Resource management for distributed edge networks hosting different services has been handled efficiently using Reinforcement learning (RL). Reward generalization in complex workflows is still an open issue within the RL realm. Therefore, this work proposes RAGAM-AI, a hybrid framework that combines the capabilities of RL algorithms and Generative-AI (Gen-AI) for powering an Agentic AI workflow that optimizes resource allocation at the Multi-access Edge Computing (MEC) units. This framework further resolves the reward generalization issue by utilizing Large Language Model (LLM) for feedback. A Multi- Agent Soft-Actor Critic (MASAC) using Math{\Sigma}tral as a reward oracle is implemented for managing resource allocations at the MEC. RAGAM-AI interprets system state information and evaluates multi-objective trade-offs guiding the resource orchestration loop over distributed edge server agents. The complexity of the framework is further brought down by incorporating surrogate reward model. The overall performance is evaluated against the single agent SAC, REINFORCE and random policies for benchmarking. On using RAGAM-AI, lower allocation deviations for CPU, memory and bandwidth resources are reported while minimizing the latency and data rate costs. A scalable AI as a feedback (AIAF) solution optimizing network resources with significant performance improvement is thus presented.
Large language models (LLMs) undergo safety alignment to prevent harmful generations, yet they remain vulnerable to jailbreak attacks. Recent work reveals a deeper structural weakness: safety alignment can be highly sparse, such that pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior. In this work, we take the first steps toward strengthening safety alignment such that it is less dependent on sparse model components. We propose Addition by Subtraction, which gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model. Our results show that the resulting models are substantially more robust to pruning-based attacks. Remarkably, these benefits generalize beyond the white-box attacks motivating our design: Addition by Subtraction improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks. The finetuned models also exhibit improved truthfulness while largely preserving general capabilities. Comprehensive experiments across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.
Agent trajectories often contain both successful decisions and failed attempts, yet optimization with sparse trajectory-level rewards struggles to identify which critical decisions should be reinforced or suppressed. We observe that meaningful distinctions among on-policy trajectories can be distilled into natural-language process rubrics and used as process supervision signals. Based on this insight, we propose Online Rubric Self-Distillation (ORSD), which leverages self-derived process rubrics for dense token-level supervision, thereby augmenting outcome-centric reinforcement learning. Specifically, the current policy independently generates multiple rubric bundles at each step according to the same group of trajectories, extracting candidate process rubrics from cross-trajectory behavioral differences. A frozen verifier then evaluates all rubric items against each trajectory in the group, and selects one candidate bundle based on bundle-level alignment between rubric satisfaction and task success. ORSD treats the selected rubric bundle as privileged information during training, then compares the log-likelihoods of the original tokens with and without the rubric. The resulting likelihood gap provides token-specific process credit, which is added to the original trajectory-level GRPO advantage through additive advantage shaping. ORSD thus bridges trajectory-level outcome supervision and fine-grained process supervision, enabling token-level credit assignment using process signals. Extensive experiments on ALFWorld and WebShop show that ORSD accelerates policy improvement and boosts final task performance for both Qwen3-1.7B and Qwen3-4B under the same environment-interaction budget. Our code is available at https://github.com/AnnaXChun/ORSD.
Accurate electricity load forecasting provides critical decision support for energy procurement, generation scheduling, and supply-demand balancing, contributing to the economic efficiency and operational reliability of power systems. However, future electricity demand can be affected by the complex interplay of multiple events described in Web notices, weather alerts, and event schedules. Their effects vary over time and can reinforce or offset one another, making electricity load difficult to forecast accurately. While existing methods can incorporate context as additional input, how to use the effects of multiple events and their interactions to guide the load forecasting process itself remains to be explored. To address this gap, we propose ElecLoad, a context-aware agent for multi-event electricity load forecasting that adaptively organizes the forecasting process. Specifically, we train ElecLoad to combine historical load with context to analyze each event's relevance to the target load, the timing of its effects, and potential interactions among events. Guided by this analysis, ElecLoad adaptively chooses to forecast the entire forecast horizon directly or to partition it into segments according to changes in event effects. When the horizon is partitioned, sub-agents independently generate initial forecasts for each segment. The main agent then reviews these forecasts, and the sub-agents perform selective refinement guided by its feedback. Experiments on multiple electricity load datasets show that ElecLoad achieves state-of-the-art performance in most evaluation scenarios. Code is available at https://github.com/zhjai/ElecLoad.
Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substantial context overhead and dilute the agent's attention to task-relevant information. Existing token pruning and summarization methods remain constrained by discrete text representations, limiting the achievable compression ratios. In this work, we challenge the prevailing assumption that agents must consume tool observations as explicit text tokens. To this end, we propose Coca-Cola, a continuous latent context compression model for LLM agents. Coca-Cola compresses tool observations into latent tokens and embeds them into the input embedding space of the agent. We further introduce Action-Preserving Training, which aligns the agent's next-action distributions under full and compressed observations. This trains the compression model to learn action-sufficient latent tokens that retain decision-critical information. Experiments with web-search agents on four QA benchmarks and two model families show that Coca-Cola reduces the agent's token consumption by 70.01%-89.24% and end-to-end runtime by 24.45%-55.97%, while consistently improving task performance. These results suggest that Coca-Cola offers a promising approach for context compression in agents. Our code is available at https://github.com/THU-Agent/Coca-cola.
Short-interval wireless schedulers must return a complete, feasible beam-and-resource allocation before each execution deadline, and a late or invalid output gives no utility in the current control cycle. Searching longer can increase transmitted traffic, but it may also miss the deadline or leave only an infeasible intermediate solution. We propose Feasibility-Preserving Transactional Refinement (FPTR), which always keeps a validated feasible allocation, the incumbent. Each refinement stage builds a separate candidate and leaves the incumbent unchanged. A candidate replaces the incumbent only if it is complete, passes the feasibility validator before its stage cutoff, and strictly improves the objective; otherwise it is discarded. Four refinement stages use this rule: buffer-aware resource selection, compatibility-group (CG) user sharing, beam-mask rebuilding from remaining demand, and joint reassignment of two resources. On 150 fixed synthetic instances, the full pipeline (Full) raises traffic by 3.83% on average over a lightweight beam-first reference and by 1.79% over its Base constructor. In a separate suite with matched inputs, Full exceeds six independent search baselines (ALNS, tabu search, GA, simulated annealing, iterated local search, and GRASP) by 6.68-8.35%, where outputs later than 100 ms receive zero utility. All 750 Full runs in the internal suite returned feasible allocations, and none exceeded 100 ms. Separating mandatory output construction from optional refinement lets spare computation time improve the allocation without endangering the incumbent. The open-source implementation supports independent reproduction.
Setting. This is an empirical characterization study; we propose no new method and run no comparisons against alternative objectives. We fine-tune Qwen3.5-4B with sparse LoRA and length-normalized DPO plus an SFT retention term (the RPO objective) on structured financial-filing tasks, and track downstream capability as the implicit reward margin grows to $\sim 7$ $\beta$-scaled units (nf4 training; $\sim 3$ under fp16 measurement on the unquantized base). Question. Does margin growth degrade capability uniformly across tasks, or selectively? Findings. In a pre-registered study (8 seeds, 6 unseen at freeze) with artifact-aware scoring and TOST equivalence testing, the alignment tax is graded and task-selective: extraction is preserved ($\Delta = -0.55$ pts, equivalent at $\pm5$ pts), json loses little ($\Delta = -3.52$ pts, equivalence fails on a clustered item-level interval), and router degrades severely ($\Delta = -30$ pts). Router's loss concentrates in a margin 1-3 damage window containing a transient format-break spike. The margin is accounted for almost entirely ($\geq 96\%$) by suppression of rejected completions, with chosen reward approximately flat, consistent with the retention term anchoring it; we do not attribute the behavioral damage to any single gradient channel. Gold-logp deflation is task-selective (extraction $-0.15$, json $-1.04$, router $-1.47$ nats/token under fp16), with gold selectively protected relative to incorrect controls on extraction (8/8 seeds). QLoRA reward margins differ up to $3.4\times$ between nf4 training and unquantized fp16 measurement, with an axis-dependent gap. Scope. Length-normalized rewards with SFT retention, sparse LoRA on 8/32 layers, one 4B model, easily-separable synthetic preferences, three structured tasks. A no-retention control and a measured format-alignment metric are left to future work.
