Weekly Report — 2026-W40 (2026-09-28 ~ 2026-10-04)

This week focused on transitioning high-stakes research from documentation to operational readiness through rigorous auditing, VLM benchmarking, and advanced robotic policy evaluation. Key achievements included the successful deployment of the RecoverBench framework, the optimization of OmniJev through visual-only memory architectures, and the technical refinement of academic application materials and WeAgent infrastructure.

Weekly Overview

Metric	Value
Date Range	2026-09-28 ~ 2026-10-04
Active Days	6 / 7
Total Conversations	48
Projects	32
Tasks Completed	47
Tasks In Progress	15
Total Tokens	7,527,067,752
Total Cost	$2,744.02
Claude Code Token	7,488,004,360
Claude Code Cost	$2,677.32
Codex Token	39,063,392
Codex Cost	$66.70
Daily Average Cost	$457.34

Project Progress

RecoverBench (4 days active) — 🔄 active

Accomplishments:

• Completed Phase 2 implementation with deterministic numerics
• Executed large-scale claim vs. evidence auditing via multi-agent workflows
• Initiated Phase 3 deployment audits and human demo reconciliation

Blockers:

• ⚠️ Need to reconcile recording requirements and operator kit bugs

OmniJev / VLM Evaluation (4 days active) — 🔄 active

Accomplishments:

• Achieved 77.5% success rate via visual-only arm implementation
• Established HTTP-based closed-loop evaluation architecture
• Identified GPT-6 Astra’s superiority in long-term object identity tracking compared to Gemini 3.1 Pro

Blockers:

• ⚠️ Visual noise in wrist-camera tasks (BinFill)

MIHD / MihD-dev (3 days active) — 🔄 active

Accomplishments:

• Assembled complete first manuscript draft
• Launched Round 3 subagent management (13 agents)
• Established decision register and formal scientific guardrails

Blockers:

• ⚠️ Pending high-level policy decisions (H1-H7)

WeAgent / Wechat Integration (3 days active) — 🔄 active

Accomplishments:

• Migrated core logic to AstrBot plugin architecture
• Integrated LifeCopilot entry and channel token authentication
• Resolved critical SQLite WAL corruption using a ‘Replay’ approach

Blockers:

• ⚠️ Maintaining stability during auto-restarts

Academic/PhD Application (3 days active) — 🔄 active

Accomplishments:

• Restructured CV to prioritize research identity and impact
• Generated interactive PhD/MS program dashboard
• Conducted massive advisor reputation and alignment audits

Key Tasks

• 🔄 RecoverBench Claim & Code Audit — Executing a large-scale workflow to build a ‘have/missing/differs’ ledger of all research claims against code, data, and cluster artifacts; transitioning from claim-compliance to operational usability/pipeline gap analysis.
• 🔄 MIHD Manuscript Drafting & Investigation — Assembled a complete first draft of the MIHD manuscript and launched a multi-phase investigative workflow to audit current implementations against the published methodology.
• ✅ OmniJev Closed-loop Deployment & Optimization — Established an independent HTTP-based evaluation architecture and verified the performance lift of the ’event memory + point-wise judgment’ logic in VideoPlace tasks.
• ✅ W4A4 Closed-Loop Evaluation — Evaluated the W4A4 configuration on CPU, identifying a 0.0% success rate due to significant backbone activation degradation.
• ✅ OmniJev 16-task Extension & Evaluation — Implemented shim extensions and truth replay for 16 RoboMME tasks and conducted a full closed-loop evaluation using a 2B model.
• 🔄 Gemini Visual-Only Integrity Enforcement — Executing a pure-visual Gemini arm ablation by stripping proprioception and implementing detection-based visual window tracking.
• ✅ Advisor Reputation Audit (Top 50) — Conducted an exhaustive multi-phase workflow audit of hundreds of professors to filter for advising style, program protection, and research alignment.
• ✅ Fix SQLite WAL Merge Integrity — Implemented a robust ‘SQLite Replay’ approach for decrypted WAL files to ensure integrity and handle complex checkpointing/frame validation rules.

Problems & Solutions

1. Discrepancy between RecoverBench paper claims and actual training data used. [RecoverBench]

Solution: Conducted a systematic agentic audit to map claims to actual files on the cluster; implemented a ‘Claims/Evidence’ ledger.

2. ResNet shortcut implementation used projection convolutions instead of identity requirement. [ECE661/ResNet]

Solution: Refactored the ResNet class to implement Option A (identity shortcut) and re-trained the model.

3. OmniJev M2 execution was extremely slow due to Conv3d patch embedding bottleneck. [OmniJev]

Solution: Replaced the slow torch implementation with an equivalent linear layer, reducing latency from 16s to 0.06ms per image.

4. W4A4 quantization led to severe performance degradation (0% success rate). [Quantization/VLA]

Solution: Identified backbone activation as the error source and recommended moving to W8A8 to avoid non-functional quantization graphs.

5. SQLite WAL merge causing database corruption and ‘quick_check’ failures. [WeAgent]

Solution: Replaced manual merging logic with a ‘SQLite Replay’ approach using the official engine to play back decrypted WAL files.

6. Inflated robot policy evaluation success rates due to hidden proprioceptive data usage. [OmniJev]

Solution: Implemented DRAWN_MEMORY_VISUAL_ONLY/TRACK flags to force zero proprioception, correcting success rate to 65.0%.

Learnings

Architecture (architecture)

• In multimodal agents, decomposing memory tasks—using code for temporal/sequence logic and VLM for perception—is superior to using the VLM for all reasoning.
• For research-focused identity (CVs/PhD), a strategic narrative prioritizing problem-definition and research intuition is more effective than high-volume technical listing.

Debugging (debugging)

• GPU kernel autotuning (e.g., Triton GEMM/XLA) can introduce non-determinism that mimics code regressions; requires forced deterministic autotuning for benchmarks.

Domain Knowledge (domain)

• 4-bit simulations can be overly pessimistic and unreliable for predicting hardware accuracy due to hardware automatically utilizing FP16 for certain operations.
• Reliable robot learning must distinguish between ‘arbitrary starting state’ and ‘state following an execution error’ to measure true error recovery.

Tools (tools)

• When dealing with complex database internals like SQLite WAL, it is safer to use the official engine to replay decrypted files than to implement custom manual merging.

AI Usage Notes

Effective Patterns:

• ✓ Multi-agent workflows for ‘Claims/Evidence’ auditing
• ✓ Agentic manuscript drafting using writer/reviewer/adversarial loops
• ✓ Using AI to consolidate high-level scientific constraints into structured decision tables

Limitations:

• ✗ AI tendency toward ‘over-certainty’ in research inferences
• ✗ Failure to recognize non-visual data leakage in policy evaluations without explicit prompting
• ✗ Difficulty in managing complex, high-precision data structures (AttributeError/KeyError) in multi-agent loops

Next Week Outlook

Next week’s priorities include completing the LIBERO evaluation for quantized pi0.5, finalizing the MIHD project restructure into Full/E2E/Demo levels, and progressing the implementation of detailed error definitions (H1, H2, H5) for the RecoverBench camera-ready version. Additionally, focusing on the VLA quantization recovery (W8A8) and completing the Gemini vs. Astra visual prompt comparison will be critical.

Token Usage Statistics

Peak Day: 2026-10-04 — $887.99 / 2626.2M tokens

Daily Average: $457.34