Tom FENG

AI researcher & builder

I research mechanistic interpretability and build infrastructure for autonomous AI agents. My work spans two layers: programming multi-step reasoning circuits directly inside Transformer weights (Weight Space Programming), and building the protocols that let autonomous agents coordinate, negotiate, and ship real work together.

I'm the author of In-Weight FPGA and creator of Concord, a platform where AI agents share a room and build software in real time.


Research & Writing

Why does fine-tuning (SFT / LoRA) cause severe collateral over-triggering (80–100% false refusals on XSTest benign technical prompts and 28–100% hallucinated tool calls on math), and why do continual learning and model merging (Task Arithmetic, TIES-Merging) suffer destructive cross-task interference? We introduce In-Weight OS: partitioning dormant SwiGLU channels of the terminal MLP (layers[L-1].mlp) into strictly disjoint skill banks ($\mathcal{C}_1 \cap \mathcal{C}_2 \cap \mathcal{C}_3 = \emptyset$) gated by a 2-Fold Out-of-Fold (OOF) Shrinkage-Calibrated Guard-Nullspace Solver. In 12–15 seconds with zero gradient steps, we simultaneously compile Surgical Safety, 3-Way JSON Tool Routing, and System-1/2 Compute Routing into just 51–57 neurons across Qwen3-8B and Llama-3.1-8B-Instruct.

Llama-3.1 3-Skill Success
98.3%
vs 73.8% TIES / 51.7% TA
Qwen3-8B 3-Skill Success
100.0%
vs 80.6% TA / 62.5% Cont-SFT
XSTest False Refusal
0.0%
vs 80–100% on Safety SFT
Edited Scope & Time
51 Neurons
12–15s · 0 Gradient Steps

Why do LLMs capitulate to misleading critiques (sycophancy), while reinforcement learning (SelfD-RL) makes them blindly stubborn (collapsing valid critique acceptance by -30.8%)? And why do mid-layer weight edits fail in multi-step free generation? We uncover the Read-Write Decoupling & Terminal-Layer Zero-KV-Drift Principle: in any pre-norm Transformer, the terminal MLP (layers[L-1].mlp) is the unique nonlinear block that executes strictly after all self-attention KV caches are written. In 30 seconds of closed-form linear algebra with zero gradient steps (editing ~0.035% of parameters), we compile multi-step conditional state machines into dormant SwiGLU neurons across Qwen3-8B and Llama-3.1-8B-Instruct.

Qwen3-8B Dialectical
86.82%
+13.08% (vs 71.1% RL)
Llama-3.1-8B Dialectical
73.12%
+18.22% (Hold & Yield ↑)
Program 2 (3-Way+2-Sub)
98.75%
0.00% False Triggers
Paired Bootstrap 95% CI
12 / 12 Passed
3,780 Real Dialogues

Projects

Concord

2026

A collaboration platform for AI agents. Multiple agents join a shared room to negotiate APIs, write code, and fix issues together in real time — no human relaying messages between them. It ships with coordination primitives (decaying signals, quorum votes, exclusive claims), end-to-end encryption with client-side keys, and human-gated destructive operations. Agents join with plain REST calls — no SDK, no configuration.