Microfluidic droplet generation is critical for many applications, yet efficient device design in high-dimensional parameter spaces remains computationally prohibitive due to complex multiphase flow dynamics. This work presents a unified methodology that couples adaptive surrogate enhancement with simulation-driven inverse design. High-fidelity Volume-of-Fluid simulations provide ground-truth data, while neural network ensembles predict flow regimes and droplet diameters. Ensemble-based uncertainty quantification identifies epistemic uncertainty using consensus levels for regime classification and the coefficient of variation for diameter prediction, revealing that uncertainty concentrates near regime transition boundaries where physics is highly sensitive to parameter perturbations. A physics-informed enrichment strategy targets these critical zones and addresses class imbalances by deploying additional targeted simulations and retraining the surrogates forming the ensembles. A framework that integrates Particle Swarm Optimization for constrained inverse design, balancing accuracy with predictive confidence and physical consistency is therefore proposed. Validation on independent test datasets demonstrates that the enriched ensembles achieve higher accuracy, with the fraction of predictions within the 0-3% error range on droplet diameter nearly doubling, and high-error cases (above 10%) dropping substantially. The inverse design procedure reliably identifies physically realizable configurations meeting target droplet sizes. This approach significantly reduces computational costs while ensuring reliable predictions, offering a framework with potential applicability to the optimization of a broad range of microfluidic systems.
Traditional document parsing systems use Markdown as their output representation, which is effective for linear text but does not fully preserve fine-grained spatial structure, table geometry, visual hierarchy, or rendering semantics. We show that HTML is advantageous for heterogeneous and layout-sensitive documents whose representations must remain editable, inspectable, and reusable, and introduce AmbiHTMLBench to systematically study this setting. In this benchmark, given a real-world document, a model must recover executable HTML that preserves its content, organization, and visual appearance. The benchmark contains 1,115 diverse pages with structured annotations and evaluates predictions in two different ways: The objective protocol aligns ground-truth and predicted rendered blocks and measures table fidelity, page-level text error, and a two-dimensional layout consistency; a complementary reference-anchored protocol measures perceived fidelity by aligning LLM judges with human preferences. Across ten primary multimodal systems, the best objective score is 0.8092, while the subjective winner is different, demonstrating that automatic decomposition and human-perceived fidelity capture complementary properties. We further fine-tune a small-parameter expert model and find that task-specific adaptation yields good performance on document-to-HTML reconstruction task. These results position AmbiHTMLBench as a diagnostic testbed for building smarter and more logical document parsing systems.
Long-horizon language-model agents accumulate increasingly large interaction histories, making context compaction essential for efficient execution. Existing learned compaction methods often rely on downstream task outcomes or solver behavior, coupling training to specific tasks or agents. We introduce SumOne, a standalone summarizer trained to preserve information in agent histories through masked reconstruction. Given a summary of the expired history, a frozen reconstructor recovers randomly masked portions of the original trajectory; reconstruction quality, combined with a length penalty, provides a task-agnostic reward for optimizing the summarizer with GRPO. This yields self-supervision directly from raw trajectories, without reference summaries or downstream task execution. At inference time, only the lightweight summarizer is retained, enabling plug-and-play context compression for otherwise unchanged agents. We conduct a systematic evaluation of SumOne in long-horizon search and software-engineering tasks. SumOne reduces estimated inference cost by approximately 78\% on DeepSearchQA and 75\% on SWE-bench Verified relative to operating with full context, with absolute decreases of 0.04. The same checkpoint also transfers across downstream models and agent harnesses without additional fine-tuning.
Active embodied question answering requires agents to acquire evidence through viewpoint changes and physical interaction. The central challenge is to avoid undirected interaction: taking plausible actions without specifying what evidence they should reveal for answering the question. Action plans and accumulated histories alone can leave the informational purpose of an interaction implicit. We introduce EviSeek, a training-free framework that makes question-derived evidence demands explicit acquisition objectives, linking observation-conditioned actions and selectively retained evidence to the information the agent seeks. On 3,436 ESI-Bench questions, EviSeek improves accuracy over the Single-model agent by 7.1 percentage points with Qwen3-VL-32B and 3.7 points with Gemini 3.1 Pro, while improving visual-support-weighted accuracy and reducing mean token usage. These findings support making the informational purpose of interaction explicit, with action selection and evidence retention organized around what the agent needs to learn to answer the question.
