A compressed student has two shapes that need not agree: the weight it deploys at inference and the weight family its training can reach. We show that a state-of-the-art weight-inheritance distiller, Low-Rank Clone (LRC), deploys a full-width student MLP but ties training to a teacher-induced slice, leaving 62.5-81.4% of each deployed matrix's independent linear degrees of freedom unreachable-paid for at inference, never trainable. Our principle is one line: train what you deploy. From the identical LRC warm start, we make the training object the entire deployed matrix, with no change in deployed shape, deployed parameter count, or inference FLOPs, via two mergeable realizations (Dense-LRC and CORE-LRC) that both collapse to one deployed weight. This recovers stranded capacity: taking the stronger realization per teacher, +2.36/+2.71/+10.45 Avg9 over matched-budget plain-LRC baselines across three teachers (Llama3.2-3B, Llama3.1-8B, Qwen2.5-3B), with the largest gain on the widest teacher (Qwen), where it reaches the original recipe's approx. 20B-token accuracy at 10B tokens (2x token efficiency); there the strictly same-lineage arm still recovers +6.39, the fully controlled figure. Controls strongly support attributing the gain to the enlarged reachable set, rather than to added parameters or the recipe. From approx. 10B distillation tokens plus a short SFT, a half-parameter 1.5B student matches its approx. 9T-token teacher's 9-task macro-average, within evaluation noise and with a residual MMLU deficit, and a 2.7B student beats Meta's own official compression of Llama3.1-8B at ~900x fewer compression tokens (a token count under unmatched recipes, not a compute claim). All results are from single-seed runs on the LRC backbone.
Large Language Model (LLM) agents based on the ReAct paradigm have demonstrated remarkable capabilities in tool use and task execution. However, ReAct suffers from a fundamental efficiency problem: every query triggers a complete reasoning loop from scratch, and similar queries repeat identical steps without leveraging historical experience. We propose TRIAGE,a three-level routing framework that reduces token consumption by reusing historical execution trajectories. Its core innovation is TaaS (Trajectory-as-a-Skill), which abstracts historical execution trajectories into reusable skills, realizing 'experience as a service'. TRIAGE classifies queries into three levels: (1) Direct Reuse-identical queries, 0 tokens; (2) Skill Substitution-similar queries, 0 tokens via deterministic parameter substitution; (3) Full ReAct-novel queries, automatically stored for future reuse. In large-scale experiments on 1,007 security monitoring queries, TRIAGE achieves 62.3% token savings, with 56.0% of queries at Level 2 and 5.5% at Level 1, both executing at zero cost. Cross-domain validation on ToolBench (15 domains, 345 queries) achieves 76.3% token reduction, confirming the generalizability of semantic routing. An online learning experiment demonstrates cold-start-to-mature evolution: the L2 hit rate rises from 0% to 57% within the first 100 queries, and the average token cost drops from 198 to 74.7. We also propose an automatic Skill extraction mechanism that distills high-frequency trajectory patterns into deterministic Skills, creating a positive feedback loop of 'the more you use it, the more efficient it becomes'.
Production agent skills are directory bundles, not isolated prompts. The root is loaded at activation; references, schemas, scripts, assets, and nested subskills are loaded only when an execution path needs them. Compressing only the root misses most deployment cost and may move branch-specific details into the always-loaded context. Flattening instead destroys progressive-loading boundaries. We introduce \method, an evaluation-free compressor for complete, progressively loaded skill bundles. It leaves the agent harness unchanged and emits an ordinary directory. The method combines two safeguards. First, it compresses \emph{across files}, removing content from a reference or subskill when the root or a declared environment contract already provides it. Second, it preserves routing, so every required file and directly callable entry remains reachable after rewriting. Users can configure \method along two independent axes. \emph{One-Shot} mode rebuilds the full bundle; \emph{Continual} mode reuses state and applies Zip-on-Write after each evolution patch. \emph{Persistent} compression rewrites the shipped bundle to reduce storage and runtime context. \emph{Transient} compression keeps that bundle byte-identical and builds a task-specific view, reducing only per-run context after build cost. Entry contracts mark private, public, and conditional resources; a multi-entry audit preserves standalone public subskills. On a production content-moderation skill evaluated by our industrial multi-round harness, \method removes \hl{38\%} of skill bundle tokens and \hl{10.4\%} of end-to-end per-run tokens with no quality loss, while an unprotected 71\% configuration loses up to 26 accuracy points to one-sided false positives. On a multi-entry bundle, \method effeciently reduces token cost while near-perfectly preserving every route and public entry.
