anupamagarwal001/amc_allocator_env

1# AI Investment Committee Environment: Short Slide Deck Outline2 3## Slide 1: Problem4 5- LLMs are still weak at resolving structured disagreement across multiple actors with different incentives.6- Real portfolio decisions are committee decisions, not single-step predictions.7- We want an environment that can train an LLM to ask for information, balance opportunity against risk, and adapt under changing constraints.8 9## Slide 2: Environment10 11- OpenEnv environment with three roles:12  - Portfolio Manager (trainable)13  - Research Analyst (scripted)14  - Risk Officer (scripted)15- Hidden market regimes, noisy signals, transaction costs, and mandate updates.16- Four deterministic tasks:17  - guided allocation18  - research-risk conflict19  - regime-shift recovery20  - mandate drift21 22## Slide 3: Why This Is Novel23 24- Not a toy game or grid-world clone.25- Multi-agent professional workflow with partially observable incentives.26- The PM must resolve conflicts between research conviction and risk discipline.27- The research question is whether verifier-driven RL can improve strategic coordination inside a realistic investment committee workflow.28 29## Slide 4: Reward + Training30 31- Reward is decomposed into verifier-style components:32  - return33  - drawdown34  - compliance35  - information usage36  - risk response37- Training stack:38  - OpenEnv39  - HF TRL GRPO40  - LoRA adapters41  - Qwen/Qwen3-0.6B42  - Colab T4 smoke run43 44## Slide 5: Results45 46- Baseline overall score:47  - heuristic `0.4084`48  - random `0.2504`49- Smoke training reward:50  - start `0.0193`51  - end `0.0275`52  - delta `+0.0082`53  - best step `4`54- Key judge-facing artifacts:55  - `reward_curve.png`56  - `judging_report.md`57  - `onsite_demo_summary.md`58 59## Slide 6: Why It Matters60 61- This environment teaches structured multi-agent reasoning, not just asset ranking.62- It is useful for training LLMs to operate in professional decision loops with partial observability and changing incentives.63- The result is a trainable, research-grade workflow benchmark rather than a finance-themed toy problem.64