Context Compression

Knowledge distillation (KD) is a widely utilized technique for transferring knowledge from a large model (the teacher) to a smaller model (the student). Owing to its flexibility and broad applicability, KD has been extensively applied in the compression of server-side models to meet the Quality of Service (QoS) requirements of client users. Despite significant advancements, the performance of distillation is substantially compromised when a large disparity exists between the capabilities of the server and the requirements of the client. To alleviate this problem, we propose a novel distillation approach, named Progressive$^2$, which operates through the combination of a progressively stronger teacher and a progressively smaller student. On the side of the teacher, rather than involving all layers simultaneously, we progressively select additional layers for distillation following a raw-to-rich semantic progression, establishing a systematic learning curriculum. Furthermore, we design a teacher-side multi-feature fusion adapter for the teacher to improve training stability, which is theoretically supported by the framework of Lipschitz continuity. On the side of the student, rather than directly training a tiny model, we gradually reduce the size of the network to facilitate an iterative co-evolution with the teacher. Progressive$^2$ serves as a flexible framework; the progressive strategy of the teacher can be deployed independently to achieve an optimal balance between accuracy and training efficiency, while the joint integration of the teacher and the student yields further improvements in overall performance.

Practitioners accept a compressed language model once it clears a stack of data-cheap quality guards: perplexity within a small factor of the original, downstream accuracy (for example MMLU) inside a confidence interval, and data-free output-fidelity signals that compare the compressed and original network's internal representations under random probe inputs. This stack has a blind spot. Across three model families, gently-compressed models clear every guard and then invent procedure steps that were never in the instructions when they run a standard operating procedure (SOP) as an agent. The effect is operator-specific: coherent low-rank (SVD) truncation induces it, and magnitude pruning matched to the same perplexity does not. One dissociation isolates the cause. The same compressed weights that CI-win a paired output-fidelity test CI-fail the invented-step canary. The governing axis is the coherence of the compression error times its rate; the magnitude of the damage does not predict it. The data-free fidelity probe is a fidelity oracle by construction, so it cannot see this axis. We characterize the blindspot and dissociation with paired confidence intervals on a pre-registered, powered canary across three architectures. Operator-specificity replicates on all three, and the perplexity-guard evasion appears where the model admits in-guard low-rank headroom. We then give a data-free screen: a two-axis statistic of the compression error (coherent-fraction and error-rate) that flags the failing builds with fixed thresholds across architectures and matches the coherence-times-rate mechanism. Perplexity, MMLU, and fidelity acceptance do not certify agent safety. Screen gently-compressed low-rank builds before agentic deployment

Context compression discards most of a document before a language model reads it, and is normally evaluated by downstream task accuracy - which makes another model the judge of what mattered. Naturalistic social highlighting offers a non-circular reference: many people independently marking passages on the same page. But the obvious metric, the fraction of crowd-marked sentences a compressor keeps, is confounded twice: crowd marks are front-loaded and crowd-marked sentences are longer, so any method favouring early or long sentences scores well regardless of readers. We remove both by matching each marked sentence against unmarked sentences of the same document at equal relative depth and equal within-document length rank, and we calibrate every estimator on synthetic nulls built from position and length alone - a step that matters, since depth-only stratification returns a false positive on 20-36% of nulls containing no effect. On 120 web documents (at least 12 independent readers each), a language-model importance ranking keeps 38.4% of crowd-marked sentences against 19.9% of their matched neighbours: an enrichment of +0.196 [+0.148, +0.239], at p = 0.0005 under an exact randomization test that assumes nothing about clustering, and replicated cross-vendor. Naive truncation, whose keep rule is position, correctly falls to +0.003. To give the number a scale: scored identically, on the same budget, against a crowd label recomputed to exclude them, a single human reader reaches +0.182 - indistinguishable from GPT-5.4 (+0.002 [-0.081, +0.088]) and below Claude Opus 5. Classical methods are not null - Luhn's 1958 heuristic reaches +0.088 - so reader selection is partly recoverable by counting words; conditioning additionally on lexical centrality removes only 0.010, so the agreement is not centrality. We also report that a claim in our own prior work does not reproduce on this corpus.

Prompt compression shortens LLM input to reduce inference cost, yet existing methods score token importance through LM forward passes. It remains questionable whether such nuanced, costly token selection is necessary. Compression requires identifying informative content, a problem that linguistic research has long addressed through cues that can be operationalized as deterministic rules. We therefore ask: can \textbf{linguistic rules alone} serve as effective prompt compressors, without LM-based scoring at compression time? To address this, we conduct offline evolutionary search over lexical, syntactic, semantic, and discourse seeds to find competitive rule combinations. The resulting linguistic compressor requires no LM forward pass at deployment and uses only CPU-side processing for compression. We evaluate it with a dual-path protocol to balance compression quality and reconstruction fidelity. Across short passages, multi-document reasoning, and dialogue-memory QA datasets, evolved compressors achieve performance similar to that of recent advanced prompt-compression strategies. Performance is strongest under light-to-moderate compression and degrades as compression becomes more aggressive, while the Direct and Reconstruction paths exhibit distinct patterns. Evolutionary analysis reveals that effective compression fuses signals across linguistic levels and, as the compression ratio increases, rules shift from token pruning to sentence extraction.

We study the problem of lossless compression of source code, motivated by the storage demands of large-scale software archives, such as Software Heritage (https://www.softwareheritage.org/). General-purpose compressors (e.g., zstd, bzip2) offer a good trade-off between compression ratio and speed, but fail to exploit all special regularities inherent in source code. Recent approaches leverage Large Language Models (LLMs) within Shannon's symbol-ranking framework, relying on a scheme in which the predicted rank can grow arbitrarily. While effective at reducing space, this setting incurs significant throughput degradation, and leaves open the question whether it is necessary to explicitly encode all ranks. In this work, we introduce LLM-based compressors deploying two novel symbol-ranking variants that bound predictions to the top-$T$ ranks ($T=1$ or $63$), with out-of-threshold symbols stored as exceptions and compressed jointly with the rank stream via general-purpose compressors. We conduct the first large-scale evaluation of LLM-based source code compression across 30 LLMs, including general-domain, code-specialized, and quantized models. Our $T$-bounded approach outperforms prior LLM-based compressors both in compression ratio (up to 37% relative improvement) and compression throughput (40% faster). Compared to general-purpose compressors (e.g., zstd, bzip2), we obtain up to 82% relative compression gain but at a lower speed, thus offering a new trade-off point in the compression-speed spectrum. We also show that these gains are stronger on source code than on natural language, suggesting an interesting indication, namely that source code exposes regularities captured by LLMs but missed by general-purpose exact-match-based compressors. We conclude by commenting on open problems that offer theoretical and practical avenues of research.

Graph compression reduces the computational cost of graph learning, but its effect on signal propagation remains largely underexplored. Existing work evaluates compression through downstream task performance or structural preservation, neither of which directly captures how propagation dynamics change after compression. We study two fundamental compression paradigms, coarsening and sparsification, and ask whether they preserve the propagation behavior of the original graph. Across five datasets, varying compression rates, and propagation depths, we measure signal behavior through three complementary metrics. Our results reveal a consistent tension between the two compression families. Sparsification retains higher signal diversity and mitigates oversmoothing, but its propagation trajectory progressively diverges from that of the original graph. Coarsening more faithfully preserves propagation behavior, but at the cost of stronger smoothing and rank collapse. These findings demonstrate that two propagation-centric objectives, preserving signal diversity and preserving propagation fidelity, are distinct and empirically at odds under graph compression, highlighting the need for evaluation protocols that jointly consider both dimensions. The code and results are available at: https://github.com/KawshikBanerjee/Compression-Propagation-Duality

Data-adaptive sparse attention masks substantially outperform fixed patterns (e.g., BigBird and Longformer) and can even exceed dense attention on long sequences. Existing adaptive approaches---including SBM-Transformer, Dynamic Mask Attention, and NSA---typically require additional learnable parameters, custom gradient estimators, or specialized CUDA kernels. We show that classical data compression provides an effective masking signal with \textbf{no additional parameters}. By computing per-block gzip compression ratios, we identify non-redundant content blocks and route long-range attention selectively through them. Intuitively, blocks that gzip cannot compress contain information not predictable from local repetition, making them natural long-range attention targets. Because the compression profile is input-dependent, the resulting sparse mask adapts dynamically to content without learned parameters, auxiliary losses, or custom kernels. On PG-19 byte-level language modeling at 92M parameters with 8K context, our method achieves 1.71 bits-per-byte (BPB), outperforming dense attention (2.89), BigBird (2.34), Longformer (3.21), and a reimplemented SBM-Transformer (3.38)---the only learned-mask baseline---by up to 1.67 BPB while adding no parameters. The advantage grows with sequence length, with the gap over BigBird widening from 0.05 BPB at 4K context to 0.63 BPB at 8K, while convergence is 3.3$\times$ faster.