Jie Liang, Zhengxin Yu, Hamid Nasiri +1cs.AI cs.CL
LLM agents need to sustain goal-consistent reasoning across long multi-turn interactions under strict resource constraints. However, as the multi-turn context accumulates, it can destabilize the underlying LLM's internal representation of task-relevant information from earlier turns, blurring the boundary between constructive reasoning and representation drift. We formulate multi-turn reasoning as a hidden-state trajectory of the underlying LLM that is characterized via two complementary signals: temporal curvature that captures the directional consistency of turn-to-turn updates, and variance slope which measures the expansion or contraction of the exploration space. Across four tasks and three underlying LLMs, we observed that these geometric signals distinguish between correct and incorrect episodes prior to completion. We further decompose each episode into three-action chains formed from four actions (Read, Write, Respond, Transfer) and show that separability is action-dependent, with different signals distinguishing various chain patterns. Our experiments demonstrate that trajectory geometry can identify critical turns in the reasoning process, increasing task success rates on $τ$-Bench from 24.1% to 39.6% while reducing token cost by 11.2%.
Cheolseung Baek, Dhammiko Arya, Eunki Kim +40cs.AI cs.CL
We introduce A.X K2, a 688B-parameter Mixture-of-Experts (MoE) language model trained from scratch as a high-performance foundation for \emph{agentic} applications. Trained on approximately 8.5T tokens---fewer than its predecessor, A.X K1---on a smaller but higher-quality mixture with substantially expanded agentic and software-engineering data, it nonetheless improves over A.X K1 across the board, by over 30 percentage points on some benchmarks, reflecting large gains in token efficiency. To support long contexts efficiently, we introduce Sparse Gated Attention (SGA), which combines sparse attention with gated attention, and adopt Gated Norm (GN) to stabilize large-scale training. SGA is trained natively at 128K through a \emph{sparse} indexer warmup that optimizes the indexer against its own sparse top-$k$ selection rather than the dense attention distribution, making adaptation markedly cheaper: each query reads only 2,048 positions, yet long-context quality is unchanged and A.X K2 scores 94.6 on RULER out to 256K. The outlier suppression of GN in turn keeps 4-bit NVFP4 serving within one point of FP8 accuracy. A simple yet effective Think-Fusion recipe further lets users switch between thinking and non-thinking modes within a single unified model. Extensive evaluations show that A.X K2 performs competitively against strong open-weight baselines, matching or exceeding them on math and Korean-language benchmarks.
Enterprise data agents answer business queries by chaining many tool calls over multiple reasoning steps, routinely accumulating hundreds of thousands of context tokens per session. Existing compression strategies typically allocate retention budgets without accounting for the downstream consequences of removing individual tool outputs. Aggressive compression may therefore trigger costly tool re-invocations that offset the initial savings. We call this the compression--consequence gap. To close it, we propose TRACER, which formulates compression as a sequential per-tool decision problem. A lightweight REINFORCE policy assigns query-conditioned retention ratios using only information available at each compression event. Its consequence-aware objective jointly accounts for task success, total token consumption, and post-compression tool re-invocations. To improve credit assignment, TRACER uses a learned outcome model to compare the predicted consequences of the selected retention ratio with those of fully retaining each tool output. On held-out production queries across three compressor backends, TRACER reduces total token consumption by 29--46% relative to keeping all context while maintaining comparable or higher task success. Compared with a tool-type-conditional static policy, TRACER provides an additional 15--18% of token savings. Interventional rollouts show that the learned per-tool credit scores correlate with measured single-tool consequences. The learned policy also yields positive savings when transferred across agent backbones and compressor architectures, and reduces token consumption by 18--25% on five held-out LOCA-bench environments. These results demonstrate the value of consequence-aware, per-tool context retention for improving the efficiency of long-horizon language agents.
