{"id":"01296f5a-1c62-4af0-851a-1e316e365a9e","arxiv_id":"2504.19806","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A tri-level PPO-based training scheme with an adaptive gradient-aware task weighting module improves multi-task semantic broadcast communication performance over conventional and prior learned baselines.","lead":"This paper proposes a three-level reinforcement learning system for broadcasting one image to multiple receivers that run different tasks, such as classifying or reconstructing the image. The authors report that the system outperforms standard compression and previous learned codecs in noisy channels by adaptively balancing task priorities.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's convergence guarantee rests on Assumption 1's c-strong convexity of the PPO encoder loss, which the clipped, quantized, stochastic-policy objective in Eq. (14) cannot satisfy; the theoretical claim therefore does not cover the implemented system.","rationale":"The reader's weakest assumption correctly identifies the load-bearing flaw: Theorem 1's convergence proof depends on Assumption 1, which asserts strong convexity of the encoder loss with respect to theta. This assumption is not satisfied by the actual clipped PPO surrogate over a Gaussian policy with binary quantization, and the paper provides no argument that it holds. The theoretical claim is therefore not established for the simulated system. Because the empirical comparisons and the adaptive-weighting ablation are plausible and could be independently valuable, the appropriate response is conditional acceptance with a required revision of the convergence claim or verification under realistic assumptions, rather than rejection of the entire paper. The reader already reached CONDITIONAL, and my analysis agrees with that assessment, so no verdict change is needed.","tokens_in":26299,"tokens_out":4789,"duration_ms":54495,"concrete_test":"Take one trained checkpoint of SemanticBC-TriRL on MNIST at CBR=0.02 and one on CIFAR-10, and compute the minimum eigenvalue of the Hessian of L_TX from Eq. (14) with respect to theta over a fixed large batch, using the same straight-through quantization and channel sampling as in training. If any checkpoint yields a negative eigenvalue whose magnitude is comparable to the gradient norm, Assumption 1's c-strong convexity is violated exactly at the operating point used for Theorem 1. To make this decisive, evaluate at several points along the training trajectory and report the spectrum; a uniformly positive-definite Hessian would be the only outcome consistent with the stated assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical contribution is Theorem 1 in Section IV-D, whose non-asymptotic rate (Eq. 19) is derived from Lemma 2 and Lemma 3 under Assumption 1. Assumption 1(1) requires f(v_l) = L_TX(v_l; {phi*_n,l}) to be c-strongly convex in theta_l and L_f-smooth. However, the optimized objective in Eq. (14) is a weighted sum of the clipped PPO surrogate (12), the value loss (13), and supervised decoder losses, evaluated through a stochastic Gaussian policy (10), 1-bit quantization QTX (2), and channel noise. The clipped surrogate is a minimum of ratio terms and is generally neither convex nor smooth in network parameters; binary quantization is non-differentiable unless a straight-through estimator is silently introduced, which itself breaks the smoothness premise. The claim that boundedness of mu(t) and A(t) implies strong convexity does not follow. Consequently Lemma 2, imported from [43]'s strongly-convex bi-level analysis, has no valid premise here, and Theorem 1's rate cannot be asserted for the simulated SemanticBC-TriRL. The analogous singleton condition on decoders (unique phi*_n,l after kappa=100 local updates) is also not guaranteed for nonconvex decoder networks, so the first-level exactness used in (8a) is approximate at best. The empirical results may still stand, but the formal convergence claim is unsupported by the actual objective's properties.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SemanticBC-TriRL, a tri-level alternating optimization framework for semantic broadcast communications with one shared transmitter-side encoder and multiple receiver-side task-specific decoders. The three levels are: (i) supervised local decoder updates, (ii) PPO-based shared encoder updates with a stochastic Gaussian policy and binary quantization, and (iii) a multi-gradient aggregation weight module that adaptively assigns task weights. The authors formulate the system as a constrained tri-level optimization problem, propose Algorithm 1, and state Theorem 1 with a non-asymptotic convergence rate of O(Γ(H/2) + √η + √(1/(Lη))). Simulations on MNIST and CIFAR-10 over AWGN, Rayleigh, and Rician channels compare against Deep JSCC, BPG+LDPC, and an equal-weight variant, reporting higher SSIM and classification accuracy at low SNR, together with scalability up to six receivers.","tokens_in":26739,"tokens_out":4413,"duration_ms":44127,"significance":"If the