Token cramming compresses sequences into learned embeddings with near-perfect reconstruction, but fixed token budgets and 99% accuracy thresholds leave it unclear whether residual errors reflect optimization failures or fundamental limits. We introduce progressive cramming, which grows the target prefix token-by-token, stopping only when reconstruction is no longer achievable within a fixed optimization budget. Progressive trajectories occupy low-dimensional structure in embedding space. Prepending a crammed embedding causes a moderate but consistent accuracy drop on multiple-choice benchmarks even with the original prefix in context, and collapses capability almost entirely under generative evaluation. Causal attention-knockout interventions trace this degradation to the embedding's interactions in the model's early layers. These results position progressive cramming as a tool for studying compression limits and show that perfect reconstruction - achievable through brittle steering rather than transferable semantics - is insufficient for meaningful compression.

Spoken language models (SLMs) enable natural human-computer interaction, but their reasoning ability still lags behind that of text-based large language models, especially on spoken mathematical question answering tasks. One important reason is that SLMs reason over purely verbalized mathematical expressions, which are harder to interpret than symbolic text. However, directly transferring text-based reasoning to SLMs is nontrivial due to architectural constraints and the additional computational requirements. To address this challenge, we propose Efficient Chain-of-Modality Reasoning (ECoM Reasoning), the first framework to introduce compressed reasoning into SLMs. By compressing the textual component so that it jointly serves as speech guidance and reasoning representation, ECoM Reasoning improves reasoning accuracy while using a smaller token budget than the standard Chain-of-Modality (CoM) architecture, which generates intermediate text before speech. To train this capability, we further propose Progressive Compression, a curriculum-based strategy that gradually trains the model from full-form reasoning to compressed reasoning. Experiments on spoken mathematical question answering benchmarks show that ECoM Reasoning improves accuracy by 21% over standard CoM without explicit reasoning, and by 3% over CoM with full reasoning traces while using only 40% of the text tokens, demonstrating that it enhances SLM reasoning while remaining inference-efficient.

Large language model (LLM) agents are vulnerable to security risks, such as prompt injection attacks from untrusted context that manipulate downstream reasoning and tool use. Existing secure-by-design approaches mitigate this risk by separating untrusted observations from privileged execution and careful control of information flow, but often degrade utility and require extensive task-specific engineering. We thus propose Twin Agent, a general privilege separation design pattern inspired by residual coding in the agent context. Twin Agent consists of two nearly symmetric agents: an Explore Agent that inspects untrusted information and a Safe Agent that executes privileged actions. The Explore Agent is conditioned on the Safe Agent's current context and communicates only compact hints to the Safe Agent about the next action to take. This design reduces the information needed to preserve task utility and thus achieves a better security--utility tradeoff, which we empirically verify by measuring how utility and attack success change as the length of hints varies. We evaluate Twin Agent on long-horizon software engineering tasks with SWE-bench Lite and on heterogeneous multi-tool interaction tasks with AgentDojo and DecodingTrust-Agent. Across both benchmarks, Twin Agent preserves high task utility while preventing prompt injection attacks, outperforming both undefended agents and privilege separation baselines.

Continuous surveillance video creates a growing storage, transmission, and inference burden for enterprise video analytics systems. While modern codecs such as H.265 reduce bitrate for human-viewable video, aggressive compression can degrade downstream computer-vision performance and does not necessarily reduce the number of vision-language model (VLM) inference calls required for semantic video understanding. This paper evaluates BLUE, a fixed-camera surveillance compression approach that suppresses static-background redundancy while preserving foreground activity, for its effect on VLM-based event and anomaly understanding. We compare raw H.265 and BLUE-compressed H.265 video on two surveillance datasets: VIRAT, comprising 227 paired event samples from 106 clips, and CHAD, comprising 54 human-activity anomaly clips. For each pair, the same frame index is evaluated using a VLM captioning pipeline, and outputs are scored against annotation-derived ground truth using a blind judging protocol. The results show no measurable degradation in semantic inference quality. On VIRAT, the mean VLM score remains effectively unchanged between raw H.265 and BLUE, with a mean difference of approximately -0.01 on a 0-10 scale. On CHAD, raw H.265 and BLUE obtain near-equivalent mean scores of 4.31 and 4.26, respectively. Compression saving is also uncorrelated with VLM score change on VIRAT (r = 0.004), indicating that higher BLUE compression does not predict semantic quality loss. Beyond storage reduction, BLUE increases the share of skip-heavy P-frames on CHAD from 1.4% to 53.2%, enabling an estimated 53% reduction in VLM calls through packet-size-based frame skipping. These findings suggest that BLUE functions as a machine-centric compression layer for surveillance video, reducing bandwidth and inference cost while preserving VLM semantic performance.

Detectors for AI-generated video are evaluated offline. A clip is decoded to pixels and scored once, increasingly by a large vision-language model. Detection, however, is deployed online. We recast the task as streaming perception and score the motion field the codec already wrote into the bitstream. Reading that field is a parse, not a pixel-domain forward pass. Because the running aggregate is monotone, one end-calibrated threshold is anytime-valid at the data-dependent decision time. Recalibrating at each prefix is not. Escalation is priced in closed form. A compute budget maps to a deferral window, on a frontier monotone exactly where the deferral condition holds. On matched GenVidBench the codec stage reaches full-length AUC 0.64 at five orders of magnitude less compute than a pixel CNN, on CPU. Its gate holds the stopping-time false-positive rate at target while the real data match its calibration, and drifts above it under distribution shift. Deferring 15% of clips lifts accuracy from 0.75 to 0.78 at $7\times$ less compute (paired: McNemar $p<10^{-6}$). The stage-1 ordering replicates on AIGVDBench. We introduce no new detector. The contribution is the reframing, two guarantees, and the measured frontiers. Code, configurations, and evaluation splits: https://github.com/KurbanIntelligenceLab/streamdet.

Recent advances in differentiable Gaussian splatting have highlighted the potential of primitive-based approaches as alternative scene representations for interactive, high-quality, volume visualization (VolVis) of large datasets. However, the explicit nature of current primitive-based methods, combined with isolated optimization for each VolVis scene, results in redundant, non-compact representations. We present ECoNGS, an efficient compressive neural Gaussian splatting framework for VolVis scene representation. ECoNGS employs lightweight neural networks to dynamically predict implicit, editable Gaussian splats from explicit anchor points, effectively combining model compactness and parameter efficiency of implicit representations with high-performance rendering of explicit primitives. We explore a joint learning strategy that clusters geometrically similar scenes and shares parameters across them, significantly reducing overall training time and model size while maintaining reconstruction fidelity. To achieve a more compact scene representation, we further compress the explicit anchor attributes using a neural entropy model that estimates their probability distributions, enabling compact storage via entropy coding. We systematically investigate Gaussian initialization strategies and propose a simple yet effective scheme tailored for VolVis scenes, improving reconstruction accuracy and accelerating convergence. We evaluate ECoNGS qualitatively and quantitatively across various univariate and multivariate VolVis scenes, highlighting its superior performance over prior methods in training time, reconstruction quality, and model size. In particular, compared with the prior method iVR-GS, ECoNGS improves reconstruction quality by up to 2.2 dB in PSNR while reducing the model size by up to 6.1x and the training time by up to 5.9x. The code is available at https://github.com/TouKaienn/ECoNGS.

Large-scale scientific simulations generate volumetric data at rates that far outpace advances in storage and network bandwidth, making effective lossy compression increasingly critical. However, conventional compressors often struggle to preserve fine structural details at high compression ratios (CRs), and implicit neural representations (INRs) require costly per-volume optimization and produce models with fixed CRs. To respond, we present EVOLVE, an autoencoder (AE)-based volume-compression framework that targets high CRs for offline compression, with three key contributions. First, we construct a large-scale cross-domain database of 6,376 volumes from 21 scientific simulations, curated via perceptual hashing to ensure diversity, enabling the optimized model to extract features that generalize across volumes within the covered scientific simulation domains. Second, we reexamine the design space of AE-based compressors and incorporate several macro- and micro-designs into a vanilla AE to develop EVOLVE, which substantially improves the expressive power and compression capability. Third, we develop a learnable gain mechanism with a three-stage training strategy to enable variable-rate encoding, allowing a single model to support continuous CR adjustment at inference time. Experiments on multiple unseen scientific simulation datasets demonstrate that EVOLVE achieves substantially higher CRs than conventional compressors at comparable reconstruction quality, while delivering compression speeds that are orders of magnitude faster than INR-based methods, highlighting its promise as a strong alternative for compressing scientific data. The code, model weights, and results are available on our project page at https://evolve-vis.github.io.

