Here is the baby stream cipher I used for my course. Take linear feedback shift registers of widths $\ell$ and $\ell'$ generating keystreams of maximum period $2^\ell-1$ and $2^{\ell'}-1$, respectively. Let $k_0k_1k_2 \ldots $ and $k'_0k'_1k'_2 \ldots $ be the keystreams for chosen keys $k_0\ldots k_{\ell-1}$ and $k'_0 \ldots k_{\ell'-1}$.
For $m \in \mathbb{N}$, define the $m$-quadratic stream cipher to have keystream $u_0u_1u_2 \ldots$ where
$$ u_i = k_ik'_i \oplus k_{i-1}k'_{i-1} \oplus \cdots \oplus k_{i-(m-1)}k'_{i-(m-1)} \quad\text{for $i \ge m-1$} $$
and $u_0 = \ldots = u_{m-2} = 0$. Thus after the first $m-1$ bits, each bit is defined by adding up the product of $m$ consecutive bits produced by the LFSRs.
The expected correlation between the keystream $u_0u_1u_2 \ldots $ and the LFSR keystream $k_0k_1k_2 \ldots $ is about $1/2^m$. This can be proved as follows: if $X_1, \ldots, X_m$ and $X_1' \ldots, X_m'$ are independent and unbiased bits then the correlation between $X_1X_1' + \cdots + X_mX_m'$ and $X_m$ is
$$\mathbb{E}[(-1)^{X_1X_1' + \cdots + X_mX_m' + X_m}] = \mathbb{E}[(-1)^{X_1X_1'}] \ldots \mathbb{E}[(-1)^{X_{m-1}X_{m-1}'}]\mathbb{E}[X_mX_m'+X_m] = \frac{1}{2^{m}} $$
since each pair $X_iX_i'$ is biased to $0$ with probability $3/4$, and the same holds for $X_mX_m' + X_m$. Therefore an attack using this correlation requires at least $2^{2m}$ bits from $u_0u_1u_2 \ldots $ to distinguish a correct guess for $k_0\ldots k_{\ell-1}$ from an incorrect guess. The graphs below show the correlations for all $31$ non-zero guesses for $k_0k_1k_2k_3k_4$ in a toy example where $\ell = 5$ on $1024$ bits of keystream. The key is $00111_2 = 7$.

As it stands, the toy $m$-quadratic is easily broken since
$$u_i \oplus u_{i-1} = k_ik_i' \oplus k_{i-m}k_{i-m}'$$
cancels out most of the quadratic terms. This leads to an effective correlation attack for any $m$. As an improvement, I suggest changing the feedback function so that
$$u_i = k_ik_i' \oplus k_{i-1}k'_{i-1} \oplus k_{i-2}k'_{i-2} \oplus k_{i-4}k'_{i-4} \oplus \cdots \oplus k_{i-2^{m-2}}k'_{i-2^{m-2}}. $$
For $m \ge 4$ there is no linear combination of the keystream bits having fewer than $m$ quadratic terms, so the correlation attacks above are ineffective.