Pith. sign in

REVIEW 3 major objections 5 minor 14 references

A hybrid DQN-DDPG agent that jointly picks per-user tokenizers, subchannels, and beamforming vectors can reduce video freezing by 68% compared with conventional H.265-based transmission while improving PSNR.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Joint tokenizer/codebook selection, subchannel assignment, and beamforming for multi-user video TokenCom is posed as an MDP and solved by DQN for discrete choices and DDPG for beamforming, with simulated gains over H.265.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Interesting problem formulation, but the RL algorithm as described is internally inconsistent and the evaluation is a table lookup, so the headline numbers don't yet stand up. the 3 major comments →

arxiv 2602.12338 v2 pith:SWYQIO2R submitted 2026-02-12 cs.LG

Wireless TokenCom: RL-Based Tokenizer Agreement for Multi-User Wireless Token Communications

Farshad Zeinali , Mahdi Boloursaz Mashhadi , Rahim Tafazolli This is my paper
classification cs.LG
keywords token communicationstokenizer agreementsemantic communicationsreinforcement learningDQNDDPGvideo transmissionresource allocation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that in multi-user wireless TokenCom, the choice of tokenizer should not be fixed or identical for everyone: the base station should adaptively agree on a tokenizer/de-tokenizer pair with each user, jointly with resource-block assignment and beamforming, based on channel conditions. It formulates this as a mixed-integer non-convex optimization and solves it with a hybrid RL agent: a DQN picks discrete tokenizers and subchannels, while a DDPG generates continuous beamformers. Simulations on a Rayleigh fading downlink with four candidate video tokenizers show the learned policy keeps video quality high while using less power. The headline quantitative result is a roughly 68% reduction in video freezing events relative to a conventional H.265 scheme at 1080p with 16 users, and about a 10 dB PSNR gain in a favorable 4-user, 64-antenna setting.

Core claim

The core claim is that tokenizer agreement in TokenCom can be treated as a learnable joint decision with wireless resource allocation, rather than a fixed protocol step. The authors establish, within their simulation environment, that a hybrid DQN-DDPG policy that selects one of four pre-trained video tokenizer/de-tokenizer pairs for each user at episode start, assigns resource blocks each slot, and shapes beamforming vectors, adapts to heterogeneous user channels and semantic quality constraints. This adaptation is what yields lower freezing rates and higher PSNR than agnostic TA, fixed TA, a discretized DDPG-only baseline, and an H.265-based conventional codec.

What carries the argument

The central mechanism is a hybrid action-space RL decomposition: a deep Q-network handles the discrete variables (which compatible tokenizer/de-tokenizer pair to agree on per user, and which resource blocks to assign), while a deep deterministic policy gradient actor-critic handles the continuous beamforming vectors. The two are coupled through a shared reward equal to the system utility, which balances normalized semantic quality (a non-increasing function f(η) of the tokenizer's compression rate, instantiated by a rate-distortion table) against transmit-power cost, with penalties for violating rate, quality, and resource constraints.

Load-bearing premise

The central premise is that if the Shannon rate is met, the user actually receives the exact tabulated PSNR for the chosen tokenizer, so no packet loss, decoding errors, or delay affect quality.

What would settle it

Run the same scenario over a physical or packet-level link where transmitted tokens can be lost or corrupted, decode the received video, and measure actual PSNR/freezing; if the proposed policy does not still beat H.265 by the reported margins (or if the margins shrink below statistical noise), the central claim would be falsified in that regime. A cheaper check: replace the deterministic q=f(η) table with a stochastic quality function that depends on SINR and compare learned policies.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If tokenizer choice is optimized per user per episode, users with weak channels can be assigned low-rate tokenizers while strong users get high-rate ones, raising average semantic quality under a fixed power budget.
  • The framework's freezing rate stays near zero across resolutions and increases only marginally from 4 to 16 users, suggesting it scales to denser downlink setups.
  • The reported 68% freezing-rate reduction over H.265, if it transfers beyond simulation, would make token-based video a practical alternative to conventional codecs in bandwidth-limited multi-user networks.
  • Because the reward can use any rate-distortion/perception metric (PSNR, SSIM, rFVD), the same agent can be retargeted to different semantic quality objectives without changing the RL structure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The quality model q=f(η) is a deterministic table and the rate constraint assumes that meeting Shannon rate suffices for exact PSNR; real channel errors would add distortion that the current simulation does not model, so the absolute gains (68%, 10 dB) are upper bounds on what a physical system would likely achieve.
  • The hybrid DQN-DDPG decomposition is general beyond video: the same discrete-continuous split applies to image, audio, and multimodal TokenCom, where the tokenizer set and quality metric would change but the action-space structure would not.
  • A natural test is to run the learned policy against a live codec chain with actual packetization and decoding; if freezing is measured per received frame rather than by rate-constraint violation, the ranking among methods could shift.
  • The approach implicitly assumes the cost of the tokenizer-agreement handshake is negligible; in very short communication episodes, that overhead would reduce the resource-efficiency gain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies downlink multi-user video TokenCom, in which a base station jointly selects a tokenizer/de-tokenizer pair per user, assigns resource blocks, and designs beamforming vectors. The authors formulate the problem as a mixed-integer non-convex program and propose a hybrid DQN-DDPG algorithm: DQN handles tokenizer agreement and RB assignment, while DDPG generates beamforming vectors. Simulation results, obtained with a deterministic mapping from tokenizer compression rate to PSNR, are reported to show lower freezing rates and higher PSNR than several baselines, including a 68% freezing reduction over a conventional H.265-based scheme.