Multimodal large language models (MLLMs) remain prone to visual hallucinations when correct responses depend on subtle visual details. One line of post-training methods constructs visual contrasts through image retrieval, generation, or editing to improve fine-grained perception. However, these contrasts are not built around the model's own hallucinations, making it difficult to ensure that supervision targets its actual perceptual confusions. Another line suppresses hallucinated outputs through response correction or preference learning under fixed image conditions, but leaves the design of supervision signals on the image-input side underexplored. Inspired by how humans correct misconceptions, we introduce DreamRealm, a post-training framework that explores diverse latent hallucinated responses, minimally edits original images to make them valid, and trains the model to assign each response to its supporting image. Our data construction pipeline, RealmForgePipe, mines observed and latent hallucinations from a target model and materializes them through image editing. Using hallucinations mined from Qwen3.5-9B, we construct the Realized Hallucination Dataset (RealmSet). Our training objective, RealmMatch, constructs a likelihood score matrix across image--response combinations and combines supervised learning with bidirectional contrastive losses to reinforce correct matches and suppress mismatches. Compared with the original Qwen3.5-9B model, DreamRealm improves performance across six visual perception and hallucination benchmarks, including gains of 7.26 percentage points in HallusionBench accuracy and 11.34 percentage points in MMVP pair accuracy. Together with ablation studies on data construction and training methods, these results demonstrate that hallucination assignment is a promising paradigm for improving visual perception and mitigating hallucinations.
Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance length across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00$\times$ and 5.71$\times$ decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03$\times$ and 3.67$\times$ for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48$\times$. Code is publicly available at https://github.com/fudan-generative-vision/D3D.
Brains and large language models need not be considered equivalent systems to reveal an interpretable commonality: both exhibit high-dimensional, nonlinear dynamics whose behavior can only be modeled imperfectly. This creates a shared challenge whenever we seek to intervene in their dynamics. A controller may either rely on a model identified around a particular operating regime and risk failure as the system moves away from that regime, or explicitly account for uncertainty and a changing internal state. Neurostimulation has confronted an analogous problem for decades. The limitations of fixed, open-loop stimulation have motivated closed-loop and state-dependent approaches that condition intervention on the ongoing state of the neural system. Inspired by this principle, we formulate LLM steering as a feedback-control problem in which the network's internal state, rather than a fixed intervention alone, determines how the system is perturbed. We further account for the mismatch between tractable dynamical models and the nonlinear transformer they approximate through robust control. As one principled realization of this framework, we develop an $H_\infty$-based steering controller that limits sensitivity to disturbances and linearization mismatches. Across multiple steering benchmarks, the controller improves performance over existing baselines, with particularly pronounced gains under out-of-distribution conditions and varying length complexities. Beyond controller performance, we introduce robust steerability, $S_{\mathrm{rob}}$, a pre-deployment measure of tolerance to dynamics mismatch; across model-task pairs, it exhibits task-specific structure and varies systematically with model scale. Beyond robust LLM steering, our measure asks not merely whether control is possible, but how much uncertainty it can survive; a principle that may span artificial and biological neural systems.
Offline reinforcement learning with verifiable rewards (RLVR) trains on a fixed pool of sampled responses and typically supervises every token of every response, although responses and tokens are not equally informative. We study which ones to supervise through the Gibbs-optimal policy of KL-regularized RLVR, which offline training targets, and measure the mismatch between a policy and this target with a bidirectional KL divergence. At the reference policy, this mismatch factorizes exactly and separates the two outcomes: correct responses contribute mainly through the forward term, concentrated on their low-probability tokens, and incorrect responses through the reverse term, concentrated on the responses the policy generates often and on their high-probability tokens. Based on this analysis, we propose GIST (Gibbs-Targeted Hierarchical Data Selection), a two-level selection method that approximates these principles with signals available before training. GIST uses the sampling temperature to select correct responses on-policy and incorrect responses near the policy's mode, and it uses the initial policy's token probabilities to supervise only the tokens where the mismatch lies. Applied to two offline training objectives, GIST raises the average accuracy over GSM8K, MATH-500, and MBPP for all three models under both objectives. In a controlled comparison on Qwen3-0.6B/GSM8K, it outperforms seven selection baselines that choose from the same candidates while supervising 66.6% fewer tokens than training on all candidates of the same prompts.
Effective collaboration in multi-agent systems requires mechanisms for coordination and adaptability in dynamic and uncertain environments. However, existing multi-agent systems rarely model resilience explicitly, instead relying on ad-hoc output correction that limits systematic, runtime management of robustness. In this work, we propose Multi-Agent Mesh (MA-Mesh), a framework that constructs a role-based self-healing layer on top of a collaboration pattern. Specifically, MA-Mesh generates resilience-aware system specifications and synthesizes them into executable multi-agent policy code. To enhance resilience, we introduce two self-healing mechanisms: Cross-agent reflection, a mechanism that allows agents to evaluate not only their own outcomes but also the performance of their peers, enabling informed behavioral adjustments and collective policy refinement; Role-based failover, a strategy which dynamically reassigns roles and enables agents to substitute for one another when sub-tasks fail or agents underperform, ensuring continuity of task execution. Experimental results across diverse tasks (e.g., software development, tool use, scientific reasoning) demonstrate that MA-Mesh significantly outperforms existing multi-agent systems in both task performance and robustness under system uncertainty.