Large Language Models (LLMs) increasingly act as autonomous agents executing complex, long-running procedural skills. Existing agent runtimes maintain execution by continually appending observations, actions, and intermediate reasoning traces to an ever-growing conversation history, causing latency degradation and context-poisoning failures over long horizons. We present SKILL. state, a runtime architecture that replaces append-only conversational history with an explicit, mutable execution state. At each execution step, the model receives only the immutable skill specification, the current structured execution state, and the latest observation. Intermediate reasoning is discarded immediately after producing a validated state update, preventing prompt growth with execution history. Across diverse datasets, models, and execution environments, SKILL. state improves task accuracy while substantially reducing cumulative token consumption. Our results demonstrate that explicit execution state is an effective and architecture-agnostic abstraction for scalable long-horizon agent skills.
Large reasoning models often produce reasoning traces with verification, revision, and backtracking. When reflection merely re-checks established results, it wastes reasoning tokens and increases latency. Most existing reflection steering methods add a label-derived mean-difference direction across preset layers, but its entanglement with reasoning and length signals destabilizes the accuracy-efficiency trade-off. In this paper, we propose Reflection Steering, a training-free framework for controlling reflection-associated computation within LLMs by disentangling reflection-related activations from general reasoning. Specifically, we contrast reflective and non-reflective hidden states at each LLM layer, denoise the resulting reflection directions with PCA, and orthogonalize them against general-reasoning directions. To limit downstream amplification from early-layer interventions, we calibrate each layer across multiple intervention strengths on a small set, retain only stable layers, and apply bounded projection removal to their residual-stream activations. We conduct extensive experiments across two public benchmarks and three open-weight LLMs against state-of-the-art activation-steering baselines. Results show that Reflection Steering reduces reasoning tokens by 16.9% on average across six matched settings. Besides, our method further introduces a bounded reflection intervention-strength parameter $α$, enabling deployment-time adjustment to balance token savings, accuracy, and generation stability.
Multi-agent LLM systems coordinate through natural-language messages that consume 40--60\% of their token budget. Replacing these with structured graphs reduces cost but fails on tasks requiring adaptive reasoning. We propose \textbf{Routed Graph Handoff}, where a lightweight LLM router (155 tokens, 0.15\% overhead) selects between a typed dependency graph and natural language for each delegation. On four benchmarks (1,050+ trajectories), the routed system matches or exceeds NL-only on every task: \textbf{+12.7\,pp} on $τ$-retail at 3.2$\times$ compression ($p{<}0.01$), \textbf{+8.7\,pp} on BrowseComp at 2.2$\times$ compression ($p{<}0.05$), and parity on BFCL and AppWorld. Without the router, graph-only delegation regresses 14.6\,pp on AppWorld; the router eliminates this at near-zero cost. A graph-aware executor prompt is required: the same schema without interpretation guidance yields no gain. An oracle analysis reveals 8.6\,pp of additional headroom, motivating execution-time adaptive routing as future work.
Long-horizon agents increasingly rely on repeated access to external artifacts, yet current reading interfaces often expose entire objects even when only sparse evidence is needed. This over-reading increases token and latency costs and can dilute task-relevant evidence, while existing context-reduction methods mainly intervene after broad content has already entered the trajectory. We present SparseRead, a training-free, model-transparent reading layer that controls content admission before unnecessary evidence reaches the model context. SparseRead combines a regime-aware Read Gate, extensible Reader Backends, and a stateful protocol for bounded, source-anchored evidence acquisition with explicit refinement, verification, stopping, and fallback. Across six frontier models, including Claude Opus 5, and five workload scenarios, SparseRead reduces token volume by up to 92.9% and wall time by up to 89.0%, while preserving or improving task quality. Its consistent gains across three agent frameworks further demonstrate broad portability.