theoretical and empirical claims hold, the framework is a practically useful extension of semantic broadcast to heterogeneous receivers, and the adaptive weight module addresses a real gradient-conflict problem. The paper's strengths include the clear problem decomposition, the comparison with an equal-weight counterpart that isolates the effect of the weight assignment, and the breadth of experiments across datasets, channel models, CBR values, and numbers of receivers. The empirical curves consistently point in the claimed direction, particularly the low-SNR robustness that is relevant for semantic communications. However, the central theoretical contribution, Theorem 1, is not supported for the implemented system because Assumption 1 is implausible for the clipped PPO objective, and the proof imports key lemmas from strongly-convex bi-level literature without verifying their premises. The empirical study also lacks statistical error bars and code release, which limits confidence in the magnitude of the reported gains.","major_comments":[{"comment":"Assumption 1 and Theorem 1 do not cover the system actually simulated in Section V. The encoder loss L_TX in Eq. (14) is a weighted combination of the clipped PPO surrogate (12), the value loss (13), and decoder losses, evaluated through the stochastic Gaussian policy (10), the binary quantizer Q_TX in (2), and channel noise. This objective is not c-strongly convex in theta_l and is not differentiable unless a straight-through estimator is silently introduced; the paper neither states nor analyzes such an estimator. Since Lemmas 2 and 3 are imported from the strongly-convex bi-level analyses of [43] and [23], their premises do not hold for the clipped, quantized, stochastic-policy objective, so Eq. (19) cannot be asserted for SemanticBC-TriRL as trained. Please either prove the required properties for the actual objective, replace the theorem with a convergence statement for an idealized smooth unclipped surrogate and clearly delimit its scope, or explicitly label the convergence analysis as heuristic.","section":"IV-D"},{"comment":"The decoder singleton condition in (8a) is not satisfied by the implemented training procedure. The manuscript assumes a unique optimal decoder phi*_{n,l} after sufficient local updates, but the decoder networks in Tables IV and V are non-convex, and Table III sets kappa=100, which does not guarantee convergence to a global optimum. Because the second-level gradients in (17b) and (18) use phi*_{n,l} while Algorithm 1 uses phi^kappa_{n,l}, the first-level exactness in (8a) is approximate at best. The paper should justify or relax this condition, for example by stating an inexact lower-level bound in the spirit of [28].","section":"III-B"},{"comment":"The gradient formulas for the weight module rely on an undocumented simplification after Eq. (16). The sentence 'it is assumed that the clipping of a single sample does not impact the overall gradient computation' asserts away the non-differentiability of the min and clip operations in Eq. (12). Without specifying a surrogate gradient, the expressions for nabla_theta F(v_l) and nabla_w F(v_l) in (17a)-(17b) are not derivatives of the actual loss (14). This directly affects the descent direction d_l in Eq. (15) and the weight updates in Algorithm 1. Please state the gradient estimator used and analyze its bias, or reformulate F(v_l) to use a smooth surrogate.","section":"IV-C"}],"minor_comments":[{"comment":"The caption reads 'Notions used in this paper' but should read 'Notations used in this paper'.","section":"Table II"},{"comment":"Equation (13) has a formatting error: V_chi(s(t)) is missing a closing parenthesis in the manuscript text, which should be corrected.","section":"Eq. (13)"},{"comment":"In the comparison schemes list, the bullet 'JSCC: JSCC share the same encoder...' contains a subject-verb agreement error, and the word 'comparision' is misspelled as 'comparision'.","section":"V-A"},{"comment":"The paper would benefit from releasing code and the full set of training hyperparameters, and from reporting error bars or multiple seed statistics for the performance curves in Figs. 5, 8, and 13, since the reported differences are sometimes small (e.g., the SSIM gap between SemanticBC-TriRL and SemanticBC-TriRL-EW).","section":"V-B"},{"comment":"In Eq. (17a), the reward term uses inconsistent indices: the subscript in Theta_{phi^kappa_{1,l}} appears to be an artifact, and the notation for the N receivers should be harmonized with (11) and (17b).","section":"Eq. (17a)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript leans heavily on the authors' own prior work [12], [40] for the alternating self-critical/PPO training mechanism and on [23], [43] for the convergence machinery; the incremental contribution is the combination with the tri-level weight-assignment module. The convergence theorem needs either substantial re-scoping or a corrected set of assumptions, but the empirical system-level comparison could still be publishable if the theory is presented honestly. I do not see grounds for questioning novelty disclosure beyond the heavy self-citation pattern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful engineering recipe, not a fundamental advance, and the convergence result is the weak link. The tri-level stack—supervised decoder updates, PPO-trained shared encoder, multi-gradient task weighting—is new as a combination for heterogeneous semantic broadcast. I don't see that exact assembly in the cited literature. The comparison work is also decent: Deep JSCC, BPG+LDPC, and an equal-weight ablation, under AWGN and Rayleigh, on MNIST and CIFAR-10. The adaptive weighting clearly does something: the curves are consistent, and the equal-weight ablation is the right control.