As large language models (LLMs) process increasingly longer prompts, computation and KV-cache memory costs have emerged as major bottlenecks in inference systems. Existing input-level prompt compression methods address this, but rank each sentence by a scalar relevance score, treating the document as an unstructured pool of words and sentences. Under tight budgets, this causes theme collapse, where the dominant theme(s) of a document consumes the budget, discarding less-frequent yet task-relevant themes. Preserving thematic coverage instead requires allocating the budget across recurring themes rather than scoring sentences in isolation. To this end, we propose SALT, a model-agnostic extractive framework that organizes per-sentence keywords into a trie ordered by sentence frequency (SF), a lightweight, reusable proxy for document thematic structure. This trie-based organization smooths memory allocation and prevents dominant themes from monopolizing the budget. Multi-anchor retrieval activates trie nodes labeled by query keywords at any depth, and the trie persists across dialogue turns, supporting multi-turn use without re-encoding the document. By preserving document themes, SALT reduces the prefill computation and memory cost of long-context prompts while remaining composable with KV-cache methods that target decoding-time latency and memory.

Sparse autoencoders (SAEs) compress model activations into sparse codes, but equal reconstruction error and sparsity can preserve different linearly decodable signals. We formalize this ambiguity as a matrix-valued distortion between optimal ridge-prediction operators and train decoder-preserving SAEs by combining this distortion with reconstruction loss. In a rank relaxation, an isotropic task prior saturates per-mode omission costs without changing PCA's ordering, whereas a structured prior can change which modes are retained. A controlled sparse experiment shows that a declared prior protects held-out combinations from its task subspace. On GPT-2 small block 8, DPSAE reduces held-out decoder distortion by 10.6--11.4% across three paired runs while matching reconstruction NMSE. The same checkpoints pass an average natural-text output-KL noninferiority test, but one matched Pythia pair shows no improvement in probes restricted to a few sparse features. These results show that reconstruction quality does not determine which refitted linear readouts survive sparse compression, and that readout preservation is distinct from learning cleaner benchmark concepts or preserving every frozen-model behavior.

Production LLM deployments combine two cost-reduction primitives: prompt caching (a discounted rate for re-used token prefixes) and prompt compression (fewer tokens sent). The compression literature has standardized on query-aware methods that produce a different compressed prefix per query, mechanically invalidating the prefix-strict cache on every call. We characterize this cost empirically on Anthropic's Sonnet 4.6 API and find caching is far from the rho=1.0 ideal the literature assumes: Sonnet's cache has a two-tier architecture with a sharp threshold near 3,500 tokens, below which the hit rate plateaus at rho~0.83 across 30-call sessions. Our cost model predicts, and experiments confirm, that under realistic rho, query-aware compression beats naive caching at high compression ratios (r>=6). We propose Cache-Aware Prompt Compression (CAPC), pairing query-agnostic compression with explicit cache_control plus a tier-preserving ratio bound that prevents over-compression from pushing the cached prefix into the hot tier. CAPC is the cheapest strategy in 16/16 configurations on LongBench-v2, with mean savings of 49% over cache-only, 64% over query-aware compression, and 90% over vanilla, at quality within 0.05 of the uncompressed baseline. We validate CAPC on three production workloads: an enterprise tool-using assistant with a 94k-token schema prefix (51.7% cost reduction at r=3); a graphify knowledge-graph RAG pipeline across two codebases (9.3x vs cache-all on FastAPI, 2.4x on httpx); and the public tau-bench retail benchmark (50 tasks), where CAPC is the cheapest of four strategies with reward exactly equal to vanilla (both 36/50, p=1.00) while query-aware compression is the most expensive at +40.1% over vanilla -- the first production confirmation of the crossover model's negative-ROI prediction on a public benchmark.

Learned image compression (LIC) is bottlenecked by the need to store independent models for each rate-distortion operating point. Existing variable bit-rate (VBR) methods aim to reduce this overhead via dense parameter modulation, but forcing a shared backbone to approximate divergent mappings causes severe feature entanglement. Specifically, low-rate smoothing gradients inherently conflict with the preservation of high-frequency textural details, leading to sub-optimal performance. To resolve this, we propose MixCompress, a unified VBR framework based on sparse structural specialization. While sparsely gated Mixture-of-Experts (MoE) routing successfully mitigates gradient conflict, it operates on a fixed computational budget. To address the increased representational demands of higher bit-rates we introduce a Mixture-of-Depths (MoD) extension to dynamically scale model capacity. Combined with Conditional Auxiliary Transforms (CAT) for dynamic sub-band energy modulation, our hierarchical framework effectively dynamically scales capacity. Extensive evaluations demonstrate that MixCompress not only matches individually optimized single-rate baselines but can even surpass them, establishing a new Pareto frontier for computationally efficient image coding.

Efficient long-context inference is not only about reducing memory cost, but also about keeping useful contextual evidence accessible as generation proceeds. However, existing compression-oriented approaches, such as key-value (KV) cache compression and context compression, often either make an early decision about which contextual information to keep or rely on an external compressor. Such designs make it difficult to adapt the compressed context to the evidence needed by later reasoning steps. This paper introduces PReM (Preserve and Refresh Memory), a context-compression framework that maintains the long context as the model's internal layer-wise KV memory and learns what to preserve and when to refresh it. Specifically, PReM uses a dedicated memory layer to make memory-selection decisions, and a special memory token <m> to trigger refreshes during generation. To train this behavior, PReM introduces Phase-Separated Refresh Training, aligning memory selection with memory-conditioned generation while preserving continuity across refreshes. Experiments with 32K-token contexts show that PReM outperforms strong baselines under both 16x and 32x compression, while maintaining a favorable balance between answer quality and inference efficiency.

The rapid deployment of machine learning systems across cloud, edge, and enterprise environments has brought model optimization to the forefront of systems-engineering. Despite a rich literature spanning quantization, pruning, knowledge distillation, parameter-efficient fine-tuning (PEFT), and inference-time optimization, practitioners are often left navigating these techniques through heuristics rather than principled methodology. We argue that optimization should be formulated as a constraint-driven, multi-objective engineering decision and introduce a unified framework that characterizes any production deployment along five interacting constraint dimensions: data availability, latency budget, memory budget, accuracy tolerance, and retraining budget. Building on this taxonomy, we synthesize empirical gains reported across the research literature and map them to operational constraints rather than algorithmic categories. To ensure practical relevance, we selected these techniques by reviewing recent literature for methods that report measurable improvements against critical deployment bottlenecks. We propose a prescriptive decision framework and provide optimization pipelines for four representative industrial scenarios to illustrate it in practice. To the best of our knowledge, this work provides one of the first structured attempts to formalize model optimization as a constraint-aware, multi-objective engineering process, synthesizing quantitative evidence from the research literature.

This paper proposes a theoretical framework for understanding intelligence as a process of atomic compression and compositional reuse. We argue that cognitive, biological, computational, and organizational systems achieve scalable intelligence by decomposing complex phenomena into reusable atomic units that can be recombined into higher-order structures. Drawing on evidence from cognitive science, information theory, evolutionary biology, software engineering, medicine, legal reasoning, education, music, and artificial intelligence, the paper develops the concept of atomic units as fundamental compression layers that support efficiency, transfer, interpretability, and evolvability. The central contribution is the Compression Calculus, a formal framework for comparing surface-level representations with atomic representations and for describing how compression gains compound across abstraction layers. We introduce the Compounding Cascade thesis, according to which each additional layer of abstraction multiplicatively increases representational efficiency rather than merely adding incremental savings. The paper further argues that contemporary AI systems often operate at suboptimal levels of representation, relying on token-level processing or document-level retrieval rather than stable, concept-level atomic structures. In this view, large language models are best understood not as complete knowledge architectures, but as dynamic fusion engines capable of navigating, sequencing, and recombining atomic units. The framework provides a foundation for designing self-evolving knowledge systems that can discover, refine, and compose new primitives over time. By reframing intelligence as compression through compositional abstraction, the paper offers a unifying perspective on expertise, knowledge representation, explainable AI, and the future architecture of adaptive intelligent systems.

