Unverified 2026

Hysteretic competence-aware tool router

Implementation & benchmark of arXiv:2607.18460 — Competitive and Complementary Tools

Usefulness6/10

Difficulty4/10

Novelty7/10

Dynamics Accuracy Stability Sample efficiency Rlhf Training dynamics Moe routing Scheduler Dynamical systems Bifurcations Control theory

Source paper: Competitive and Complementary Tools arXiv:2607.18460 ⓘ · analyzed Aug 30, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Add a scalar competence state to a tool-augmented neural agent and let it control the probability of calling an external tool. Competence rises after autonomous success and decays when the agent offloads work, while tool reliance rises when competence is low; this creates a deliberate hysteresis loop that avoids both excessive tool calls and irreversible dependence. The router should be tested by temporarily removing the tool and measuring whether autonomous performance recovers.

Formulas

$$\dot{\iota}=r(1-x)(\iota_{\max}-\iota)-\mu x\iota$$

$$\dot{x}=\beta a\,\sigma(\iota_c-\iota)(1-x)-\big(\alpha(1-a)+\epsilon\big)x$$

$$\sigma(z)=\frac{1}{1+e^{-kz}},\qquad \iota_{t+1}=\operatorname{clip}\!\left[\iota_t+\Delta t\left(r(1-x_t)(\iota_{\max}-\iota_t)-\mu x_t\iota_t\right),0,\iota_{\max}\right]$$

$$x_{t+1}=\operatorname{clip}\!\left[x_t+\Delta t\left(\beta a_t\sigma(\iota_c-\iota_t)(1-x_t)-\big(\alpha(1-a_t)+\epsilon\big)x_t\right),0,1\right]$$

Mathematical statement

The paper models competence by $\iota$, reliance by $x\in[0,1]$, and tool availability by $a\in[0,1]$. Its coupled mean-field dynamics are $\dot\iota=r(1-x)(\iota_{\max}-\iota)-\mu x\iota$ and $\dot x=\beta a\sigma(\iota_c-\iota)(1-x)-[\alpha(1-a)+\epsilon]x$. Here $r$ is the autonomous competence-learning rate, $\mu$ is the competence-erosion rate under offloading, $\iota_{\max}$ is the competence ceiling, $\beta$ is the rate at which low competence increases reliance, $\iota_c$ is the reliance threshold, $\alpha$ is abandonment of the tool when unavailable, $\epsilon$ is baseline reliance decay, and $\sigma(z)=1/(1+e^{-kz})$ is a logistic with slope $k$. The positive feedback is explicit: high $x$ lowers $\iota$, and low $\iota$ raises $x$. In an ML agent, $\iota$ is an exponentially smoothed estimate of no-tool competence, $x$ is the router's tool-call probability, and $a$ is a binary or fractional tool-availability mask. A two-threshold implementation sharpens the predicted hysteresis: use $\iota_{\mathrm{on}}<\iota_{\mathrm{off}}$, calling the tool below $\iota_{\mathrm{on}}$ and suppressing it only above $\iota_{\mathrm{off}}$.

Implementation notes

(1) Integration point: insert a tool gate immediately before calculator, retrieval, code-execution, or search calls in a small tool-augmented language model. The model produces a base tool score $q_t$ from its hidden state; combine this score with the dynamical reliance state $x_t$. Maintain one competence scalar $\iota_t$ per user, task family, or model instance across episodes. Define autonomous competence as the exponentially smoothed correctness of attempts made with the tool disabled: $\hat c_t=(1-\rho)\hat c_{t-1}+\rho\,\mathbf 1[\text{no-tool answer correct}]$, then map it to $\iota_t=\iota_{\max}\hat c_t$. (2) Pseudocode: initialize $\iota,x$; for each task, compute a no-tool proposal and confidence, calculate $p_t=\operatorname{clip}(q_t+\lambda x_t,0,1)$, sample or select the tool gate, execute the tool if selected, evaluate correctness, update $\hat c_t$, and apply the two Euler recursions. Alternatively use a hysteretic gate: set $p_t=1$ when $\iota_t<\iota_{\mathrm{on}}$, set $p_t=0$ when $\iota_t>\iota_{\mathrm{off}}$, and retain the previous gate state between thresholds. Add a small tool-call cost to the answer loss. (3) Estimated quantities: estimate $r$, $\mu$, $\beta$, and $\epsilon$ by fitting observed trajectories of $(\hat c_t,x_t)$ with least squares; set $k$ from the desired transition sharpness and tune $\iota_c$, $\iota_{\mathrm{on}}$, and $\iota_{\mathrm{off}}$ on validation data. (4) First experiment: use a 1B-or-smaller instruction model on GSM8K or a synthetic arithmetic and multi-hop benchmark with calculator and retrieval tools. Compare always-tool, confidence-only routing, a learned router, and this dynamical router at equal answer quality. Run alternating blocks of tool availability and measure answer accuracy, tool calls per task, performance after tool removal, and recovery speed after re-enabling the tool. Success is fewer tool calls at equal accuracy plus materially better autonomous accuracy and recovery than confidence-only routing; failure is persistent bistability at the wrong low-competence state or no improvement over a static learned gate.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.