qwen2.5-0.5B using GRPO via HuggingFace TRL. The agent went through 3 curriculum levels:
+
+Level 1 (Beginner): Learn to block apps Level 2 (Intermediate): Handle events + manage breaks Level 3 (Advanced): Full strategy across multi-day sessions
+
+Episode reward improved from -3.50 → +2.54 The agent learned to:
+
+Block high-temptation apps before they cause distractions Defer low-urgency social messages Take recovery breaks exactly when cognitive load hits 0.75+
+
+Real World Impact This isn't just a hackathon project. The reward logic in FocusFlow maps directly to proven productivity strategies used by top students and professionals worldwide. With some engineering, this agent's policy could be embedded into:
+
+OS-level focus modes Study apps like Forest or Focusmate Student productivity tools
+
+500 million students globally struggle with distraction. FocusFlow is a step toward an AI that fights for their attention instead of stealing it.
+
+Links
+
+Live Environment: https://hannan2859r-focusflow-env.hf.space GitHub: https://github.com/abdulhannan-18/Focus_Flow_env
+
+Submitted by Abdul Hannan — Meta × Scaler OpenEnv Hackathon 2026