\n\nThe main problem is exactly where the stress-test note points. Theorem 1 imports Lemmas 2 and 3 from [43]/[23], and those lemmas require c-strong convexity of the lower-level objective and a singleton decoder optimum. The implemented encoder loss in (14) is a clipped PPO surrogate over a stochastic Gaussian policy, with binary quantization and channel noise. That objective cannot be c-strongly convex. The assertion in Assumption 1 that bounded μ(t) and A(t) imply Lipschitz gradients is a non sequitur; bounded first moments do not give strong convexity, and the rest of the paragraph merely claims Hessian boundedness, which also does not follow. The decoder singleton condition is equally unsatisfied after κ=100 local updates on nonconvex nets. So the non-asymptotic rate in (19) is a correct conditional statement about an idealized problem, but it is not a convergence proof for the system that was simulated. That claim should be reframed or removed.\n\nSecondary soft spots: no error bars, no code, and several implementation details about the quantized PPO pipeline—especially how gradients pass through QTX(·)—are left unspecified. The authors lean on their own earlier works [12], [40] for foundational pieces, which is legitimate, but the reliance is heavy enough that reproducibility suffers. Those are fixable.\n\nEmpirically the paper is plausible; I don't see the components being fraudulent or the curves fake. It is a reasonable systems contribution: it tells people how to combine alternating training with adaptive task weights and gives evidence that it helps. A serious referee should engage with it, but should send back the theory for revision. If the authors either condition the theorem on the actual objective or honestly reposition it as heuristic motivation, the paper becomes acceptable.","headline":"Useful engineering recipe for heterogeneous semantic broadcast, but the convergence theorem is overclaimed and should be revised before publication.","tokens_in":27168,"tokens_out":5101,"would_cite":true,"duration_ms":52251,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SemanticBC-TriRL claims that a tri-level alternating framework—supervised decoder updates, a PPO-driven shared encoder, and an adaptive multi-gradient task-weighting module—resolves conflicts between image reconstruction and…","keywords":["semantic broadcast communications","multi-objective optimization","tri-level optimization","self-critical reinforcement learning","proximal policy optimization","multi-gradient aggregation","joint source-channel coding","image reconstruction"],"falsifier":"Run the tri-level training on a batch and record the stationarity gap $\\psi(v_l)=\\|-\\nabla F(v_l)-\\lambda_l \\nabla g(v_l)\\|^2$ and the feasibility gap $g(v_l)$; if the strong-convexity assumption holds, these should decay roughly like $O(\\Gamma(H/2)+\\sqrt{\\eta}+1/\\sqrt{L\\eta})$, and with $\\eta=O(L^{-1/2})$ the gap should fall as $O(\\Gamma(H/2)+L^{-1/4})$. Any observed divergence or much slower convergence would indicate that the theory is not describing the implemented system.","tokens_in":26124,"feed_emoji":"📡","tokens_out":9855,"duration_ms":91442,"temperature":0.7,"pith_summary":"This paper proposes SemanticBC-TriRL, a training scheme for semantic broadcast systems in which one encoder at the transmitter serves multiple receivers with different tasks, specifically image reconstruction and image classification. The central claim is that a tri-level alternating optimization—task-specific decoders trained by supervised learning, a shared encoder trained by proximal policy optimization (PPO), and an adaptive task-weighting module built on multi-gradient aggregation—resolves conflicts among heterogeneous receiver objectives. If the claim holds, a single transmitted bitstream can simultaneously support several downstream tasks with higher SSIM and classification accuracy than end-to-end Deep JSCC, conventional BPG+LDPC coding, or an equal-weight variant, particularly at low signal-to-noise ratios. The paper also provides a non-asymptotic convergence guarantee for the alternating procedure.","feed_headline":"Tri-level RL scheme balances competing wireless receiver tasks","feed_subtitle":"Alternating decoder training, PPO encoder updates, and adaptive task weighting keeps image quality and