CAM-DF¶

Core mechanism evidence. Paper and local evaluations are separate.

Paper information¶

Field	Value
Paper	arXiv 2607.27083
Affiliations	Peking University / McGill University / Shanghai University of Finance and Economics / Tsinghua University
First public release	2026-07-29 (arXiv v1)
Authors' code	The paper mentions cached artifacts and evaluation scripts but no public repository was found.
Adapter / method	`cam-df`
Local code	`src/auto_research/agent_research/methods.py`

Original paper¶

Background and contribution¶

Router rank alone does not tell how many tools to expose. Virtually traverse ranked prefixes before any tool execution. Payoff is sufficiency less heterogeneous cost; compare stopping with the best future prefix to label each decision and weight errors by regret.

flowchart LR
  R[Ranked tools] --> P[Enumerate prefixes] --> U[Sufficiency minus cost] --> G[Regret-weighted stop] --> T[Expose selected tools]

Key figure from the paper¶

Original CAM-DF figure. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ U(A,x)=\operatorname{sufficiency}(A,x)-\lambda\sum_{j\in A}c_j, \]

\[ \Delta_t=Q_{\rm stop}(t)-Q_{\rm cont}(t),\quad y_t=\mathbf 1[\Delta_t\ge0],\quad w_t=|\Delta_t|+\epsilon. \]

Paper results¶

1,343 tasks over five tool domains; on live τ-bench Retail, exposed tools fall from 7 to 4.4 (−37%) at similar task success. No production A/B.

Local reproduction¶

ScaleMCP mini freezes ranks/costs and evaluates all prefixes. Joint success 1.0000, 480 of 980 tools exposed (−51.02%), early stops 120/120. Metrics.

auto-research agent-eval --method cam-df --benchmark scalemcp-mini \
  --episodes 120 --seed 42

Reproduction limits¶

Deterministic mini-suite, not live Retail; sufficiency comes from auditable required-tool fixtures, not a learned scorer.