Significance. If correct, this would be a timely contribution: it is among the first to formulate tokenizer agreement as part of a wireless resource-allocation problem and to apply hybrid discrete-continuous RL in that setting. The paper is clearly written in its problem formulation, the constraint structure is explicit, and the simulation setup is mostly reproducible from Tables II and III. The main weaknesses are that the RL algorithm as stated is not instantiable from the text, and the reported video-quality metric is the same deterministic function used in the reward, so the numerical claims are partly circular. These issues are load-bearing and prevent the results from being accepted at face value.

major comments (3)
  1. [Sec. III-A and III-D, Eq. (4), Algorithm 1] The DQN update is ill-posed for the tokenizer-selection action. The action space in Sec. III-A includes m_i, which is fixed at episode start and cannot be changed at t>1. However, Eq. (4) and the text in Sec. III-D take a Bellman max over the full action a' at every transition, implicitly allowing future tokenizer switches. Also, Algorithm 1 line 4 sets m_i by arg max Q(s1) without specifying the action argument of Q, while line 7 uses the same Q for RB assignment. No action masking, factored Q head, or otherwise restricted max is described. As written, the DQN target overestimates the value of the episode-start tokenizer choice, and the algorithm cannot be instantiated or reproduced. This affects all numerical results in Sec. IV.
  2. [Sec. II-A, Sec. III-C, Sec. IV, Table III, Fig. 1] The performance evaluation is circular. The paper defines quality as q=f(η), uses PSNR values from Table III computed from tokenizer compression rates, includes the normalized q̄ in the per-step reward (Sec. III-C), and then reports 'average PSNR' as an achieved outcome in Fig. 1. Since the environment produces exactly the q that the reward maximizes, the reported PSNR gains are partly a restatement of the objective rather than an external decoded-video quality measure. The freezing metric is also defined only as the Shannon-rate constraint (2c) being unsatisfied. To support the claims, the paper should evaluate end-to-end decoded video quality (or a held-out quality metric not used in the reward) and account for channel-induced errors in the token stream.
  3. [Sec. IV, Table II, Fig. 1a] The scalability and convergence claims are not supported by the described architecture. For U=16, R=16, N=64, the DDPG actor must output U·R·N complex beamforming coefficients, i.e., tens of thousands of continuous values per step, and the DQN over U·R binary variables faces an effectively intractable discrete action space. The paper does not describe how these outputs are generated, how the DQN argmax is computed, or how training is stabilized at these dimensions. Reporting a single freezing-rate curve up to 500 episodes without multiple seeds, confidence intervals, or convergence criterion is insufficient to substantiate the claimed scalability and 'stable adaptation'.
minor comments (5)
  1. [Eq. (2c)] q_i^t is written with a time index but q_i^t = f(η_{m_i}) is constant within an episode; consider writing q_i or making the time dependence explicit.
  2. [Algorithm 1, line 4] The notation m_i ← arg max Q(s1) is ambiguous. It should specify the action over which the argmax is taken, e.g., m_i^* ← arg max_m Q(s1, m, κ=0).
  3. [Page 4, after Algorithm 1] The sentence 'While complexity increases with...' is grammatically incomplete. It should say 'The complexity increases...' or be merged with the following sentence.
  4. [Table III] The row labeled 'HEVC medium' is not a tokenizer/de-tokenizer pair in the same sense as the Cosmos and BSQ-VAE entries; clarify whether it is being used only as a rate-distortion anchor, and cite the relevant source for the PSNR value.
  5. [Fig. 1] The plots show central tendencies without error bars or multiple-seed statistics. Given the large action spaces and the stochastic nature of RL, reporting seed variance would materially strengthen the comparison.

