Privileged Info Guided Dreamer

• Privileged Information Guided Dreamer is a paradigm that uses training-only signals (e.g., simulator states, TIPs) to guide latent representation learning.
• It enhances policy safety and rapid adaptation by employing asymmetric training with enriched critic and estimator objectives.
• Empirical evaluations in sim-to-real transfer, safe RL, and HIP-MDPs show significant performance gains and robust, rapid adaptation.

Searching arXiv for the named methods and closely related papers to ground the article. Searching for "Privileged Information Guided Dreamer" and related terms on arXiv. Privileged Information Guided Dreamer denotes a class of Dreamer-derived world-model reinforcement learning methods in which information available during training but unavailable at deployment is used to improve latent representation learning, value estimation, safety modeling, or rapid adaptation. Within the cited formulations, privileged signals include simulator state, extra sensors, object poses, hazard positions, friction coefficients, mass scaling factors, goal coordinates, previous actions, and LLM-derived Task-Invariant Properties (TIPs). The common principle is asymmetric training: the learning system is allowed to use more information than the deployed actor, while test-time control remains conditioned on ordinary observations or observation-derived latent states (Liang et al., 3 Apr 2026, Huang et al., 4 Aug 2025, Byrd et al., 17 Feb 2025).

1. Problem settings and motivating failure modes

The three main formulations arise from distinct but closely related settings. In quadruped sim-to-real locomotion, the environment is framed as a POMDP

$$(\mathcal S,\mathcal O,\mathcal A,\mathcal P,\mathcal R,\gamma),$$

with return

$R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$

The central diagnosis is that standard Dreamer-style learning can overfit to simulator-specific dynamics parameters because its latent state is optimized primarily for reconstructing observations and predicting transitions under the training simulator. Under shifted real-world dynamics, the learned features and policy become brittle (Liang et al., 3 Apr 2026).

In safe partially observable reinforcement learning, the relevant formalism is the constrained objective

$$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$

with

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$

PIGDreamer introduces Asymmetric Constrained Partially Observable Markov Decision Processes (ACPOMDPs) to formalize the advantage of allowing a training-time critic to access privileged information. The stated motivation is that CPOMDP critics without privilege tend to underestimate safety risk, so partial observability degrades both return and safety (Huang et al., 4 Aug 2025).

In hidden-parameter Markov decision processes (HIP-MDPs), each episode is governed by an unobserved parameter sampled once per episode and fixed during that episode. The process is written as

$$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$

and the policy objective is

$$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$

PrivilegedDreamer is motivated by the claim that treating hidden parameters merely as additional variance is often insufficient, especially when rewards are parameterized by hidden variables (Byrd et al., 17 Feb 2025).

2. Privileged information as a training-only resource

Privileged information is defined consistently as information available during training but not at deployment, but the concrete instantiations differ across methods.

Method	Privileged information	Training-time role
DreamTIP	linear velocity, elevation maps, friction coefficients, center-of-mass position, foot contact forces	compute TIPs such as contact stability and terrain clearance
PIGDreamer	underlying simulator state, object poses, hazard positions, proprioceptive readings, previous actions, velocity/acceleration/gyro sensors, goal coordinates	privileged representation alignment, privileged predictors, asymmetric critics
PrivilegedDreamer	ground-truth hidden parameter such as friction coefficient, mass scaling factor, motor scaling	supervised parameter estimation and conditioning

DreamTIP introduces Task-Invariant Properties as high-level properties that are strongly correlated with task success, generalize across related locomotion tasks, and are robust to changes in low-level dynamics. The paper explicitly gives contact stability and terrain clearance as examples. These properties are derived from privileged simulator state via

$f_t = TIP_{extractor}(s_t),$

where

$TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$

The method is therefore privileged-information-guided in the sense that privileged states are used during training to define auxiliary targets, but the final policy does not require privileged signals online (Liang et al., 3 Apr 2026).

