Muhammad Zubair and Syed Khaliq Shah of Southeast University in Nanjing tackled what they call the integrated rebalancing and faulty bike recycling problem. Instead of sending trucks out on separate missions—one to redistribute bikes, another to collect broken ones—their framework folds both tasks into a single vehicle routing model. Each service truck carries a partitioned cargo hold: a main compartment for rebalancing bikes and a dedicated section for faulty ones. This partitioned capacity design means a vehicle can deliver bikes to deficit stations, collect surplus bikes from oversupplied ones, and scoop up damaged bicycles along the way, all in one tour.

The mathematics behind the model is demanding. The researchers formulated it as a variant of the pickup and delivery problem, with an objective function that balances three competing costs: user dissatisfaction when riders cannot find a bike, operational expenses including travel and loading time, and the handling cost of faulty bicycles. Loading or unloading an operational bike takes about one minute, while a damaged bike requires roughly two minutes of handling time—values drawn from real observations of Nanjing operators. Vehicles must complete their tours within a six-hour shift, and travel times are adjusted by traffic congestion factors derived from Baidu Maps data, ranging from free-flow conditions to heavy congestion.

Exact optimization tools like IBM’s CPLEX solver can crack such models, but only at small scale. The team first validated their formulation on 38 real stations near Southeast University, solving it to near-perfect optimality with percentage gaps as low as 0.02 percent. In 35 of the 38 stations, rebalanced inventory aligned almost exactly with forecast demand, and all faulty bikes were successfully routed back to the depot for repair or recycling. But real-world networks dwarf this testbed: Shanghai operators manage more than 50,000 dockless bikes, and New York’s CitiBike runs over 1,000 stations. Exact mixed-integer programming becomes computationally intractable beyond roughly 50 stations, which is where the researchers’ metaheuristic enters.

Their solution combines a genetic algorithm with adaptive large neighborhood search, wrapped in a cluster-first, route-second decomposition. The genetic algorithm evolves populations of candidate routes using tournament selection, order crossover, and mutation operators, while preserving the best solutions through elitism. The ALNS component then intensifies the search by destroying parts of a solution—removing stations at random, targeting the worst-performing segments, or dismantling entire clusters—and repairing them with greedy, regret, or cheapest insertion strategies. An adaptive weighting mechanism learns which destroy-and-repair combinations perform best, and a simulated annealing acceptance criterion decides whether to embrace worse solutions temporarily to escape local optima.

The clustering stage is where the approach departs most sharply from prior work. Rather than dividing a city into fixed geographic zones, the dynamic clustering procedure jointly optimizes the number of clusters, their centroids, and station assignments based on expected inter-cluster traversal costs. This captures a fundamental trade-off: more clusters shrink the distance vehicles travel within each zone but inflate the cost of moving between zones. In experiments on a 200-station instance, increasing clusters from 3 to 20 cut intra-cluster travel by 65 percent while inflating inter-cluster travel by more than 700 percent. The sweet spot landed at 11 clusters, where total routing distance fell to 362 kilometers—a 26.1 percent improvement over three clusters and a 21.8 percent improvement over twenty.

The performance numbers are striking. On 200-station synthetic instances, the hybrid algorithm produced solutions within 3.2 percent of best-known values in about 15 minutes, a fourfold scale improvement over the most comparable integrated studies, which stall at 50 to 100 stations. Against 50-station benchmarks, the method matched exact solvers with gaps below 0.5 percent while running 3.75 times faster. Dynamic clustering itself shaved 6.5 percent off routing distance compared with static, geography-only approaches. Meanwhile, folding faulty bike collection into rebalancing routes added just 8 to 12 percent computational overhead—yet eliminated the separate collection trips entirely, cutting total vehicle distance by 15 to 25 percent compared with running the two operations sequentially.

The practical implications are substantial. For a 200-station city, the researchers estimate daily savings of 50 to 75 kilometers of vehicle travel, translating to roughly 10 to 15 liters of fuel, two to three hours of driver time, and annual cost reductions of 15,000 to 25,000 US dollars per vehicle. The recycling loop itself follows a four-stage pipeline: IoT sensors and GPS monitoring flag faulty bikes based on battery status, signal strength, failed QR scans, and idle periods exceeding seven days; trucks with partitioned holds collect them during routine rebalancing; capacity constraints keep the two cargo types physically separated throughout the route; and the depot sorts arrivals into immediate repair, scheduled maintenance, or parts recovery.