Context-reduction layers for API-based coding agents, including command-output compressors, retrieval rankers, and API-boundary proxies, are commonly evaluated by how much context or tool output they remove. We ask a different question: which interventions actually reduce end-to-end billed cost while preserving task success? Our primary evidence is a pre-specified, hash-frozen, paired campaign of 2,908 provider-billed Claude Code runs, of which 2,848 were analyzed, covering 103 tasks, seven repositories, and three models. The campaign compared a baseline with two generations of hook-based compression and an API-boundary proxy within a broader measured program of roughly 5,500 billed executions. Three findings emerge. First, prompt-cache traffic dominated cost composition, accounting for about 87% of reconstructed four-component cost (about 80% of the actual bill), with an 8.7% dollar-weighted residual not attributable from retained telemetry. Second, local payload reduction was not a reliable predictor of end-to-end billed cost. An arm that removed 38% of estimated raw tool-output tokens incurred 6.8% higher paired cost (95% CI: +2.8% to +11.3%), while per-task reduction showed only a weak association with cost change (Pearson r = 0.15). Third, aggressive compression can remove action-critical evidence: on SWE-bench-derived Go tasks, compression reduced successful patch application from 27/40 to 15/40 by corrupting verbatim edit anchors. We propose evaluating context-reduction systems by success-adjusted billed cost rather than token reduction alone.

With deep neural networks (DNNs) increasingly deployed on edge devices, hardware (HW)-aware optimization techniques--such as HW-aware compression and HW-aware neural architecture search (HW-NAS)--have become essential. These methods rely on real feedback from the target hardware to tailor DNN architectures for efficient deployment. While the search can be parallelized, latency measurements via hardware-in-the-loop (HIL) remain a bottleneck due to their sequential nature. Recent approaches use latency predictors to replace costly HIL feedback, but challenges persist: (1) platform-specific predictors often require tens of thousands of samples, and (2) inaccurate predictions can mislead the NAS process. To address this, we introduce HiFi-LLP, a high-fidelity, low-cost latency predictor based on graph attention networks, augmented with a confidence metric. HiFi-LLP outperforms prior platform-specific predictors by up to 9 percentage points (p.p.) in the 10% accuracy bound and achieves a Spearman's rank correlation of up to 0.996 across six devices in the LatBench dataset. We further propose a hybrid NAS framework that routes low-confidence predictions to HIL, achieving up to 8.6$\times$ speedup compared to typical NAS while maintaining a competitive Pareto front.

In open-domain multi-hop question answering (QA), LLM-based search agents offer a promising approach to knowledge-intensive QA by combining retrieval with reasoning. Existing methods mainly improve open-domain multi-hop QA through reasoning paradigms, retrieval interaction, and search strategy optimization. However, using multiple search trajectories introduces a challenging final answer selection problem. Different trajectories may support different candidates, and the retrieved information can be heterogeneous, redundant, incomplete, or conflicting. Directly comparing raw trajectories exposes the verifier to noisy and unaligned content, while comparing answer strings ignores the evidence supporting each candidate, making reliable final selection difficult. To address this challenge, we propose STEC, an evidence compression framework for final answer selection in multi-hop QA. STEC selects the final answer from the existing candidate set through two mechanisms: (1) Answer-Level Evidence Compression, which groups trajectories by normalized answer identity and converts each answer group into a candidate-specific evidence representation; and (2) Evidence-Guided Answer Verification, which compares these representations and selects the final answer from the candidate set. The design shifts final selection from raw trajectory comparison to candidate-level evidence comparison. We evaluate STEC on four open-domain multi-hop QA benchmarks against representative baselines. Experimental results show that STEC performs best overall among the compared methods, and ablation results provide evidence that answer-level evidence compression contributes to final answer selection.

Worldbuilding, the construction of coherent fictional worlds, is a foundational task in game design and literary creation. Large Language Models (LLMs) offer new possibilities for automated content generation, but their application to worldbuilding faces three challenges: context explosion that grows linearly with the building process, the tension between creative diversity and content consistency, and the absence of automated quality assurance. This paper presents AutoWorldBuilder, a multi-agent collaborative system that addresses these challenges through five integrated components: a structured concept network with conflict detection; a DAG-based hybrid batch scheduler that groups tasks by semantic locality; a four-layer context compression mechanism achieving approximately 90% token reduction; an iterative review system with specialized Auditor agents that improves proposal pass rates from 42% to over 85%; and a skill-driven agent architecture supporting zero-code extension with differentiated temperature configuration. Two experiments across 20 diverse worldbuilding tasks, using GPT-OSS 120B and DeepSeek v3.2 as LLM backends, demonstrate a 95.0% success rate. The system generated 56-103 self-consistent concepts per world in 18-31 minutes with zero-conflict delivery. The architectural patterns validated here, including layer-as-budget compression, semantic-locality scheduling, and separation of generation and review, transfer to the broader class of knowledge-intensive, multi-agent LLM applications.

While large-scale text-to-image generative models have achieved unprecedented visual performance, their inherent reliance on multi-step iterative solvers incurs severe inference latency. Few-step distillation targeting the Classifier-Free Guidance (CFG) trajectory has emerged as the prevalent dual-dimensional compression paradigm. However, existing frameworks remain subjugated by a coarse-grained blind injection paradigm that perpetually enforces a globally static guidance strength while indiscriminately sampling the supervisor timestep. This state-agnostic design completely disregards the intrinsic nature of image generation as a dynamic evolutionary process characterized by progressive entropy reduction, which not only restricts the performance boundary of few-step compression but also precipitates severe CFG over-conditioning artifacts. To transcend these limitations, we re-examine the distillation procedure through the theoretical lens of Information Theory, formally modeling it as a dynamic mutual information game constrained by the Information Bottleneck (IB) principle. Specifically, we dismantle traditional blind assumptions via a dual-track adaptive framework. To determine the injection target, we propose an instance-aware selection mechanism that transmutes the intractable KL divergence constraint into a zero-overhead closed-form solution predicated on the local vector field norm. To regulate the injection strength, we introduce an entropy-aware schedule that dynamically decays alongside the SNR, applying maximal thrust for initial structural anchoring before smoothly reverting to the natural manifold to refine micro-details. Extensive empirical evaluations corroborate that our framework fundamentally eradicates over-conditioning artifacts, shattering the performance ceiling to achieve SOTA generative fidelity under extremely stringent 2-step configurations.

Real-image diffusion inversion is governed by a tight quality-cost trade-off, with costs incurred in computation, storage, or per-image optimization. We study this trade-off through the forward Gaussian noise anchor that defines a diffusion trajectory and isolate two mechanisms behind effective stored-noise inversion. First, diffusion noise exhibits an element-wise compression asymmetry: int8 full-dimensional anchors preserve reconstruction, whereas low-dimensional subspace summaries are much less reliable, often collapsing even at comparable or smaller payloads; the element-wise over subspace ordering persists across five stored-noise inversion methods. Second, inversion is trajectory-bound and score-prior coupled: the matched forward anchor and a trained score network are both necessary, arguing against a purely algebraic-identity explanation. Together, these findings specify what to store and how to use it. They lead to Noise-Anchored Reverse Correction (NARC), a training-free inversion primitive that stores a single int8 latent anchor and reuses it with a fixed, noise-level-dependent anchor-weight schedule: strong anchoring when the reverse trajectory is noise-dominated, then relaxed anchoring as image detail emerges. On PIE-Bench++ with Stable Diffusion 1.5, NARC outperforms five modern non-exact baselines and improves PSNR by +3.24 dB over PnP DirectInv while using about 400x less inversion storage than PnP DirectInv. The compression asymmetry, anchor specificity, and editing plug-in also transfer to SDXL 1024^2.

Convolutional neural networks (CNNs) have demonstrated encouraging results in image classification tasks. However, the prohibitive computational cost of CNNs hinders the deployment of CNNs onto resource-constrained embedded devices. To address this issue, we propose EdgeCompress, a comprehensive compression framework to reduce the computational overhead of CNNs. In EdgeCompress, we first introduce dynamic image cropping (DIC), where we design a lightweight foreground predictor to accurately crop the most informative foreground object of input images for inference, which avoids redundant computation on background regions. Subsequently, we present compound shrinking (CS) to collaboratively compress the three dimensions (depth, width, and resolution) of CNNs according to their contribution to accuracy and model computation. DIC and CS together constitute a multidimensional CNN compression framework, which is able to comprehensively reduce the computational redundancy in both input images and neural network architectures, thereby improving the inference efficiency of CNNs. Further, we present a dynamic inference framework to efficiently process input images with different recognition difficulties, where we cascade multiple models with different complexities from our compression framework and dynamically adopt different models for different input images, which further compresses the computational redundancy and improves the inference efficiency of CNNs, facilitating the deployment of advanced CNNs onto embedded hardware. Experiments on ImageNet-1K demonstrate that EdgeCompress reduces the computation of ResNet-50 by 48.8% while improving the top-1 accuracy by 0.8%. Meanwhile, we improve the accuracy by 4.1% with similar computation compared to HRank, the state-of-the-art compression framework. The source code and models are available at https://github.com/ntuliuteam/edge-compress