PIGDreamer generalizes the same asymmetry to safe RL under partial observability. Privileged information may include hidden geometry, hidden dynamics, extra sensors, temporal information, and simulator state. The core idea is to use privileged information during training for representation alignment, reward and cost prediction, and asymmetric critics, while ensuring that the deployed actor uses only observation-derived latent states (Huang et al., 4 Aug 2025).

PrivilegedDreamer uses privileged labels of the true hidden parameter during training. Typical examples are contact friction, pendulum mass scaling factor, ball mass scaling factor, object mass scaling factor, and x/y motor scaling factors. At test time, the true hidden parameter disappears; the method instead performs history-based online estimation (Byrd et al., 17 Feb 2025).

3. Architectural realizations

DreamTIP builds on a Dreamer variant with an RSSM: $$h_t = f_\theta(h_{t-1}, z_{t-1}, a_{t-1}), \qquad z_t \sim q_\theta(z_t \mid h_t, o_t), \qquad \hat z_t \sim p_\theta(\hat z_t \mid h_t), \qquad \hat o_t \sim p_\theta(\hat o_t \mid h_t, z_t).$$ Its distinctive addition is a properties predictor: an MLP on top of $[h_t,z_t]$ that predicts $R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$0. In the downstream controller, the hidden state $R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$1 is combined with the current observation $R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$2 and fed to a PPO policy,

$R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$3

The paper emphasizes that DreamTIP modifies world-model representation learning, while downstream policy optimization is carried out with PPO rather than standard Dreamer latent-imagination actor-critic equations (Liang et al., 3 Apr 2026).

PIGDreamer is built on DreamerV3 and uses two coupled world models. The naive world model is trained from observations $R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$4, while the privileged world model is trained from privileged information $R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$5. In the naive model, the paper defines an oracle posterior

$R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$6

which is richer than the observation-only posterior

$R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$7

The actor uses only the naive latent state,

$R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$8

while the critics use both naive and privileged latents,

$R_t=\sum_{i=0}^{\infty}\gamma^i r_{t+i}.$9

Its imagination mechanism is explicitly named Twisted Imagination and rolls out synchronized latent trajectories for horizon $$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$0 (Huang et al., 4 Aug 2025).

PrivilegedDreamer extends DreamerV2 with a dual recurrent architecture. The baseline RSSM retains the deterministic recurrent state $$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$1 and stochastic latent state $$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$2, but an external LSTM + MLP estimator predicts an intermediate hidden-parameter estimate

$$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$3

and the RSSM adds a hidden-parameter prediction head

$$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$4

The representation model is conditioned on $$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$5,

$$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$6

and the actor and critic are conditioned on $$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$7,

$$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$8

The architecture is thus explicit system identification embedded inside Dreamer-style imagination (Byrd et al., 17 Feb 2025).

4. Objectives, adaptation mechanisms, and deployment behavior

DreamTIP augments the RSSM training objective with TIP supervision. Its world-model objective is written as

$$\max_{\pi} J_r(\pi) \quad \text{s.t.} \quad J_c(\pi) \le d,$$9

For real-world adaptation, it adapts the world model only while keeping the policy frozen. The method uses a mixed replay buffer, freezes the recurrent model, and regularizes the adaptable stochastic latent $$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$0 against a frozen reference model $$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$1 with negative cosine similarity: $$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$2 The adaptation objective is

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$3

The stated purpose is to mitigate representation collapse, catastrophic forgetting, and latent drift during low-data real-world calibration (Liang et al., 3 Apr 2026).

PIGDreamer uses a stop-gradient symmetric KL-style representation loss,

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$4

and defines

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$5

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$6

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$7

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$8

with the overall world-model objective

$$J_r(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t r_t\right], \qquad J_c(\pi)=\mathbb E_\pi\left[\sum_{t=0}^{\infty}\gamma^t c_t\right].$$9

Its actor is optimized with an augmented Lagrangian objective,

$$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$0

and the critics use TD($$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$1) targets over imagined trajectories (Huang et al., 4 Aug 2025).