Empirical research from major Chinese cities underscores why this integration matters. Faulty bike rates can exceed 2 percent during peak usage, and delayed collection measurably damages user satisfaction and safety. Earlier integrated models relied on exact optimization that buckled beyond 50 stations, and most treated vehicle capacity as a single undifferentiated constraint, implicitly assuming sequential loading and unloading. The new framework’s partitioned capacity model and jointly optimized clustering address both shortcomings, and the authors argue the approach could extend beyond bike sharing to fleet management, ride-sharing, and last-mile delivery, where decomposition is often treated as a necessary evil rather than an optimization opportunity.

Limitations remain. The model assumes deterministic demand within a single operational period and a homogeneous fleet, simplifications that real cities may strain against. The researchers point to several frontiers: incorporating stochastic demand through robust optimization, extending to multi-depot networks, adapting the algorithm for real-time rebalancing as conditions change mid-shift, integrating dynamic pricing to encourage users to self-balance the system, and pairing the optimizer with machine learning demand forecasts. For now, the study offers something operators have lacked: a computationally feasible way to keep dockless fleets balanced, functional, and profitable at the scale modern cities actually demand.

Subject of Research: Integrated rebalancing and faulty bike recycling optimization in dockless bike-sharing systems using a hybrid metaheuristic with dynamic clustering

Article Title: Integrated rebalancing and faulty bike recycling in dockless bike-sharing systems: a hybrid metaheuristic approach with dynamic clustering

Article References: Zubair, M., & Shah, S. K. (2026). Integrated rebalancing and faulty bike recycling in dockless bike-sharing systems: a hybrid metaheuristic approach with dynamic clustering. Frontiers in Future Transportation, 7, Article 1919819. https://doi.org/10.3389/ffutr.2026.1919819

Image Credits: AI Generated

DOI: 10.3389/ffutr.2026.1919819

Keywords: bike-sharing, dockless bikes, vehicle routing, metaheuristics, genetic algorithm, adaptive large neighborhood search, dynamic clustering, faulty bike recycling, urban mobility, optimization, Nanjing, transportation systems

The threat was systematized by the SafeRAG benchmark, which identified four novel attack types that abandon overt malicious signals in favor of stealth. Silver noise embeds partially correct, highly relevant information so that malicious fragments rank well in relevance scoring and slip past filters while diluting valid evidence. Inter-context conflict plants subtle contradictions across multiple retrieved passages, exploiting weaknesses in how language models reconcile conflicting sources. Soft advertisement weaves implicit promotional or toxic language into otherwise compliant text, evading standard toxicity detectors. White Denial-of-Service, or White DoS, disguises fabricated safety warnings as compliance alerts, tricking the model into refusing to answer altogether and degrading the availability of the entire system. Each of these attacks targets the blind spots of defenses built to catch obviously irrelevant or overtly violating content, rendering components such as natural language inference filters and knowledge compressors largely ineffective.

Existing multi-stage defense frameworks, including TrustRAG, which cleans retrieved evidence through clustering, integrates knowledge, and performs consistency self-verification, were never designed with these covert patterns in mind. Their clustering and filtering mechanisms rely on fixed parameters that cannot simultaneously handle multi-document clustered contamination, typical of silver noise and inter-context conflict, and single-document embedded manipulation, typical of soft advertisement and White DoS. Their generic integration rules provide no targeted handling for implicit contradictions, subtle manipulative phrasing, or induced refusals. And their post-hoc verification typically checks only factual consistency, overlooking implicit toxicity and availability risks. The result is a defense that either misses malicious chunks or discards valid evidence, and that can still be steered by a false evidence chain or quietly degraded by refusal induction.

MIRAGE, developed by researchers at Information Engineering University in Zhengzhou and the Institute of Information Engineering of the Chinese Academy of Sciences, attacks the problem in two chained stages. The first stage, clean retrieval, begins by embedding the retrieved documents and computing the local density of each document in the embedding space. Because different attacks leave different distributional fingerprints, this density analysis allows the framework to predict the attack type before filtering begins. Silver noise and inter-context conflict tend to produce multiple similar texts that cluster together with high intra-cluster similarity, since they are generated from the core propositions of genuine passages following consistent templates. Soft advertisement and White DoS, by contrast, appear as individually injected anomalies with low local density.