Gijs Kassenaar, Zhao Yang, Vincent François-Lavetcs.AI
Reasoning language models trained with reinforcement learning typically operate under a fixed token budget rather than an explicitly adaptive one, which can lead to over-computation on easy problems and insufficient computation on difficult ones. We study whether a model can learn to allocate its own reasoning effort by choosing, as the first token of its response, one of three modes: \textsc{NoThink} (answer as quickly as possible), \textsc{Short} (brief reasoning), or \textsc{Long} (extended reasoning). The choice is learned inside Group Relative Policy Optimization (GRPO) with no separate router, through a shaped reward that makes each mode worthwhile at a different response length, together with hard per-mode token caps that keep the modes distinct. On a 1.5B distilled model trained on MATH, the three modes emerge without collapsing to a single choice, and the brief modes end up more accurate than \textsc{Long}, which shows that the router sorts problems by difficulty rather than at random. Averaged over three seeds, the resulting policy stays close to the base model's accuracy on the held-out MATH500 ($0.782$ vs.\ $0.796$) while cutting the mean response length from $4{,}796$ to $2{,}811$ tokens (a $41\%$ reduction). Interestingly, it also transfers to other benchmarks without retraining, with the largest savings where problems are easier, with for instance 76\% token reduction on GSM8K and at higher accuracy than the baselines at similar response length. In short, we build a reasoning model that adaptively chooses how much to reason for each problem.
Poomphob Suwannapichat, Boonyarit Changaival, Caesar Wu +1cs.MA cs.CL cs.LG
LLM-based Multi-Agent Systems (MAS) achieve strong performance on complex reasoning tasks by coordinating multiple agents, but at the cost of substantial token consumption. Recent work on automatic topology design, ARG-Designer, has reframed this problem as autoregressive graph generation. However, its training objective provides no explicit incentive for the model to generate sparse and efficient topologies. We address this limitation by introducing a Reward-Guided Autoregressive Graph Generation (RGA-Designer) inspired by Reinforcement Learning from Human Feedback (RLHF). We train a reward model that jointly captures task correctness and structural compactness, and then fine-tune the pretrained graph generator using the reward model as feedback. Our method preserves task accuracy at the level of ARG-Designer while reducing token consumption by an average of 20.5%.
Long-horizon embodied tasks require LLM agents to iteratively decompose high-level goals, revise plans in response to environmental feedback, and ground leaf-level subgoals into valid executable actions. Recursive context-management methods such as ReCAP improve planning stability through multi-level task decomposition and parent-node refinement, but still repeatedly invoke the LLM at leaf nodes to ground atomic subtasks into exact valid actions. We refer to this final grounding step as last-mile grounding redundancy, which accumulates into substantial LLM-call and token overhead during long-horizon execution. To mitigate this issue, we propose HaReCAP (Habitual-action Grounded ReCAP), a low-intrusion leaf grounding extension for ReCAP. HaReCAP extracts frequent leaf decisions from successful trajectories and compiles them offline into auditable and abstainable one-step leaf-reflex rules. At runtime, it skips the leaf LLM call only when a rule can uniquely determine a legal action in the current valid-action set; otherwise, it falls back to the original ReCAP. This design avoids repeatedly carrying the full recursive context into the LLM for routine leaf action grounding, while preserving the original recursive control flow. We evaluate HaReCAP on Robotouille and ALFWorld with Qwen3.5-27B as the main model. On tasks solved by both ReCAP and HaReCAP, HaReCAP reduces token consumption by 14.67%, 17.93%, and 20.08% on Robotouille synchronous, Robotouille asynchronous, and ALFWorld, respectively. The results show that HaReCAP can serve as a low-intrusion extension to ReCAP-style recursive context-management frameworks, reducing last-mile grounding redundancy across environments and models on commonly successful trajectories.