PrivilegedDreamer retains DreamerV2-style imagined actor-critic training but adds a supervised hidden-variable loss: $$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$2 The imagined $$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$3-return is

$$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$4

with actor and critic losses

$$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$5

A central deployment fact is that the true hidden parameter is unavailable at test time; only the learned estimator remains active (Byrd et al., 17 Feb 2025).

A frequent misconception is that privileged-information-guided Dreamer requires privileged inputs online. In these formulations, that is not the case. DreamTIP uses TIPs only as auxiliary prediction targets, PIGDreamer discards privileged world models and critics at deployment, and PrivilegedDreamer removes the true hidden-parameter label and relies on online estimation from history (Liang et al., 3 Apr 2026, Huang et al., 4 Aug 2025, Byrd et al., 17 Feb 2025).

5. Experimental domains and reported empirical behavior

Setting	Reported result	Paper
Quadruped sim-to-real transfer	average performance improvement of 28.1% across eight simulated transfer tasks; real-world Climb task: baseline 10% success, DreamTIP 100%	(Liang et al., 3 Apr 2026)
Safe partially observable RL	state-of-the-art aggregate performance on Safety-Gymnasium; often near-zero-cost performance	(Huang et al., 4 Aug 2025)
HIP-MDP rapid adaptation	41% higher average reward over the tested baselines on average; mean $$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$6	(Byrd et al., 17 Feb 2025)

DreamTIP is evaluated on a Unitree Go2 quadruped in Isaac Gym and on real hardware with an Intel D435i depth camera and onboard Orin Nano. The observation space includes proprioceptive information and depth images; the privileged space includes linear velocity, elevation maps, friction coefficients, center-of-mass position, and foot contact forces. Simulation and real-world terrains include Stair, Climb, Tilt, Crawl, Gap, CoM Transfer, Velocity Transfer, and Compound Task. The paper reports an average performance improvement of 28.1% across eight simulated transfer tasks. On Crawl, WMP drops from about 33.51 at the easiest level to 5.66 at the hardest level, whereas the full method goes from 36.58 to 25.35. In real-world evaluation, Table I reports Stair $$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$7, Climb $$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$8, Tilt $$\boldsymbol{\theta} \sim p(\boldsymbol{\theta}), \qquad s_{t+1} \sim p_{\boldsymbol{\theta}}(s_{t+1}\mid s_t,a_t), \qquad r_t = r_{\boldsymbol{\theta}}(s_t, a_t),$$9, and Crawl $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$0 success for the full method, compared with WMP values of $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$1, $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$2, $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$3, and $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$4. The paper also reports that using $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$5 real trajectories is a practical choice because performance gains are large from 3 to 5 trajectories and show diminishing returns beyond 5 (Liang et al., 3 Apr 2026).

PIGDreamer is evaluated on Safety-Gymnasium from $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$6 RGB observations. The main tasks shown are CarGoal1, PointButton1, PointGoal2, PointPush1, and RacecarGoal1. Metrics are average episodic return and average episodic cost return over $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$7 episodes with episode length $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$8, and results are recorded at 4M environment steps. The paper reports state-of-the-art performance on Safety-Gymnasium, with the claim that PIGDreamer not only satisfies safety constraints at convergence but often achieves near-zero-cost performance. In the appendix comparison against Distill, the reported numbers are: SafetyPointGoal2 reward $$J(\pi) = \mathbb{E}_{\boldsymbol{\theta}\sim p(\boldsymbol{\theta}),\,\tau\sim p_{\boldsymbol{\theta},\pi}} \left[ \sum_{t=0}^{T-1}\gamma^t r_{\boldsymbol{\theta}}(s_t,a_t) \right].$$9, cost $f_t = TIP_{extractor}(s_t),$0; SafetyCarGoal1 reward $f_t = TIP_{extractor}(s_t),$1, cost $f_t = TIP_{extractor}(s_t),$2; SafetyRacecarGoal1 reward $f_t = TIP_{extractor}(s_t),$3, cost $f_t = TIP_{extractor}(s_t),$4; SafetyPointPush1 reward $f_t = TIP_{extractor}(s_t),$5, cost $f_t = TIP_{extractor}(s_t),$6; SafetyPointButton1 reward $f_t = TIP_{extractor}(s_t),$7, cost $f_t = TIP_{extractor}(s_t),$8. The ablation study reports that removing privileged representation alignment causes a large drop, while privileged critics alone provide only marginal improvement (Huang et al., 4 Aug 2025).