Once the distribution type is predicted, MIRAGE applies the appropriate detection strategy. For clustered attacks it selects the optimal cluster count via the elbow method and runs dynamic K-means clustering, flagging clusters whose internal semantic consistency is suspiciously high as likely products of batch rewriting. For embedded attacks it performs per-document outlier detection, combining low density with keyword checks for inducement, warning, or promotional terms. A second screening layer then computes pairwise ROUGE-L statistics across the retrieval results and constructs a dynamic threshold equal to the mean plus a scaled standard deviation, set at one standard deviation after a grid search. This statistical threshold adapts to the semantic distribution of each query, catching adversarially generated documents that reuse key phrases and sentence structures from original passages even when their embedding-space camouflage succeeds. A multi-factor reranking step, weighting relevance at 0.5, attack risk at 0.3, and density at 0.2, then prioritizes high-value clean documents for the context.

The second stage, conflict resolution, assumes that some malicious content may survive the first line of defense. MIRAGE first elicits the model’s internal knowledge as a factual anchor, on the rationale that this knowledge is uncontaminated by retrieval poisoning. The reranked documents are decomposed into fine-grained propositions, and four attack-adaptive integration rules are applied. For silver noise, each proposition is matched against the internal knowledge anchor by embedding cosine similarity and discarded if it deviates too far from the factual backbone. For inter-context conflict, conflicting proposition pairs are cross-validated against the anchor, and the more credible side is retained rather than relying on simple majority voting. For soft advertisement, a toxicity and implicit manipulation detector scores each proposition on promotional intent, implicit bias, and emotional manipulation, removing those above threshold. For White DoS, a proposition is deleted only if it both exhibits refusal-induction features and lacks factual support, a conjunctive condition designed to spare genuinely necessary risk warnings, and a forced-generation flag prevents the model from being steered onto the refusal path.

A final enhanced self-assessment scores the integrated knowledge along two dimensions, factual accuracy and safety-availability, combining them into a weighted trust score. If the score falls below a trust threshold, the framework iteratively refines the integration until it converges or reaches a maximum number of iterations, forming a closed loop of front-end purification and back-end verification. The authors are candid about the limits of the internal knowledge anchor: for temporally sensitive questions or long-tail specialized domains, the anchor itself may be wrong, and a dedicated error analysis on a knowledge-boundary subset called KB-Probe showed that a confidently incorrect anchor could drive correct external evidence below the retention threshold, occasionally emptying the evidence set entirely. Notably, an absent anchor proved preferable to an incorrect one, since the fact-checking module abstained rather than adjudicating when the anchor was empty.

The experimental results, conducted under the static, non-adaptive attack setting defined by SafeRAG, are striking. On DeepSeek-v3.2, MIRAGE achieved average improvements of 20.9 percent in Attack Failure Rate, 11.5 percent in Retrieval Accuracy, and 14.1 percent in average F1 over baselines across all four attack scenarios. Against Vanilla RAG it improved AFR by 28.7 percent, and against TrustRAG by 16.1 percent. Single-point defenses fared far worse, with the NLI filter and SKR compressor reaching AFRs of only about 0.43 compared with MIRAGE’s 0.626. Ablation studies confirmed that every component contributes: removing dynamic clustering cut AFR by 9.7 percent, removing adaptive integration rules by 9.0 percent, removing attack type prediction by 6.3 percent, and removing enhanced self-assessment by 5.8 percent. Paired t-tests over five random-seed runs confirmed statistical significance for the improvements over TrustRAG.

Generalization results add weight to the claims. Across four mainstream Chinese large language models, including DeepSeek-v3.2, Baichuan4, Qwen3, and Hunyuan-2, MIRAGE maintained strong defense, with AFR ranging from 0.558 to 0.626, though the authors stress that model-agnosticism remains a design assumption rather than an empirically validated cross-lingual result, since no English-language open-source models were evaluated. On the canonical benchmarks Natural Questions, MS-MARCO, and HotpotQA under mixed attack injection, MIRAGE kept Retrieval Accuracy above 0.56 and AFR above 0.58, outperforming RobustRAG, InstructRAG, AstuteRAG, and TrustRAG, with the largest gains on multi-hop HotpotQA where fragmented evidence exacerbates conflict defense. The framework also naturally covered traditional attacks such as noise injection, knowledge poisoning, fact tampering, and simple denial of service, improving AFR by an average of 31.0 percent over Vanilla RAG, without requiring dedicated modules for each legacy threat.