The use of multimodal LLMs (MLLMs) for egocentric video understanding with wearable devices is constrained by the token budget. Memory and compute cost scale with the number of visual tokens, and high-resolution video quickly becomes expensive to transmit and process at scale. Prior work (GazeLLM) addresses this by cropping the video around the camera wearer's gaze. This reduces the number of visual tokens by about tenfold while maintaining or improving the quality of full-resolution descriptions. However, this compression strategy depends on dedicated eye-tracking hardware, which is unavailable on consumer smart glasses. Building a software-only substitute poses a joint constraint: the predictor must be accurate enough to preserve downstream description quality, yet light enough to run on-device, within the power and compute budget of a smartphone. We address this with EgoGazeLite, a lightweight dual-process gaze predictor for egocentric video. Across two MLLMs, three automated metrics, and two LLM judges, predicted-gaze crops show no significant difference from ground-truth-gaze crops. Equivalence is confirmed in all ten cases. EgoGazeLite achieves this at 15.7M parameters, 6.71 GFLOPs, and runs the full gaze-and-crop pipeline end-to-end in real time (21.6 ms/frame) on consumer accelerator hardware. Together, these results remove the need for eye-tracking hardware for token-efficient, gaze-conditioned egocentric video understanding with MLLMs.
Agent skills are often injected in full on every request, increasing token cost. We compare four content-preserving loading methods: Full, Skill Block, Reference, and Hybrid. Across SearchQA, SpreadsheetBench, ALFWorld, ScienceWorld, and SynthProc, we measure token usage using raw input for single-turn tasks and cache-correct effective input for multi-turn tasks. Results show no universal winner. Hybrid reduces input by 27.4% on SearchQA and 39.8% on SpreadsheetBench. On large multi-turn skills, Skill Block and Hybrid achieve substantial reductions, reaching 62.5% and 52.8% on ScienceWorld and 73.0% and 66.6% on SynthProc. ALFWorld shows smaller gains because procedures are short and repeatedly needed. Paired outcome tests detect no quality differences, though they do not establish equivalence. Overall, conditional loading is most beneficial when large portions of a skill are not needed on every turn.
Pretraining LLMs on artificial languages ("pre-pretraining") is a technique that could reportedly increase token efficiency by 33%, i.e., save up to 33% of training tokens needed to reach a certain performance. We validate this prior result for English on a larger set of natural languages across four language families, using two different tokenizers and varying model sizes. We also relate the observed gains (or losses) in token efficiency to quantified linguistic properties of the languages, such as sentence length, morphological richness, and features of dependency syntactic trees (tree depth, number of children, number of crossing dependencies). Our empirical results indicate that the reported gains depend heavily on the experiment setup and the choice of random seed, although we can confirm the trend of stable gains with 128-Dyck pretraining of small models with the Llama tokenizer for most of the examined languages. On a general note, we argue that multiple training runs should be carried out at least for a subset of experiments to avoid the community adopting unstable approaches.
Skill evolution improves agent skills through feedback over time, with failed trajectories often providing informative signals by revealing incomplete or misleading behaviors. However, existing methods mainly rely on single trajectories, where early reasoning errors can propagate through subsequent steps and weaken the feedback available for skill refinement. Consequently, improving skills requires repeated cycles of rollout, diagnosis, and update, incurring substantial token costs. To address this challenge, we introduce Branch2Skill, an efficient framework that transforms a single reasoning tree into dense supervision for skill evolution. For each task or problem, Branch2Skill performs Monte Carlo tree search under a fixed budget to obtain diverse reasoning trajectories, then compares an elite path with sibling alternatives sharing the same prefixes to extract step-wise evidence about which reasoning patterns to retain, revise, or avoid. Finally, Branch2Skill distills multi-step evidence into reusable updates, allowing one reasoning tree to provide supervision across multiple reasoning steps and reducing the need for repeated rollout-update cycles. Across six benchmarks covering reasoning and agentic tasks, Branch2Skill consistently improves task performance while enhancing skill evolution efficiency. For example, with GPT 5.5 as the target model, Branch2Skill uses 73.2% fewer tokens than SkillOpt, while achieving superior performance. These results demonstrate that reasoning trees can support not only more effective trajectory search, but also richer supervision for more efficient skill improvement. Code will be published.