Scaling down the resolution of input images can greatly reduce the computational overhead of convolutional neural networks (CNNs), which is promising for edge AI. However, as an image usually contains much spatial redundancy, e.g., background pixels, directly shrinking the whole image will lose important features of the foreground object and lead to severe accuracy degradation. In this paper, we propose a dynamic image cropping framework to reduce the spatial redundancy by accurately cropping the foreground object from images. To achieve the instance-aware fine cropping, we introduce a lightweight foreground predictor to efficiently localize and crop the foreground of an image. The finely cropped images can be correctly recognized even at a small resolution. Meanwhile, computational redundancy also exists in CNN architectures. To pursue higher execution efficiency on resource-constrained embedded devices, we also propose a compound shrinking strategy to coordinately compress the three dimensions (depth, width, resolution) of CNNs. Eventually, we seamlessly combine the proposed dynamic image cropping and compound shrinking into a unified compression framework, Smart Scissor, which is expected to significantly reduce the computational overhead of CNNs while still maintaining high accuracy. Experiments on ImageNet-1K demonstrate that our method reduces the computational cost of ResNet50 by 41.5% while improving the top-1 accuracy by 0.3%. Moreover, compared to HRank, the state-of-the-art CNN compression framework, our method achieves 4.1% higher top-1 accuracy at the same computational cost. The codes and data are available at https://github.com/ntuliuteam/smart-scissor

Speech large language models (Speech LLMs) typically encode speech into sequences far longer than text, creating a major efficiency bottleneck during autoregressive decoding. A common remedy is to compress the speech sequence at the adapter level to remove temporal redundancy before it enters the LLM; however, such early downsampling risks discarding fine-grained information that cannot be recovered. We propose SpeechKV, which applies a learned pooling to the KV cache of speech tokens inside the LLM. This design allows the LLM to fuse speech and text internally while directly accelerating decoding. Trained on 71K hours of speech data, SpeechKV compresses the speech to approximately text-level granularity yet maintains performance on par with or even slightly better than the uncompressed baseline, with relative gains of 6.6% on out-of-domain entity recognition and 2.3% on OpenASR, while delivering at least 1.49 times decoding speedup that scales with audio length.

Federated learning (FL) over mobile and edge devices increasingly involves multimodal models in which clients differ in both sensing capability and computational capacity. Existing update compression schemes typically apply uniform policies across layers and devices, without accounting for modality-specific differences in spectral structure and compressibility. We propose MESH-FL, an entropy-guided matrix product state (MPS) update-compression framework for modality-heterogeneous FL on resource-constrained devices. MESH-FL estimates the spectral entropy of each layer-wise update via truncated singular value decomposition and allocates MPS compression ranks adaptively across layers, modalities, and devices under per-client payload budgets. We show that higher spectral entropy necessitates a higher reconstruction rank under the majorization order on singular-value energy distributions. Building on this result, we prove that the proposed entropy-guided allocation solves a convex surrogate rank-allocation problem, preserves monotonicity under the exact payload model, and achieves convergence with an explicit compression-dependent error term. Experiments on a 15-node heterogeneous Raspberry Pi~4/5 cluster with modality-heterogeneous clients show that MESH-FL achieves up to $56.8\times$ compression while surpassing the uncompressed FedAvg baseline in final accuracy by up to 2.01%, and reduces total transmitted data to reach convergence by up to $66\times$.

Why do intelligent systems need to perform explicit symbolic reasoning? Computer science has traditionally regarded symbolic reasoning as a defining component of intelligence. Yet the remarkable success of modern foundation models raises a fundamental question: if increasingly capable AI systems can operate with little explicit symbolic reasoning, what role do symbolic methods actually play? This article argues that explicit symbolic reasoning is not a fundamental property of intelligence, but a computational consequence of operating on simplified models of reality. We propose the Compression Principle: every computational model is a simplified representation of reality, and explicit symbolic reasoning compensates for information omitted during model construction. From this principle, we derive the Modeling--Reasoning Trade-off: as computational models preserve richer representations of the world, the need for explicit symbolic reasoning correspondingly decreases. This perspective provides a unified explanation for both the historical success of symbolic methods and the remarkable effectiveness of modern foundation models. Paradoxically, the same development makes symbolic methods increasingly important for humans. As intelligent systems become more capable and more opaque, symbolic representations increasingly serve as interfaces through which humans specify requirements, verify behavior, regulate autonomous systems, and establish trust. We therefore argue that the future of symbolic methods lies not primarily as the computational engine of intelligent systems, but as the symbolic interface between increasingly capable AI systems and the humans who build, govern, and depend upon them.

Neural networks are thought to represent concepts as directions in their activation space, and superposition lets them encode more concepts than they have dimensions. It is natural to ask whether they can also compute more functions than they have neurons, i.e., perform computation in superposition. In this regime many functions of sparse inputs are evaluated by a layer with fewer neurons than there are functions to compute. Representation in superposition is by now fairly well understood, but computation in superposition is not, and there are few toy models of it arising through training rather than being hand designed. As a toy model of computation in superposition we study the compressed-computation setup: a single-hidden-layer ReLU network with 50 neurons that must compute the ReLU of each of 100 sparse input features. We show that training it under an $L^4$ loss (the mean fourth power of the error), rather than the usual $L^2$, elicits a solution that appears to compute all features in superposition. We then reverse-engineer this solution. We find that the network assigns each feature a sparse binary codeword over neurons and decodes it with a pseudoinverse of the encoder. Given these codewords, a description with only three scalars recovers most of the network's performance, and we validate it by building equivalent networks from hand-designed codes.

Deep neural networks often contain substantial hidden-state redundancy, but most compression methods operate directly on weights, neurons, or quantised representations without explicitly characterising the dynamical role of internal states. This paper proposes a controllability-observability framework for empirical state-order reduction of deep neural networks. By viewing a trained network as a depth-indexed nonlinear dynamical system, we construct data-driven reachability, observability, and balanced Gramians from hidden-state snapshots and output Jacobians. The resulting A/B/C tests estimate layer-wise reachable, observable, and jointly reachable--observable ranks. These ranks are then used not only as diagnostic measures of hidden-state redundancy, but also as actual compressed layer widths for realised reduced networks. Experiments on MNIST and CIFAR-10 compare the proposed balanced realisation against projection-based reduction, unstructured pruning, structured pruning, low-rank SVD, dynamic INT8 quantisation, and linear baselines. On MNIST, a four-layer SiLU DNN is reduced from state order 1024 to 277, giving 72.95% state compression and 73.48% parameter compression, while maintaining 95.45% accuracy compared with 96.60% for the full model. On CIFAR-10, a larger SiLU DNN is reduced from state order 4608 to 1339, giving 70.94% state compression and 83.09% parameter compression, while preserving accuracy from 54.45% to 54.44% and reducing CUDA inference latency by approximately 3X. The results show that balanced reachable-observable ranks provide a principled empirical minimal-realisation criterion for designing compact neural architectures with little or no loss in accuracy.

We present Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super optimized for interactive deployment. We designed the model to maximize server throughput under high user throughput constraints. In interactive serving workloads on a single 8xB200 node, Puzzle-75B-A9B achieves approximately 2x higher server throughput than Nemotron-3-Super at matched user throughput constraints. In ultra-long-context deployment on a single H100 GPU, the compressed model increases 1M-token concurrency from 1 request to 8 requests. Puzzle-75B-A9B is constructed using a multi-stage pipeline that combines the Iterative Puzzle compression framework with knowledge distillation, reinforcement learning, quantization, and a Multi-Token Prediction head. The compression process jointly optimizes heterogeneous MoE pruning, active parameter budget, and Mamba pruning to improve inference efficiency while preserving model quality. We evaluate Puzzle-75B-A9B on a broad suite of reasoning, coding, multilingual, long-context, and agentic benchmarks. Despite substantial compression, the model retains strong downstream accuracy relative to the parent model across a wide range of tasks. These results demonstrate that large hybrid MoE models can be substantially optimized for deployment efficiency while maintaining strong downstream capability. Our model is publicly available on Hugging Face.