The costs and caveats are real. MIRAGE’s average per-sample inference latency on DeepSeek-v3.2 is roughly 5.6 seconds, compared with 3.8 seconds for TrustRAG and 1.2 seconds for Vanilla RAG, though the authors suggest parallelizing proposition processing or capping self-evaluation iterations for latency-sensitive services. The security analysis also acknowledges that a fully informed adaptive adversary could potentially bypass clustering detection through semantic perturbation, defeat ROUGE-L filtering through deep paraphrasing, or craft malicious propositions aligned with the model’s internal knowledge, and systematic evaluation under adaptive attacks is flagged as key future work. Still, by projecting diverse attacks onto a unified framework of embedding-space distribution modeling on the retrieval side and knowledge-distortion modeling on the generation side, MIRAGE offers the most complete answer yet to a threat that grows more urgent as RAG attacks industrialize, go black-box, and expand into multimodal and indirect injection forms. For domains such as finance, government, and healthcare, where a single poisoned passage can corrupt decisions at scale, that progress matters.

Subject of Research: Adaptive defense of retrieval-augmented generation systems against covert textual evidence manipulation attacks

Article Title: MIRAGE: manipulation-immune RAG with adaptive guard against textual evidence attacks

Article References: Yuan, Q., Li, F., Hu, H., Mu, Y., Zhou, H., Liu, Y., & Zhang, Y. (2026). MIRAGE: manipulation-immune RAG with adaptive guard against textual evidence attacks. Cybersecurity, 9(1), Article 240. https://doi.org/10.1186/s42400-026-00678-0

Image Credits: AI Generated

DOI: 10.1186/s42400-026-00678-0

Keywords: retrieval-augmented generation, large language models, RAG security, textual evidence manipulation, SafeRAG benchmark, silver noise, inter-context conflict, soft advertisement, White DoS, dynamic clustering, self-assessment, AI cybersecurity

The core difficulty the researchers set out to solve lies in a family of techniques known as Multi-Attribute Decision-Making, or MADM. These methods are designed to rank alternatives—here, scenic spots—against a set of weighted criteria. They work elegantly when the number of attributes is small and the data are homogeneous. But modern tourism evaluation data, harvested at scale from platforms such as Ctrip, violate both assumptions. Hundreds of attributes may be involved, many of them correlated, some redundant, and some outright contradictory in how they rank the same destinations. Feeding such high-dimensional data directly into a decision model can distort results, because redundant or conflicting attributes exert an outsized influence on the final score.

The team’s solution unfolds in three stages. The first stage is a consistency-driven dynamic clustering algorithm. Instead of clustering the scenic spots themselves, the method clusters the evaluation attributes, grouping together those criteria that produce similar ranking patterns across the destinations being compared. In other words, if two attributes—say, visitor ratings of cleanliness and ratings of facility maintenance—consistently order the scenic spots in nearly the same way, they are placed in the same cluster. The word dynamic is important: the algorithm adjusts cluster assignments as it processes the data rather than fixing them in advance, and it uses measures of rank consistency to decide when attributes genuinely belong together. This step reduces the dimensionality of the problem in a principled way, without discarding any information outright.

The second stage applies a well-established decision tool within each cluster: TOPSIS, the Technique for Order Preference by Similarity to Ideal Solution. TOPSIS, introduced in the operations research literature in the 1980s and refined through the 1990s, ranks alternatives by measuring how close each one is to an ideal positive solution and how far it is from a negative ideal solution. Within a single cluster of similar attributes, this is a reasonable and well-conditioned calculation, because the attributes behave coherently. Running TOPSIS separately inside each cluster yields a set of local rankings—one ordered list of scenic spots per cluster—each reflecting the perspective of a coherent family of evaluation criteria.

The third stage is where the method borrows from an unexpected field: voting theory. The researchers use the Borda count, an aggregation procedure devised in the eighteenth century by Jean-Charles de Borda, to fuse the local rankings into a single global ranking. Under the Borda count, each local ranking awards points to each scenic spot according to its position, and the spot with the greatest accumulated total across all clusters wins the top place. This is the same arithmetic that determines winners in some ranked-ballot elections, and its appeal here is the same as in politics: it aggregates many partial, imperfect orderings into one socially robust verdict while dampening the influence of any single anomalous cluster. The authors note that prior theoretical work has argued the Borda count is in an important sense the optimal ranking method, which makes it a natural choice for fusion.