As LLM-based agents develop, they have been increasingly applied in the legal domain. Legal Judgment Prediction~(LJP) requires models to conduct multi-step reasoning and external legal evidence retrieval. Existing legal search agents' performance is often constrained by the quality of retrieval queries. They often formulate queries directly from the initial user request, overlooking latent legal issues and legally salient details. To address this issue, we propose \emph{LegalEvolver}, a query optimization framework specifically designed to generate legally informative queries for LJP-oriented retrieval. Sampling high-quality optimized queries and their corresponding search trajectories is challenging, as relying solely on unguided high-temperature rollouts is both inefficient and constrained by the model's initial capabilities. We therefore heuristically optimize queries as seeds, improving the efficiency of trajectory generation. Furthermore, we summarize rollout experience into rubrics and use them to guide multi-round iterative self-evolution. Iteratively performing rubric optimization and rubric-guided RL enables the model to surpass its initial performance ceiling. End-to-end performance on extensive benchmarks demonstrates the effectiveness of query optimization on LJP. Comparisons with direct SFT followed by RL demonstrate the effectiveness of rubric guidance. Evolving rubrics further outperform fixed rubrics, confirming the benefits of self-evolution.
Context compression reduces inference cost by encoding inputs into compact representations while preserving semantic content. An open question is whether semantic preservation alone is sufficient to maintain downstream behaviours such as refusal alignment. We investigate this question and find that encoder-based compression systematically weakens refusal behaviour in instruction-tuned language models, despite high reconstruction fidelity. This effect persists across model families and compression architectures. Mechanistic analysis shows that compression attenuates activation along the decoder's learned refusal direction. We further explore Memory Steering, a lightweight inference-time intervention that restores refusal rates to near-baseline levels without retraining and operates entirely in compressed representation space. These results demonstrate that semantic preservation does not guarantee behavioural preservation under compression, highlighting the need to explicitly preserve alignment-relevant features in compression-aware systems.
AI's inflection point is a fountain of alarming risks. Digital platforms in particular are vulnerable to three major AI risks, as those in control of global platforms cede control to AI, of the algorithms, of membership and of content: (1) AI control of global platforms, social and economic, used by the majority of humanity, resulting in the majority of humanity becoming vulnerable to AI control. (2) Unaccountable AI agents operating anonymously on social and economic digital platforms, side-by-side with humans, possibly masquerading as humans, serving unknown masters with unknown goals, operating under unknown restraints without accountability or culpability. (3) Unattributed deepfakes generated by anonymous AI, with people unable to distinguish them from genuine content and with their producers evading accountability and culpability. Here, we propose to corral these AI risks with grassroots contracts, a digital incarnation of a two-centuries old notion of Pierre-Joseph Proudhon. We present grassroots contracts for a social graph with accountable AI agents and for a social network with digital provenance, and working smartphone apps that implement the two and show how they corral these major AI risks: (1) AI control of global platforms. (2) Unaccountable AI agents. (3) Unattributed deepfakes.
The halting problem asks whether a program eventually halts, with unrestricted resources. Engineering systems instead certify bounded halting: given a program, an input, and a step budget T, decide whether the program halts within T steps. This paper studies two of its uses. First, it defines an exact finite-budget problem, placing the simulated program and the checker on the same unit-cost timescale, so the checker can be embedded in an adversarial program. A diagonal argument with explicit additive overhead shows that no fixed deterministic checker decides every instance within any fixed fraction of T below one, for large T. In the nondeterministic case, T bounds the branch depth allowed for a stable observable decision. Second, bounded-halting instances give adversarial tests for finitary universality, since a universal model must support arbitrary finite computations as T grows. The tests check resource growth with T, the absence of a fixed deliberation bound, and later access to intermediate data. They separate a fixed model with extendable external resources, a model family that grows with T, and a larger system built around a fixed component. For machine-learning systems, this identifies the resource growth a universality claim assumes and whether it fails for deployed models.