accuracy high.","key_machinery":"The load-bearing mechanism is the tri-level alternating learner with a multi-gradient aggregation weight module. The first level relies on the singleton condition: for a fixed encoder parameter $\\theta_l$, each decoder has a unique optimum $\\phi^*_{n,l}$ after $\\kappa$ local updates. The second level reformulates the encoder as a reinforcement-learning agent: its policy is a Gaussian distribution over channel symbols, and PPO (a policy-gradient update that clips how far the policy moves in one step) maximizes the task-weighted reward $r=\\sum_n w_n \\Theta_n(m,\\hat{m}_n)$ with an auxiliary supervised loss. The third level computes the gradient of the encoder loss with respect to the weights and the encoder, subject to the approximated feasibility constraint, and updates the weights so that tasks whose gradients point against the common descent direction receive lower weight. This closed-form quadratic-program update is what lets the scheme rebalance conflicting objectives during training.","core_discovery":"The paper's discovery is that the conflict among heterogeneous receiver tasks in semantic broadcast can be converted into a constrained tri-level optimization problem with a bottom-up alternating solution. At the first level, each receiver's decoder performs $\\kappa$ local gradient steps toward its own task loss, giving an approximate decoder optimum $\\phi^*_{n,l}$ for the current encoder parameter $\\theta_l$. At the second level, the encoder is updated by PPO, treating the encoder as an agent whose Gaussian sampling policy produces channel symbols and whose reward is the task-weighted receiver performance; the PPO loss is augmented with a decoder-side supervised loss to stabilize the update. At the third level, a weight-assignment module finds a descent direction for the joint variable $v_l=(\\{w_{n,l}\\},\\theta_l)$ by minimizing the first-order Taylor term of the encoder loss subject to a constraint on an approximated feasibility function, solved with a Lagrange multiplier. The paper proves in Theorem 1 that, under strong-convexity and Lipschitz-gradient assumptions, the procedure reaches a weak stationary point with error $O(\\Gamma(H/2)+\\sqrt{\\eta}+1/\\sqrt{L\\eta})$, and reports that this scheme outperforms the baselines on MNIST and CIFAR-10, with the largest gains at low SNR.","pith_inferences":["Inference, not stated in the paper: the multi-gradient weighting module could transfer to other multi-objective joint source-channel coding settings, for example detection plus reconstruction, whenever each objective supplies a differentiable reward, since the quadratic-program update does not depend on the specific task losses.","Inference, not stated in the paper: because the encoder is frozen during decoder updates, the scheme could be deployed incrementally, adding a new receiver's decoder and then fine-tuning the encoder with adjusted weights while old receivers keep working.","Inference, not stated in the paper: a natural stress test is to initialize the two task gradients in direct opposition and check whether the learned weights track the Pareto front computed by explicit multi-objective search; the paper reports weight trajectories but does not compare them to the true Pareto set.","Inference, not stated in the paper: if the strong-convexity assumption is the bottleneck, replacing it with a Polyak-Łojasiewicz condition or a smoothed relaxation would likely preserve a similar non-asymptotic rate and make the proof applicable to the actual deep-network objective."],"forward_implications":["A single transmitter can serve receivers with different tasks from one encoded bitstream, avoiding per-task dedicated transmissions and reducing the transmitter's computational load.","At low SNR, the adaptive weighting should yield graceful degradation for both reconstruction and classification, in contrast to the abrupt cliff effect of conventional BPG+LDPC transmission.","Adding a new receiver task requires training only its decoder and letting the weight module rebalance the encoder loss, which is the scalability claim validated with up to six receivers.","For fixed inner steps $H$, increasing the outer iteration count $L$ and choosing $\\eta=O(L^{-1/2})$ drives the tri-level procedure to a weak stationary point at rate $O(\\Gamma(H/2)+L^{-1/4})$.","The performance difference from the equal-weight variant supports the claim that learned task weights, not just the larger model, drive part of the improvement."],"supporting_citations":[{"why":"Supplies the alternating self-critical learning strategy that the tri-level scheme extends.","marker":"[12]"},{"why":"Provides the first-order multi-gradient algorithm for multi-objective bi-level optimization that underwrites the weight-assignment and convergence analysis.","marker":"[23]"},{"why":"Supplies Lemma 2 and Lemma 3, the gradient-norm and feasibility bounds from which Theorem 1 derives its rate.","marker":"[43]"},{"why":"Defines the Deep JSCC baseline for image recovery and classification