The pipeline was tested on evaluation data for China’s 5A scenic spots, collected from publicly accessible tourism review information on the Ctrip platform. The 5A designation is the pinnacle of China’s tourist rating system, so the destinations involved are already elite; the question was whether a data-driven ranking built from raw, high-dimensional visitor evaluations could reproduce and refine the official assessments. The empirical results showed a high degree of consistency between the method’s output rankings and existing tourism ratings, suggesting that the clustering-and-fusion pipeline captures the same underlying quality signals that expert evaluators respond to, but derives them transparently from visitor data rather than from a fixed administrative rubric.

Why not simply cluster the destinations instead of the attributes? The authors’ design choice reflects the structure of the problem. Tourists evaluate the same scenic spot along many dimensions, and the statistical relationships among those dimensions—how they correlate, which ones move together across destinations—carry the real information. By grouping attributes with similar ranking behaviour, the method effectively discovers latent evaluation perspectives: bundles of criteria that visitors treat as a single experiential quality. Each cluster then functions as a coherent judge, and the Borda count lets these judges vote. This architecture mirrors recent work on large-scale group decision-making, in which clustering and consensus techniques are used to manage hundreds or thousands of participants, and it connects to a growing literature on ranking fusion for massive online review datasets.

The technical payoff is robustness against heterogeneity. Real-world evaluation data are rarely generated by a single consistent process: different visitors weight different concerns, different attributes suffer different amounts of noise, and some criteria may be systematically inflated or deflated on particular platforms. A monolithic MADM model must average all of these effects together, which can wash out genuine distinctions between destinations. The cluster-then-rank-then-fuse design isolates each coherent evaluation channel, ranks destinations within that channel, and only then combines the verdicts. Conflicting attributes are no longer forced into a single weighted sum; they are allowed to speak in their own cluster and then reconciled at the voting stage. The authors position the framework as particularly valuable for scenarios involving high-dimensional and heterogeneous evaluation data, where conventional approaches struggle.

The practical implications extend beyond holiday planning. Tourism management authorities could use such a framework to monitor how the data-driven ranking of attractions evolves as visitor sentiment shifts, identifying destinations whose reputations are improving or deteriorating and tracing those changes back to specific clusters of attributes. Travel platforms could integrate the method into recommendation engines that must reconcile millions of reviews across many criteria. And the underlying algorithmic idea—cluster the criteria, rank within clusters, fuse by voting—is domain-agnostic, potentially transferable to any multi-attribute decision problem drowning in high-dimensional data, from product selection to service quality assessment.

The study also offers a candid account of its data foundations. The datasets were drawn from publicly accessible tourism evaluation information on Ctrip, and the full raw dataset cannot be redistributed because of third-party platform terms of use and intellectual-property restrictions. To support reproducibility, the published paper documents the data source, structure, preprocessing, variable construction, clustering procedure, parameter settings and experimental design in detail, with derived statistical information and non-restricted processed data available from the corresponding author on reasonable request. The research was supported by the Hunan Provincial Natural Science Foundation of China and the Scientific Research Fund of the Hunan Provincial Education Department. As online reviews continue to swell into one of the largest bodies of human judgement ever collected, methods like this one—part statistics, part social choice theory—point toward a future in which the wisdom of crowds is not merely accumulated, but properly counted.

Subject of Research: A dynamic clustering and ranking fusion method for multi-attribute evaluation of 5A scenic spots using tourist review data

Article Title: Dynamic clustering and ranking fusion for tourism recommendations: A data-driven approach to 5A scenic spot evaluation

Article References: Li, W., Xiao, Z., Yu, B., & Xu, Z. (2026). Dynamic clustering and ranking fusion for tourism recommendations: A data-driven approach to 5A scenic spot evaluation. Applied Intelligence, 56(15), Article 433. https://doi.org/10.1007/s10489-026-07488-4

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07488-4

Keywords: tourism recommendation, dynamic clustering, ranking fusion, TOPSIS, Borda count, multi-attribute decision-making, high-dimensional data, 5A scenic spots, online reviews, Ctrip, consistency-driven clustering, Applied Intelligence