The growing demand for image-to-video creation on mobile devices has increasingly focused on cinematic motion effects like bullet time, dolly zoom, slow motion, etc. While Diffusion Transformers (DiTs) exhibit strong performance in video generation, their large parameter sizes and multi-step iterative denoising processes lead to substantial computational overhead, making efficient generation on mobile devices challenging. We propose CineMobile to bridge the gap. In particular, CineMobile adopts a three-fold optimization strategy: (1) leveraging a distillation-guided pruning approach to derive a compact yet efficient model that retains the essential video generation capabilities required for cinematic effects; (2) optimizing the compressed model into a 4-step generator via a combination of diffusion distillation and reinforcement learning; (3) employing a hybrid post-training quantization strategy to compress the model footprint to under 1 GB. Experimental results show that compared to the teacher model with the Wan 2.1 architecture, CineMobile achieves a 40x speedup in generation while maintaining comparable visual quality. Specifically, CineMobile generates 49-frame 480p videos with a per-step denoising latency of 0.6s on an NVIDIA H200 GPU and 20s on the MediaTek Dimensity 8400 Ultimate 5G platform, with a peak memory usage of 1.8 GB, demonstrating its practical applicability for mobile-based image-to-video creation.

Point cloud compression relies on techniques to compress both geometry and attributes. Motion-based approaches for dynamic solid point cloud geometry compression within the geometry-based point cloud compression (G-PCC) framework have achieved significant reductions in geometry rate. However, motion-based techniques for attribute compression remain underexplored, making it challenging to achieve significant reductions in the temporal redundancy of attributes. Firstly, this paper proposes a geometry-based inter-coding scheme to compress the attributes of dynamic solid point clouds. Secondly, a graph-based motion-estimation scheme for point-cloud attribute compression is proposed. Thirdly, an interpolation-free fractional-voxel motion estimation method is proposed to refine motion accuracy to fractional-voxel precision. Our experimental results on the MPEG point cloud dataset show that the proposed scheme outperforms G-PCC, GeS-TM, and V-PCC in lossless and lossy geometry conditions. We achieve average bitrate savings of $55.3\%$, $42.3\%$, and $16.5\%$ over G-PCC, GeS-TM, and V-PCC, respectively, under lossy-geometry conditions.

The ever-increasing collection of personal data has created mounting pressure to develop technologies that protect sensitive aspects of individual identity. Differential privacy (DP) provides a principled framework with strong formal guarantees and has already achieved practical success. However, releasing high-dimensional data, such as images, has remained elusive: releasing uncompressed privatized data requires significant storage. At the same time, no effective data compression scheme exists that can compress high-resolution data with privacy guarantees. We address this challenge with DP-DiPP, a compression pipeline that combines stochastic codes with diffusion models. DP-DiPP is highly flexible: the practitioner has direct control over the compression rate-privacy-utility tradeoff. As the theoretical backbone, we extend the Poisson private representation (PPR) to encode the outputs of privacy mechanisms. We then combine it with DiffC, a diffusion-based lossy data compression method, to obtain a differentially private image compressor. Our experiments on privatized image classification on CIFAR-10 demonstrate that DP-DiPP significantly outperforms the baseline, achieving a 10-30 times better compression while retaining comparable privacy guarantees and utility.

Omni modal large language models (OmniLLMs) have attracted wide attention for their ability to jointly process audio and video, but they generate large token sequences under audio-visual inputs, leading to substantial inference cost. Existing audio-visual token compression methods often rely on unimodal guidance, overlooking the temporal locality of query-relevant evidence in audio-visual inputs and implicitly assuming that the two modalities share a temporally aligned information density distribution. We propose \textbf{OmniFocus}, a training-free query-guided token compression method for OmniLLMs that performs independent importance estimation for video and audio, enabling a modality-symmetric compression design that preserves modality-specific salient evidence while maintaining audio-visual alignment, thereby mitigating the modality bias issue that can arise from unimodal-guided compression. Experiments on the Qwen2.5-Omni model family across four audio-visual benchmarks show that OmniFocus maintains strong compressed performance at low token retention ratios and outperforms existing baselines on several major benchmark scores at 25% token retention. On DailyOmni with Qwen2.5-Omni-7B at 25% token retention, OmniFocus maintains 59.40 accuracy while delivering up to 1.38$\times$ prefill speedup relative to the full-token baseline, highlighting a favorable practical accuracy-efficiency trade-off.

LLM-based coding agents solve software-engineering tasks through iterative interactions with development environments, where returned observations accumulate in the context and become a major source of inference cost. Observation compression reduces this cost by shortening observations before they are appended to the context. However, existing methods still exhibit an unsatisfactory efficiency-effectiveness trade-off, as they do not explicitly model how compression affects the agent's subsequent behavior. This paper proposes CoACT, an action-preserving observation compression method for coding agents. CoACT is built on next-action preservation (NAP), which requires a compressed observation to induce the same next action as the raw observation. By checking the agent's immediate next action, NAP provides a practical signal for whether a compression preserves the information needed for continued task solving. During training, a teacher model first generates multiple compressed candidates of each observation. CoACT then uses an action-preservation reward based on NAP to filter out candidates that would change the agent's next action, and uses a length-reduction reward to choose compact candidates as supervision for a lightweight compressor. Experiments on SWE-bench Verified with three agentic models show that CoACT reduces average total token consumption by 33.0% while maintaining task-solving effectiveness close to the uncompressed agent.

How can Large Language Models (LLMs) solve problems they currently cannot? Repeated sampling scales test-time compute but GPU cost grows linearly with attempts, while reinforcement learning (RL) with verifiable rewards improves single-attempt accuracy at the expense of sample diversity. Both strategies ultimately fail when the base policy has near-zero probability of producing a correct solution: no amount of sampling or gradient signal can overcome a search space that is simply too large. We take a different approach: rather than sampling harder, we make the task easier by decomposing problems into smaller, independently solvable sub-functions whose implementations can be recombined. Since off-the-shelf models are not trained for this modular generation, we introduce DecompRL, an RL algorithm that explicitly learns to decompose and implement hierarchical code structures. Recombining $k$ implementations of $n$ modules yields up to $k^{n}$ candidate solutions, shifting the bottleneck from GPU inference to cheap CPU evaluation and cutting GPU token cost by $\sim$50$\times$. On LiveCodeBench and CodeContests (Qwen2.57B, Code World Model~32B), DecompRL outperforms standard and diversity-optimized RL baselines beyond $10^5$ tokens per problem, solving problems that standard generation cannot reach.

In this paper, we consider the nonsmooth nonconvex decentralized optimization problem, where inter-agent communication is compressed. We propose a general framework that unifies various decentralized stochastic subgradient-type methods with unbiased compression and contractive compression with error compensation. By relating the consensus-error iterates and the averaged iterates to the trajectories of continuous-time differential inclusions, we establish global convergence for all methods encompassed by our framework when the objective functions are nonsmooth and lack Clarke regularity. Based on our framework, we further develop several compression-based methods, including decentralized stochastic subgradient methods utilizing sign-based regularization and gradient-tracking momentum. Preliminary numerical experiments empirically support our theoretical results and highlight the communication-accuracy trade-off of the newly developed methods.

Transformer inference on long sequences is expensive because softmax attention repeatedly reads from a large KV cache. The prevalent approach to this bottleneck is KV cache compression, which replaces the full cache with a compact summary. Despite its practical importance, the design of such summaries is largely driven by empirical experimentation. On the theoretical side, existing results show that KV cache compression can be impossible in the worst case, but offer little systematic guidance for designing algorithms in regimes where accurate compression is possible. We bridge this gap by characterizing the minimax risk of KV cache compression in terms of the intrinsic compressibility of a cache, revealing when and how accurate compression is possible. These results yield novel design principles for KV cache compression under causal masking that map efficiently to prefill and autoregressive decoding while achieving minimax-optimal risk. We instantiate these principles in a practical algorithm and report promising performance on LongBench in targeted experiments. Overall, our results provide a principled avenue for practical KV cache compression with theoretical guarantees.

Recent work has shown that implicit neural representations (INRs) can be trained to effectively compress structured and unstructured volume data, allowing for direct data querying with a reduced memory footprint. However, as existing INRs for unstructured volumes do not encode geometry, they require partial mesh storage for later sampling, limiting achievable compression. At the same time, novel view synthesis methods have shown that explicit collections of 3D Gaussians can be used to accurately visualize volume data. In this work, we introduce an explicit model for volume data compression based on 3D Gaussian primitives. We reinterpret collections of 3D Gaussians as an explicit representation of a scalar field and use a sampling strategy that reconstructs scalar values at spatial locations through weighted aggregation of intersecting Gaussians. We develop optimized CUDA-accelerated pipelines for structured and unstructured model sampling, loss functions that encourage accurate domain encoding by our models, and a novel sampling-error based densification strategy. Our explicit formulation naturally encodes domain geometry, eliminating the need for mesh storage in unstructured volumes and introducing significantly higher compression opportunities. Compared to existing INRs, we demonstrate that our explicit model achieves competitive reconstruction quality with significant training speedups on structured volumes, while markedly outperforming in all metrics on unstructured volumes.

