Reading notes on:

Modern LLM RL mostly walked away from critics. The reason was pragmatic: actor-critic systems are harder to scale, harder to keep numerically aligned, and much easier to destabilize than group-relative baselines. That trade made sense in the throughput-driven era mapped out by StreamRL and RL Systems Mind the Gap.

But the cost of dropping the critic is real. Group baselines give one blunt sequence-level signal to an entire response, which is exactly the wrong shape for long reasoning traces. Le Critique asks whether token-level credit assignment can come back without resurrecting the instability that made the field abandon PPO-style critics in the first place. Its answer is elegant: let the critic see privileged but admissible information, then mix that value estimate with the reliable group baseline using a learned coefficient.

Privileged critic branch and TETHER's adaptive mixture of group and value baselines


1. What the PPO to GRPO Shift Bought, and What It Lost

The generic token-normalized policy-gradient estimator is

\[\hat{\nabla}_\theta J = \frac{1}{\sum_{i=1}^B T_i} \sum_{i=1}^B \sum_{t=1}^{T_i} w_{i,t}\,\hat{A}_{i,t}\,\nabla_\theta \log \pi_\theta(y_{i,t} \mid h_{i,t}).\]

Everything interesting is packed into the advantage estimator $\hat{A}_{i,t}$.

In actor-critic methods, that advantage can vary token by token. In group-relative methods, it is effectively sequence-level. The leave-one-out baseline

\[b_i^{LOO} = \frac{1}{K - 1}\sum_{j \ne i} R_j\]

and the GRPO baseline

\[A_i^{GRPO} = R_i - \bar{R}\]

are equivalent up to a constant scale:

\[A_i^{GRPO} = \frac{K - 1}{K} A_i^{LOO}.\]

That is why critic-free RL scales cleanly: the baseline is training-free and synchronization-light. But it is also why it struggles to tell which reasoning step mattered. The whole trajectory shares a single reward contrast.

This explains the historical split:

Method Credit assignment Infra burden Failure mode
PPO / actor-critic token-level high critic lag, instability
GRPO / RLOO sequence-level low blunt temporal credit

The paper’s aim is not to revert to old PPO. It is to recover critic granularity without paying the full systems tax.


2. The Key Theoretical Boundary: Admissible Privilege

The crucial idea is that a critic can condition on more information than the policy, as long as that extra information does not leak the sampled action itself.

The sufficient admissibility condition is

\[\mathbb{E}\!\left[b(h_{i,t}, z_{i,t}) \nabla_\theta \log \pi_\theta(y_{i,t} \mid h_{i,t}) \mid h_{i,t}\right] = 0 \quad \Longleftarrow \quad z_{i,t} \perp\!\!\!\perp y_{i,t} \mid h_{i,t}.\]

So the question becomes: what counts as admissible privilege?

Safe examples:

  • gold answers or solved boards;
  • verifier rubrics or proof sketches;
  • sibling rollouts and rewards from the same prompt group, excluding the current trajectory.

Unsafe examples:

  • future tokens of the trajectory being scored;
  • the realized terminal reward of that exact trajectory;
  • environment feedback that only exists because the current action was taken.

This distinction is load-bearing. It turns privileged critics from a hand-wavy systems trick into a proper control-variate construction.


3. Privileged Value Functions

The proposed critic is a Privileged Value Function (PVF):

\[V^\pi(h_{i,t}, z_{i,t}) = \mathbb{E}_{\tau_i \sim \pi}[R_i \mid h_{i,t}, z_{i,t}].\]

The paper’s motivating example is Sudoku. A standard critic observing only a partial board has to implicitly solve a difficult latent inference problem to predict the return. A privileged critic that also sees the completed board turns that into verification rather than search.

The gain is formal:

\[\mathbb{E}\Big[(R_i - \mathbb{E}[R_i \mid h_{i,t}, z_{i,t}])^2\Big] \le \mathbb{E}\Big[(R_i - \mathbb{E}[R_i \mid h_{i,t}])^2\Big].\]

So, with an optimal predictor, adding admissible privilege can only reduce variance.

That makes PVFs quite different from self-distillation or reward shaping. They are not changing the RL objective. They are estimating a better baseline.


4. TETHER: Let the Baseline Decide When to Trust the Critic

Even a principled critic can be wrong early in training. The paper’s fix is TETHER, an adaptive mixture between the stable group baseline and the learned value baseline:

\[b_{i,t}^{TETHER} = (1 - \rho)\, b_i^{LOO} + \rho\, V_{i,t}.\]

Define the group advantage and value correction as

\[A_B = R - B, \qquad \Delta = V - B.\]

Minimizing the least-squares prediction error gives the optimal mixture coefficient

\[\rho^* = \text{clip}_{[0,1]} \left( \frac{\mathbb{E}[A_B \Delta]}{\mathbb{E}[\Delta^2]} \right).\]

This is a good design for two reasons:

  1. When the critic is useless, $\rho$ collapses toward zero and the algorithm falls back to the safe group baseline.
  2. When the critic becomes predictive, $\rho$ increases smoothly and token-level credit gradually takes over.

In implementation, $\rho$ is smoothed with an EMA using decay 0.95, and the paper insists on an evaluate-before-train protocol so that the critic is scored on data it has not just memorized.


5. What the Results Actually Say

The most direct metric is critic explained variance:

Task Standard VF EV PVF EV
Sudoku ($K=4$) 0.032 0.245
CodeIO ($K=4$) 0.594 0.751
Reasoning Gym ($K=8$) 0.499 0.519

That is a big jump, especially on Sudoku where the latent search problem is hardest.

But the headline is more nuanced than “critic wins everywhere.”

  • On Sudoku, TETHER mitigates the damage from a weak standard critic, but the plain group baseline still slightly outperforms it in final reward.
  • On MiniF2F and Reasoning Gym, TETHER wins, which suggests token-level credit matters most when partial prefixes already contain useful proof structure or semantic evidence.

The paper also highlights a long-context tuning detail that is easy to miss and likely more general: for long CoT, $\lambda_{GAE}$ may need to sit extremely close to 1. The reported sweet spot is

\[\lambda_{GAE} = 0.999888,\]

chosen so the first token in an 8,192-token trajectory still retains roughly 40% of the terminal reward signal. A default value like 0.95 simply washes the credit out.


6. Why This Matters

The paper restores an old idea in a more modern form:

  1. Critics do not need to be policy clones. They can be privileged verifiers as long as the privilege is admissible.
  2. Group baselines remain the right bootstrap. TETHER is persuasive precisely because it does not force the critic to be trusted before it earns that trust.
  3. The real value of critics is temporal resolution. They help most when long reasoning traces contain locally diagnostic partial structure.
  4. This is a control-variate story, not a reward-model story. The objective remains the same; only the variance changes.

My read is that this is the right way to bring critics back into large-scale LLM RL. It does not pretend the GRPO era was wrong. It accepts the systems constraints that killed old actor-critic pipelines, then reintroduces value functions only where they can be mathematically defended and operationally hedged.