Circularity Check

2 steps flagged

Reported PSNR and freezing gains partly restate the paper's own reward/constraint-penalty terms, so the headline performance numbers are not independent external benchmarks.

specific steps
  1. self definitional [Sec. II-B (q^t_i = f(η_mi), normalized qbar), Sec. III-C (reward r_t = U_t - Σλ_pen Υ), Sec. IV (training/evaluation on PSNR from Table III)]
    "The distortion/perception quality achieved for user i is then given by qt i = f (ηmi). We normalize qt i using ¯qt i = qt i −qmin qmax−qmin ... The instantaneous reward follows the utility function rt = Ut − Σ8 n=1 λpenΥt n ... In this simulation, we train the RL agent using PSNR as the quality metric ... For performance evaluations, we use the PSNR."

    The 'achieved PSNR' reported in Fig. 1 is the same q = f(η) that enters the normalized quality ¯q in the per-step utility and reward. Because the agent is trained to maximize this reward, the reported PSNR gains are a direct restatement of the objective's quality term rather than an independently measured end-to-end video reconstruction quality. The paper does not simulate actual decoding/packet loss; the 'achieved' PSNR is set by the chosen tokenizer's Table III value, which is exactly the quantity the reward maximizes.

  2. fitted input called prediction [Sec. II-B constraint (2c), Sec. III-C reward penalty, Sec. IV freeze definition and Fig. 1a-b]
    "Freezing of the video can occasionally occur for a user in a time step, if the wireless transmission data rate for that user falls below the bitrate required for its selected tokenizer, i.e., constraint in (2c) is unsatisfied in that time step. ... The instantaneous reward follows the utility function rt = Ut − Σ8 n=1 λpenΥt n, where Υt n = 1 ... if constraint n in time slot t is not satisfied, and 0 otherwise."

    The freeze event is defined as violation of constraint (2c), and constraint violations are the exact penalty terms in the reward. Therefore the reported 68% freezing reduction relative to H.265 is a report of the penalty indicator the agent is trained to minimize, not an independent measure of playback freezes from a real video pipeline. The 'freezing rate reduction' is thus partly the training objective itself, not an external result.

full rationale

The paper's central contribution—a hybrid DQN-DDPG framework for joint tokenizer agreement, RB assignment, and beamforming—is not itself circular: the optimization over channels, powers, and constraints is a genuine MDP with independent content. However, the two headline evaluation metrics coincide with terms already in the reward: PSNR is q = f(η) from Table III, which enters the utility via ¯q, and freezing is the unsatisfied constraint (2c), which is penalized directly in r_t. Consequently, the abstract's claims of 'highest PSNR' and '68% freezing reduction' are at least partly restatements of the objective rather than externally validated video-quality measurements. This is partial circularity (score 6), not full circularity, because the agent still has to learn feasible allocations and beamforming. The separate issue that the DQN Bellman update in Eq. (4) ignores the one-shot nature of the tokenizer action is an algorithmic correctness concern, not a circularity concern, and is therefore not counted in this score.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central contribution is an RL formulation; the load-bearing external inputs are the precomputed rate-distortion values of four tokenizers, the assumed Shannon-rate/beamforming model, and hand-tuned utility weights. No new physical entity is postulated. The paper supplies no code or data, so re-implementation would require guessing network and channel details.