Driven by Artificial Intelligence-Generated Content (AIGC), the authenticity of audio-visual content is facing severe challenges. Temporal Forgery Localization (TFL) aims to precisely identify manipulated segments within untrimmed sequences. However, existing methods are limited by CNNs' local receptive fields or Transformers' quadratic complexity, while emerging linear models often struggle to balance global authentic context compression with local abrupt forgery perception. To address this, we propose MG-RWKV, a multi-granularity framework that leverages the data-dependent state evolution of RWKV to achieve efficient full-sequence processing with O(T) complexity. Our framework features three core innovations: (1) a Bidirectional RWKV architecture that captures bidirectional temporal contexts without quadratic overhead; (2) a Multi-Granularity Mixture of Experts (MG-MoE) that performs dynamic routing over explicit temporal receptive fields, adaptively selecting granularities based on forgery duration to significantly enhance decision interpretability; and (3) Cross-Granularity Consistency (CGC), which aligns adjacent feature pyramid levels through hierarchical scale-wise pairing and spatial boundary-aware weighting, effectively reducing false positives in authentic regions. Extensive experiments on Lav-DF, TVIL, and Psynd datasets demonstrate that MG-RWKV achieves state-of-the-art performance with low computational cost.

We propose the first compression approach for image-to-shape Diffusion Transformers (DiTs) that substantially reduces model size while preserving geometric fidelity. Despite remarkable progress in 3D shape generation, large DiT-based models remain computationally prohibitive in resource-constrained settings. Furthermore, it is difficult to directly transfer existing diffusion model compression strategies developed for different domains to 3D generation, and prior 3D efficiency approaches focus primarily on inference speed rather than backbone compression. To address this limitation, we build a geometry-aware compression framework tailored to image-to-shape DiTs. Guided by the observation that 3D DiT layers exhibit non-uniform importance for geometry synthesis, we introduce a vitality-guided framework integrating structured pruning, adaptive quantization, and targeted fine-tuning. Our method achieves up to 66% model-size reduction across state-of-the-art image-to-3D models while maintaining synthesis fidelity comparable to full-sized counterparts. This highlights the potential of our framework as a plug-and-play solution for efficient 3D shape generation across diverse models.

Deploying large convolutional neural networks (CNNs) on resource-constrained devices is challenging due to their high computational cost. While dynamic execution methods are promising, existing approaches for CNNs typically require specialized training or fine-tuning, limiting their effectiveness when applied to pre-trained models and requiring data access. To address this gap, we propose HASTE (Hashing for Tractable Efficiency), a plug-and-play convolution module that enables training-free, dynamic compression of large pre-trained CNNs. At inference time, HASTE uses locality-sensitive hashing to identify and merge redundant channels of latent feature maps on a patch-wise basis. This process simultaneously compresses the depth of both input features and their corresponding filters, resulting in computationally cheaper convolutions. We conduct extensive experiments on CIFAR-10 and ImageNet across a range of architectures, demonstrating a 46.2% FLOPs reduction in a ResNet34 on CIFAR-10 with only a 1.25% drop in accuracy, without any retraining. We support our claims by comprehensive ablation studies to validate our core design choices, an analysis of the method's properties and limitations, and a discussion that connects our channel merging scheme to the conceptually related task of token merging in Vision Transformers. Our results demonstrate that HASTE provides an effective solution for steerable compression of pre-trained CNNs at runtime, opening new possibilities for the deployment of efficient deep learning methods.

Large language models achieve high reasoning performance via explicit chain-of-thought and reinforcement learning, but require long output sequences and extended inference time. Latent reasoning reduces this cost by shifting computation into a latent space; however, continuous latent methods are hard to train, suffering from unstable and uninterpretable reasoning trajectories. We argue these issues stem from a misalignment between continuous-space reasoning and discrete symbolic supervision, as continuous states lack explicit anchors for step-by-step alignment. To resolve this, we propose \textbf{Discrete Latent Reasoning~(DLR)}, the first method that converts continuous latent states into explicit discrete tokens. Inspired by render-based compression, we render textual chains of thought into images, extract visual features, and construct a discrete latent vocabulary via clustering-based fine-tuning. Expanding the vocabulary and output head enables standard autoregressive modeling over both natural language and latent tokens, supporting pretraining alignment, SFT, and RL. Experiments on five reasoning benchmarks and two model series~(Qwen3-VL and LLaMA-3) confirm that \textbf{DLR} outperforms prior latent reasoning baselines with up to \textbf{20$\times$ compression}. Furthermore, the learned latent trajectories retain an interpretable semantic structure. Overall, discrete latent tokens provide a controllable and interpretable basis for efficient latent reasoning.

Learned video codecs based on continuous latent representations struggle to operate reliably below 0.1 bits per pixel~(bpp): without a differentiable rate signal, Lagrangian optimisation cannot effectively trade reconstruction quality for bitrate at extreme compression ratios. We demonstrate that discrete latent representations sidestep this limitation entirely. In a vector-quantized~(VQ) codec, the codebook size~$K$ imposes a hard information ceiling of $\log_2 K$ bits per symbol; a learned autoregressive prior then exploits the non-uniform distribution of code usage -- which we show follows a power law -- to push actual bitrates well below this ceiling, without any rate-penalty tuning. Building on the MS-VQ-VAE architecture introduced in~\cite{kotthapalli2026msvqvae}, we sweep $K \in \{128, 256, 512, 1024\}$ under a uniform training protocol to trace four operating points on the rate-distortion~(RD) curve. We identify and resolve a critical training instability: gradient-based VQ collapses catastrophically at $K \leq 512$, whereas EMA-stabilised codebook updates with dead-code restart maintain full utilisation across all configurations. On 500 UCF101 test clips ($64\!\times\!64$, 32frames), our models operate at 0.043-0.064bpp -- 3.3-5$\times$ below H.264's practical floor and $5$-$7.6\times$ below H.265's floor at this resolution. Every MS-VQ-VAE configuration outperforms H.265 CRF,36 on perceptual quality (LPIPS) despite using $5$-$7.6\times$ fewer bits. At $K{=}1024$, the model surpasses H.265 CRF,36 on LPIPS by a margin of 0.072 absolute while using $5.1\times$ fewer bits. Codebook analysis confirms power-law index distributions and 70-85% entropy efficiency, establishing the pipeline as a principled learned entropy coder.

Deep learning problems rarely involve objectives that are equal in importance. A primary objective defines the goal, whilst secondary objectives, such as sparsity, compression, or robustness constrain the solution. While existing multi-objective methods have proven effective in practice, they have a clear symmetry problem and neglect the inherent objective hierarchy built into these objective spaces. We introduce Priority-Constrained Descent (PCD), a gradient-based optimization framework designed to explicitly exploit hierarchical objective structures. PCD preserves the direction of primary descent whilst allowing for the minimal distortion necessary to guarantee progress on secondary objectives, controlled by a single $τ\in [0, 1]$ that dictates the strength of the distortion. The resulting formulation is invariant to objective scaling and admits exact closed-form solutions for problems with two and three objectives. We evaluate PCD within structured network compression settings, unstructured sparsity and low-rankness, and across a variety of synthetic experiments, showing Pareto dominance and better per-objective performance with secondary progress guarantees over existing methods, further exhibiting the interpretable trade-off that $τ$ provides.

Financial decision-makers face more information than they can directly inspect, making context compression necessary. Yet when large language models (LLMs) compress financial source material, they can alter the investment judgment supported by the original source. We frame this problem as information fidelity: compression loses fidelity when it changes the decision induced by the source. In agentic systems, such losses may recur across intermediate steps and amplify throughout the decision process. Across financial filings and earnings-call transcripts, we find that LLM-based compression can produce fluent and factually plausible compressed contexts that nevertheless alter downstream decisions. We analyze two diagnostic patterns associated with fidelity loss: decontextualization, where salient evidence is retained but separated from the caveats and contextual qualifiers needed for correct interpretation, and model dependency, where different compressors expose different views of the same source. We then propose Agentic Context Compression, which generates multiple candidate compressions and audits their disagreements against the original source. Our results suggest that financial compression should be evaluated not only by efficiency or factuality, but also by its ability to preserve decision-relevant context.