PrivilegedDreamer is evaluated on five HIP-MDP tasks: Walker Run, Pendulum Swingup, Throwing, Kuka Sorting, and Pointmass. Hidden parameters include contact friction, pendulum mass scaling factor, ball mass scaling factor, object mass scaling factor, and x/y motor scaling factors. The paper reports training for 2 million timesteps, evaluation over 100 runs per seed, and mean/std over 3 seeds. Table 2 gives mean scores of $f_t = TIP_{extractor}(s_t),$9 for PrivilegedDreamer, $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$0 for DreamerV2, $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$1 for SAC, $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$2 for PPO, and $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$3 for RMA. The largest gains occur on Sorting and Pointmass, where rewards are parameterized by hidden parameters: on Sorting, DreamerV2 scores $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$4 while PrivilegedDreamer scores $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$5; on Pointmass, DreamerV2 scores $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$6 while PrivilegedDreamer scores $TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$7. The paper also notes that SAC is best on Throwing (Byrd et al., 17 Feb 2025).

6. Interpretation, limitations, and relation to adjacent methods

These works suggest three distinct uses of privilege inside Dreamer-style world models. DreamTIP uses simulator-only signals to construct semantically meaningful auxiliary targets, PIGDreamer uses privileged information to align observation-based latents and train asymmetric critics for safe RL, and PrivilegedDreamer uses privileged labels to learn explicit hidden-parameter estimation for rapid adaptation. In relation to adjacent literature, DreamTIP is described as resembling learning with privileged information, teacher-student distillation, and simulator-only auxiliary supervision; PrivilegedDreamer is positioned against domain randomization, domain adaptation, system identification, and meta-RL; and PIGDreamer is presented as DreamerV3 plus safe constrained optimization, training-only privilege, and ACPOMDP theory (Liang et al., 3 Apr 2026, Huang et al., 4 Aug 2025, Byrd et al., 17 Feb 2025).

Theoretical support is strongest in PIGDreamer. ACPOMDP introduces asymmetric value estimation by allowing the training-time critic access to underlying states. The paper states Theorem 3.3 as

$TIP_{extractor}=\mathrm{LLM}(I_{text}, I_{priv}).$8

and interprets this as showing that asymmetric information yields a better or equal optimal policy than purely symmetric partial-observation learning. At the same time, the paper explicitly notes what it does not provide: finite-sample error bounds, explicit safety violation probability reductions, and explicit PAC sample-complexity gains (Huang et al., 4 Aug 2025).

The limitations are method-specific. DreamTIP explicitly notes performance degradation over prolonged operation due to compounding world-model prediction errors, and its behavior depends on TIP construction. PIGDreamer notes that privileged information does not always help; PointButton1 is the cited case where PIGDreamer does not outperform SafeDreamer because the available privileged information does not provide meaningful information gain for the behavior required. It also highlights two-world-model complexity and the absence of formal hard-safety guarantees at deployment. PrivilegedDreamer assumes an episode-wise stationary hidden parameter, requires privileged labels during training, and can suffer from misestimation, non-identifiability, and long-horizon model bias; the Throwing task is the example where SAC outperforms the model-based approach (Liang et al., 3 Apr 2026, Huang et al., 4 Aug 2025, Byrd et al., 17 Feb 2025).

A broader synthesis suggested by these papers is that privileged information is most useful when it disambiguates reward-, cost-, or dynamics-relevant latent factors that ordinary observations do not expose directly. Under that reading, privileged-information-guided Dreamer is less a single algorithm than a design principle: use training-only information to shape latent world models and critics, but preserve deployment-time autonomy by restricting the executed policy to observation-derived state (Liang et al., 3 Apr 2026, Huang et al., 4 Aug 2025, Byrd et al., 17 Feb 2025).