free parameters (3)
  • Utility weights α, β = α=2.0, β=1.0
    Hand-set in Table II to balance quality and power; no sensitivity analysis is provided, and all reported trade-offs depend on them.
  • Constraint penalty weight λ_pen = 2.0
    Chosen in Table II; controls how strongly constraint violations (including freezing) are penalized, so the freezing-rate comparison is sensitive to it.
  • Normalization range [qmin, qmax] = [18, 36] dB
    Quality is normalized as q̄_i=(q_i−q_min)/(q_max−q_min), which changes reward scale and DQN rankings; values appear arbitrary.
axioms (5)
  • domain assumption Video quality is given by a deterministic function q=f(η) of tokenizer compression rate; the PSNR values in Table III are exact achieved qualities whenever (2c) is satisfied.
    Sec. II-A defines q=f(η); Sec. IV uses Table III PSNR as the training and evaluation metric. No end-to-end video decoding is simulated.
  • domain assumption Each tokenizer/de-tokenizer pair is pre-trained and compatible only with itself, requiring TA; all modeled users/devices can run any pair in M_i.
    Sec. II-A: 'each tokenizer is compatible only with its corresponding de-tokenizer'; Sec. II-B assumes 'sufficient computational capabilities'.
  • domain assumption The Shannon-rate model in (1), with Rayleigh block fading and linear beamforming, describes the actual achievable rate with no channel estimation error or packet loss below rate.
    Sec. II-B Eq. (1); Sec. IV uses 'Rayleigh block fading channels'. Constraint (2c) treats rate as exact.
  • domain assumption The token ID stream for a chosen tokenizer requires exactly ρ×η×H×W bps with no protocol, coding, or packetization overhead.
    Sec. II-A: 'rate required in bits per second is ρ×η×H×W'; used in constraint (2c).
  • ad hoc to paper The hybrid DQN-DDPG converges within 500 episodes for the stated action-space dimensions (e.g., U=16, N=64, R=16 implies thousands of continuous outputs).
    No convergence proof or sample-complexity analysis is given; the reported results depend on this unstated assumption.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Wireless TokenCom: RL-Based Tokenizer Agreement for Multi-User Wireless Token Communications." pith.science (2026). https://pith.science/paper/SWYQIO2R

@misc{pith2026260212338,
  author       = {Pith},
  title        = {Pith review of: Wireless TokenCom: RL-Based Tokenizer Agreement for Multi-User Wireless Token Communications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWYQIO2R}},
  note         = {Machine review of arXiv:2602.12338}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Token Communications (TokenCom) has recently emerged as an effective new paradigm, where tokens are the unified units of multimodal communications and computations, enabling efficient digital semantic- and goal-oriented communications in future wireless networks. To establish a shared semantic latent space, the transmitters/receivers in TokenCom need to agree on an identical tokenizer model and codebook. To this end, an initial Tokenizer Agreement (TA) process is carried out in each communication episode, where the transmitter/receiver cooperate to choose from a set of pre-trained tokenizer models/ codebooks available to them both for efficient TokenCom. In this correspondence, we investigate TA in a multi-user downlink wireless TokenCom scenario, where the base station equipped with multiple antennas transmits video token streams to multiple users. We formulate the corresponding mixed-integer non-convex problem, and propose a hybrid reinforcement learning (RL) framework that integrates a deep Q-network (DQN) for joint tokenizer agreement and sub-channel assignment, with a deep deterministic policy gradient (DDPG) for beamforming. Simulation results show that the proposed framework outperforms baseline methods in terms of semantic quality and resource efficiency, while reducing the freezing events in video transmission by 68% compared to the conventional H.265-based scheme.

Figures

Figures reproduced from arXiv: 2602.12338 by Farshad Zeinali, Mahdi Boloursaz Mashhadi, Rahim Tafazolli.

Figure 1
Figure 1. Figure 1: Performance results. cross-modal context-aware semantic communications,” IEEE Wireless Communications, vol. 32, no. 5, pp. 80–88, 2025. [2] L. Qiao, M. B. Mashhadi, Z. Gao, R. Schober, and D. G ¨und ¨uz, “ToDMA: Large model-driven token-domain multiple access for semantic communications,” arXiv preprint arXiv:2505.10946, 2025. [Online]. Available: https://arxiv.org/abs/2505.10946 [3] W. Yang, H. Du, Z. Q. … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 3 linked inside Pith

  1. [1]

    L. Qiao, M. B. Mashhadi, Z. Gao, R. Tafazolli, M. Bennis, a nd D. Niyato, “Token communications: A large model-driven fra mework for 0 100 200 300 400 500 Episodes 0 20 40 60 80 100Freezing rate Proposed-TA DDPG-TA Agnostic-TA Fixed-TA Conventional (a) Freezing rate reduction. 360p 480p 720p 1080p Video resolution (H W) 0 10 20 30 40 50 60 70 80Average of...