Accurate and timely channel state information (CSI) is essential for next-generation wireless systems, yet existing works treat CSI compression and CSI prediction as separate problems, both in academia and in current 3GPP studies. Consequently, channel aging remains insufficiently addressed within standardized CSI feedback pipelines. In this article, we propose a unified compression-prediction framework that integrates Contrastive Predictive Coding (CPC) directly into the 3GPP-compliant CSI compression architecture. Instead of predicting high-dimensional CSI matrices, our approach forecasts future latent representations and jointly optimizes reconstruction fidelity and temporal predictive coherence via a combined 1-SGCS and InfoNCE objective. This design enables temporal representation learning without increasing feedback overhead. We present two variants: CPC-before-Compression, which performs autoregressive modeling on encoded features prior to quantization, and CPC-after-Compression, which shifts temporal modeling to the base-station to reduce the complexity of users' devices. Evaluations on 3GPP-compliant datasets from Nokia, Oppo, and CATT show that CPC-before-Compression achieves over 90% reconstruction accuracy with 32x lower decoder GFLOPs than the 3GPP baseline, while CPC-after-Compression preserves an identical encoder footprint and the same 64-bit feedback overhead. By unifying compression and prediction within a standardized pipeline, the proposed framework provides an age-aware, computationally efficient CSI feedback solution. The source code is publicly available at: https://github.com/AhmedRadwan02/cpc-3gpp

Implicit neural representations (INRs) have recently emerged as a promising approach to video compression, delivering competitive rate-distortion performance alongside rapid decoding. However, existing neural video codecs struggle to balance complexity and scalability. Lightweight models often suffer from degraded compression performance when scaled to different bitrate/quality levels, whereas high-performance models exhibit limited scalability, as their model complexity typically increases with quality. This lack of a unified architecture capable of maintaining consistent complexity across a wide range of bitrates severely limits their diverse real-world deployment. To address these challenges, we introduce NVRC++, a novel INR-based video codec that utilizes a lightweight INR with multiple high-resolution feature grids, providing high scalability at any given complexity level. This is paired with an optimization framework that enables efficient overfitting on high-resolution grids for long video sequences, thereby exploiting spatio-temporal redundancies without prohibitive computational or memory overhead. Additionally, an advanced entropy model is designed for efficiently compressing the high-dimensional grid parameters. As a result, NVRC++ provides four complexity levels (from 7kMACs/pixel to 360kMACs/pixel), each spanning wide bitrate and quality ranges while supporting real-time decoding. The experimental results show that NVRC++ offers a much faster decoding speed (up to 7.6x) compared to the SOTA INR-based video codec, NVRC, while delivering comparable performance.

Training-free compression methods for large language models (LLMs) often use calibration data to guide compression decisions. ROCKET, a recent method combining sparse-dictionary factorization with multi-choice knapsack problem (MCKP) allocation, derives its per-layer factorization from an output reconstruction objective but uses weight-space Frobenius error as the MCKP allocation cost. We investigate whether aligning the allocation cost with the output-space objective improves compressed model fidelity. On Qwen3-8B at 50% compression, our ROCKET-ActCost achieves +0.8 percentage points higher average accuracy across 8 zero-shot benchmarks (53.1% vs 52.3%), but increases WikiText perplexity by 16% (61.46 vs 52.98). This accuracy-perplexity tradeoff reveals that different allocation objectives favor different downstream metrics. The high correlation ($>$0.99) between weight-space and output-space errors limits allocation divergence, explaining the modest effect size. On Llama-3.2-1B at 20% compression, the two methods produce near-identical results (53.3% vs 53.5% accuracy, 14.45 vs 14.66 PPL), suggesting that the effect of the cost function is minor at lower compression ratios.

Recent work on neural texture compression has demonstrated that it is possible to learn small, per-material texture representations (composed of latent textures and a small Multi-Layer Perceptron decoder) that can be decoded in real-time during shading to reproduce the input to a physically based shading model. However, existing methods require performing gradient-descent optimization per material for a given MLP and latent configuration. In this work, we train a single hypernetwork that outputs both the latent features and the MLP's weights and biases. Though the solution space is high-dimensional, this approach produces results comparable in quality to the current reference neural texture compressors. We further extend this approach to infer multiple decoders at once or even produce decoders that learn super-resolution.

We present HiReLC, a hierarchical ensemble-reinforcement learning framework for automated joint quantization and structured pruning of deep neural networks. The framework decomposes the compression search across two levels of abstraction: low-level agents (LLAs) operate independently per block, selecting per-kernel configurations over a multi-discrete action space spanning bitwidth, pruning keep-ratio, quantization type, and granularity, while high-level agents (HLAs) coordinate global budget allocation via ensemble voting guided by Fisher Information-based sensitivity estimates. To mitigate the computational cost of policy evaluation, an iterative active learning loop interleaves surrogate-guided RL optimization with post-compression fine-tuning, using a lightweight MLP surrogate to amortize expensive evaluations and a logit-MSE proxy during cold-start. The surrogate is used for reward shaping rather than as a replacement for final post-compression evaluation. The controller is architecture-agnostic by design, with a modular layer abstraction decoupling the RL environment from the underlying network topology. Experiments across Vision Transformer and CNN benchmarks demonstrate effective parameter-storage compression ratios of 5.99 - 6.72$\times$ with a 3.83 % gain in one setting and 0.55 - 5.62 % accuracy drops elsewhere, supporting hierarchical policy decomposition and sensitivity-aware guidance as practical design choices for joint neural network compression.

"Talk short. Drop grammar. Save token." This caveman style is widely promoted as a way to cut inference cost, but whether it actually saves anything depends on which channel (the user's prompt or the model's response) is being compressed. We present Cavewoman, a two-channel evaluation protocol that scores every generation on task accuracy, realized per-item cost, and reference-text agreement against the model's unconstrained reference. We evaluate eight models on five datasets at five reduction levels, with both channels measured on the same items. Output compression cuts realized cost on most API models (1.4-2.4x per model, up to 3x in the best case) and on all four open-weight models under public-tier pricing. Input compression has the opposite effect, a strict lose-lose: it raises net cost rather than lowering it (~1.15x on the five-benchmark mean, up to 1.8x on the worst dataset and 2.7x under stronger compression), because models compensate with longer responses even as accuracy collapses. Under the same setting, surface text diverges from the unconstrained reference: on the non-reasoning models, roughly half of all generations are correct yet their surface text no longer entails the model's own unconstrained baseline generation. The divergence survives length-controlled re-scoring, multiple-comparisons correction, and replication under complementary semantic measures. Code and data are available at https://github.com/danielle34/cavewoman.

We perform a realistic KiDS-Legacy mock analysis with field-level neural compression and simulation-based inference using just 60 $N$-body simulations. The weak lensing shear field encodes substantially more cosmological information than standard two-point summary statistics such as the power spectrum. Field-level inference can fully exploit this information, but physical realism at the field-level requires very high-fidelity simulations. This poses a major challenge for simulation-based inference (SBI): accurate empirical density modelling and deep-learning-based neural compression require tens of thousands of training samples, but achieving physical realism at the field level makes each simulation extremely costly. We demonstrate that multifidelity SBI can alleviate this tension by substantially reducing the number of high-fidelity simulations needed for accurate cosmological inference. We pre-train neural inference models on realistic KiDS-Legacy-like shear mocks using fast log-normal \texttt{GLASS} simulations and fine-tune them on a small set of high-fidelity $N$-body simulations. We show that $60$ high-fidelity simulations are sufficient to obtain informative and well-calibrated cosmological posteriors, enabling at least an order-of-magnitude reduction in simulation cost for accurate field-level inference in a realistic setting.

Video world models hold promise for simulating interactive environments, yet maintaining consistent long-term memory across complex camera trajectories remains a critical challenge. Existing methods typically rely on computationally expensive context scaling or rigid heuristic retrieval mechanisms, which lacks generalization to varying camera trajectories and environments. In this paper, we propose Compression and Retrieval (CaR), an attention-driven implicit memory retrieval mechanism to overcome these limitations. By injecting viewpoint information via positional encoding, our method performs flexible memory retrieval through attention computation. To efficiently process extended contexts with minimal computational overhead, we further introduce a lightweight context compression network. Furthermore, we construct SceneFly, a large-scale synthetic dataset featuring realistic camera trajectories and frame-level annotations to train and evaluate long-horizon video world models. Extensive experiments demonstrate that our approach achieves state-of-the-art results on established benchmarks and exhibits strong generalization to open-domain scenes.

Deep neural networks have witnessed remarkable advancements in recent years and have become integral to various applications. However, alongside these developments, training and deployment of neural network models on embedding and edge devices face significant challenges due to limited memory and computational resources. These problems can be addressed with deep neural network compression, which involves a trade-off between model size and performance. In this paper, we propose a novel method for model compression through two phases. First, we utilize model compression techniques, such as pruning and quantization, to significantly reduce the model size. Then, we use Mixture of Experts to route the previously compressed models to enhance performance while maintaining a balance in inference efficiency. MoEs consist of multiple expert models (i.e., compressed models) that are moderately sized and deliver stable performance. Experimental results on several benchmark datasets show that our method successfully compresses CNN models which achieves substantial reductions in FLOPs and parameters with a negligible accuracy drop.