that the experiments compare against.","marker":"[15]"},{"why":"Provides the original Deep JSCC architecture used as a comparative baseline and as the encoder structure in the JSCC scheme.","marker":"[29]"},{"why":"Defines the conventional BPG+LDPC baseline whose cliff effect at low SNR is contrasted with the proposed scheme.","marker":"[44]"},{"why":"Motivates viewing multi-task learning as Pareto multi-objective optimization, the perspective behind the adaptive task weighting.","marker":"[21]"},{"why":"Supplies the barrier/constraint method used to set the control parameter $\\rho_l$ in the quadratic-program update.","marker":"[42]"},{"why":"Provides the singleton condition that justifies treating each decoder's locally optimized parameters as unique optima at the first level.","marker":"[28]"}],"fun_headline_variants":["RL resolves task conflicts in semantic broadcast","Tri-level RL steers multi-task semantic transmission","Adaptive task weights optimize semantic broadcast encoder","SemanticBC-TriRL: RL balances conflicting receiver tasks","Reinforcement learning settles multi-task semantic broadcast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence theorem requires the encoder loss to be strongly convex in the encoder parameters with Lipschitz-continuous gradients, and each decoder to reach a unique optimum after its local updates; these conditions are not guaranteed by the deep-network PPO objective with binary quantization and channel noise, so the stated rate may not apply to the system actually simulated.","fun_headline_variants_meta":{"raw":{"variants":["RL resolves task conflicts in semantic broadcast","Tri-level RL steers multi-task semantic transmission","Adaptive task weights optimize semantic broadcast encoder","SemanticBC-TriRL: RL balances conflicting receiver tasks","Reinforcement learning settles multi-task semantic broadcast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1483,"prompt_tokens":1046,"completion_tokens":437,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":366}},"tokens_in":662,"tokens_out":437,"duration_ms":4510,"temperature":1.0,"reasoning_tokens":366,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:43:47.647441+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the tri-level training on a batch and record the stationarity gap $\\psi(v_l)=\\|-\\nabla F(v_l)-\\lambda_l \\nabla g(v_l)\\|^2$ and the feasibility gap $g(v_l)$; if the strong-convexity assumption holds, these should decay roughly like $O(\\Gamma(H/2)+\\sqrt{\\eta}+1/\\sqrt{L\\eta})$, and with $\\eta=O(L^{-1/2})$ the gap should fall as $O(\\Gamma(H/2)+L^{-1/4})$. Any observed divergence or much slower convergence would indicate that the theory is not describing the implemented system.","supporting_citations":[{"cited_title":"Self- critical alternate learning based semantic broadcast communication,","cited_arxiv_id":null,"evidence_quote":"Supplies the alternating self-critical learning strategy that the tri-level scheme extends."},{"cited_title":"A first-order multi-gradient algorithm for multi-objective bi-level optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the first-order multi-gradient algorithm for multi-objective bi-level optimization that underwrites the weight-assignment and convergence analysis."},{"cited_title":"Bome! bilevel optimization made easy: A simple first-order approach,","cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 2 and Lemma 3, the gradient-norm and feasibility bounds from which Theorem 1 derives its rate."},{"cited_title":"Semantic communications for image recovery and classification via deep joint source and channel coding,","cited_arxiv_id":null,"evidence_quote":"Defines the Deep JSCC baseline for image recovery and classification that the experiments compare against."},{"cited_title":"Deep joint source- channel coding for wireless image transmission,","cited_arxiv_id":null,"evidence_quote":"Provides the original Deep JSCC architecture used as a comparative baseline and as the encoder structure in the JSCC scheme."},{"cited_title":"Design of low-density parity check codes for 5g new radio,","cited_arxiv_id":null,"evidence_quote":"Defines the conventional BPG+LDPC baseline whose cliff effect at low SNR is contrasted with the proposed scheme."},{"cited_title":"Multi-task learning as multi-objective opti- mization,","cited_arxiv_id":null,"evidence_quote":"Motivates viewing multi-task learning as Pareto multi-objective optimization, the perspective behind the adaptive task weighting."},{"cited_title":"Bi-objective trade-off with dynamic barrier gradient descent,","cited_arxiv_id":null,"evidence_quote":"Supplies the barrier/constraint method used to set the control parameter $\\rho_l$ in the quadratic-program update."},{"cited_title":"A generic first-order algorithmic framework for bi-level programming beyond lower-level singleton,","cited_arxiv_id":null,"evidence_quote":"Provides the singleton condition that justifies treating each decoder's locally optimized parameters as unique optima at the first level."}],"review_version":1}