    2025

  2. [2]

    ToDMA: Large model-driven token-domain multiple access f or semantic communications,

    L. Qiao, M. B. Mashhadi, Z. Gao, R. Schober, and D. G¨ und¨ u z, “ToDMA: Large model-driven token-domain multiple access f or semantic communications,” arXiv preprint arXiv:2505.10946 , 2025. [Online]. Available: https://arxiv.org/abs/2505.10946

    Pith/arXiv arXiv 2025

  3. [3]

    Semantic communications for future in ternet: Fundamentals, applications, and challenges,

    W. Y ang, H. Du, Z. Q. Liew, W. Y . B. Lim, Z. Xiong, D. Niyato, X. Chi, X. Shen, and C. Miao, “Semantic communications for future in ternet: Fundamentals, applications, and challenges,” IEEE Communications Surveys & Tutorials , vol. 25, no. 1, pp. 213–250, 2023

    2023

  4. [4]

    Beyond transmitting bits: Context, s emantics, and task-oriented communications,

    D. G¨ und¨ uz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Y ang,A. Y ener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, s emantics, and task-oriented communications,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 1, pp. 5–41, 2023

    2023

  5. [5]

    Generative AI for semantic communication: Architecture, challenges, and outlook,

    L. Xia, Y . Sun, C. Liang, L. Zhang, M. A. Imran, and D. Niyat o, “Generative AI for semantic communication: Architecture, challenges, and outlook,” IEEE Wireless Communications , vol. 32, no. 1, pp. 132– 140, 2025

    2025

  6. [6]

    DRL -Driven Dynamic Resource Allocation for Task-Oriented Semantic Co mmunica- tion,

    H. Zhang, H. Wang, Y . Li, K. Long, and A. Nallanathan, “DRL -Driven Dynamic Resource Allocation for Task-Oriented Semantic Co mmunica- tion,” IEEE Transactions on Communications , vol. 71, no. 7, pp. 3992– 4004, 2023

    2023

  7. [7]

    Optimizing Semantic Spec tral Efficiency in Wireless Image Transmission: A PPO-Driven Res ource Allocation Scheme,

    S. Y ang, B. Shen, and X. Huang, “Optimizing Semantic Spec tral Efficiency in Wireless Image Transmission: A PPO-Driven Res ource Allocation Scheme,” IEEE Communications Letters , vol. 29, no. 6, pp. 1466–1470, 2025

    2025

  8. [8]

    Performance Optimization for Semantic Communicat ions: An Attention-Based Reinforcement Learning Approach,

    Y . Wang, M. Chen, T. Luo, W. Saad, D. Niyato, H. V . Poor, and S. Cui, “Performance Optimization for Semantic Communicat ions: An Attention-Based Reinforcement Learning Approach,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 9, pp. 2598–2613, 2022

    2022

  9. [9]

    Beamforming design for semantic-bit coexisting communic ation sys- tem,

    M. Zhang, G. Zhu, R. Jin, X. Chen, Q. Shi, C. Zhong, and K. Hu ang, “Beamforming design for semantic-bit coexisting communic ation sys- tem,” IEEE Journal on Selected Areas in Communications , vol. 43, no. 4, pp. 1262–1277, 2025

    2025

  10. [10]

    The perception-distortion tr adeoff,

    Y . Blau and T. Michaeli, “The perception-distortion tr adeoff,” IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2018

    2018

  11. [11]

    On the rate-distortion-perception function,

    J. Chen, L. Y u, J. Wang, W. Shi, Y . Ge, and W. Tong, “On the rate-distortion-perception function,” IEEE Journal on Selected Areas in Information Theory , vol. 3, no. 4, pp. 664–673, 2022

    2022

  12. [12]

    Com- parison of the coding efficiency of video coding standards—i ncluding high efficiency video coding (HEVC),

    J.-R. Ohm, G. J. Sullivan, H. Schwarz, T. K. Tan, and T. Wi egand, “Com- parison of the coding efficiency of video coding standards—i ncluding high efficiency video coding (HEVC),” IEEE Transactions on Circuits and Systems for Video Technology, vol. 22, no. 12, pp. 1669–1684, 2012

    2012

  13. [13]

    Cosmos world foundation model platform for physical A I,

    “Cosmos world foundation model platform for physical A I,” 2025. [Online]. Available: https://arxiv.org/abs/2501.03575

    Pith/arXiv arXiv 2025

  14. [14]

    Image and video tokenization with binary spherical quantization,

    Y . Zhao, Y . Xiong, and P . Kr¨ ahenb¨ uhl, “Image and video tokenization with binary spherical quantization,” 2024. [Online]. Avai lable: https: //arxiv.org/abs/2406.07548

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.