Long-horizon research agents solve open-ended tasks through iterative retrieval, aggregation, and synthesis, but context grows rapidly while the marginal value of additional evidence often declines. This leads to unnecessary token cost, higher latency, and noisier inputs for final report generation. We study marginal value estimation for context management in deep research agents and present the first systematic stage-aware comparison of pruning strategies across the pipeline. We evaluate lightweight heuristic criteria and a learned value model at pre-retrieval, post-retrieval, and pre-synthesis stages. Our results show that pruning effectiveness depends more on where pruning is applied than on the specific scoring rule: early pruning yields the largest end-to-end savings, while later pruning mainly refines the final synthesis context. Lightweight heuristics reduce token usage by up to 73% with little quality degradation, learned pruning remains competitive on selected trade-offs, and no single method dominates across quality, efficiency, and faithfulness. These findings provide practical guidance for designing efficient long-horizon agentic systems.
Agentic coding faces growing problems of affordability and wasted tokens. We introduce Blast Radius, a predictive memory management layer that estimates an incoming prompt's reach through coupled context and code channels. NECROPHORESIS enables reversible eviction by archiving dead context verbatim, while Recurring Dead Matter (RDM) identifies and buries repeatedly occurring transcripts. We formulate reversible context eviction over a Polish context space, providing a measurable foundation for retention, recurrence, and eviction while connecting context entropy to resurrection probability. Across seven OpenAI models, Blast Radius reduced token consumption by 17-26%, achieved the lowest overflow rate among tested policies, and remained byte exact reversible. Of 450 buried bodies, 378 were recurring dead matter and zero were recalled. Blast Radius operates beneath HCRC, determining which records to bury and how far an incoming prompt may reach into the codebase. This work contributes to the broader goal of Algosophy: making large language models and agentic coding more reusable and sustainable.
Many methods for automated multi-agent system design optimize prompts and topologies during an initial design stage and then deploy the resulting system unchanged on subsequent samples. Experience from these samples is rarely consolidated into reusable system updates, while accuracy-oriented designs may incur high token costs. We introduce EMAS (Evolving Multi-Agent System), which uses this experience to revise MAS topology and prompts without updating LLM parameters, either to improve accuracy or to reduce cost. EMAS converts traces into structured diagnoses that specify a revision operation and target. It generates a candidate revision only when the same diagnosis recurs across samples and applies it only if paired validation against the current MAS meets the corresponding acceptance criterion. Across four benchmarks and two LLMs, EMAS attains the highest task-weighted overall accuracy for both backbones and is best or tied in six of eight model--benchmark settings. Within two evolution epochs, EMAS achieves relative gains of 6.30% and 20.10% in task-weighted accuracy on Kimi-K2-6 and Qwen3.6-27B, respectively. On MBPP with Qwen3.6-27B, EMAS raises accuracy from 55.09% to 89.12% while reducing token use per task by 62.2%. These results show that EMAS can turn experience from new samples into reusable updates to MAS topology and prompts.
Video large language models (Video-LLMs) have made strong progress in open-ended video understanding. However, their visual interfaces remain token-intensive and provide limited explicit structure for linking recurring object evidence across time. We introduce SlotNarrative, a slot-based interface that organizes a video into persistent object narratives represented by compact object-state tokens. Rather than compressing frame-wise features before establishing temporal correspondence, SlotNarrative first groups visual features into object-like slots and then associates recurring observations with clip-level object entries through a lightweight, parameter-free memory that integrates multiple complementary matching cues. Each retained entry is serialized into two token types: an identity token that summarizes persistent object appearance and a set of state tokens that encode segment-level appearance, geometry, visibility, and trajectory information. This design yields an interface of only 144 allocated visual-token positions for a frozen Video-LLM, independent of the number of sampled frames. Across multiple datasets, SlotNarrative achieves a favorable trade-off between accuracy and visual-token count compared with prior compact Video-LLM interfaces. Experimental results establish persistent object narratives as a compact, structured, and temporally organized visual interface for Video-LLMs. Our code will be made publicly available.
Arslan Battalov, Karim Kramin, Alexander Markotenko +1cs.LG
Muon is a recent optimizer that orthogonalizes the update to each weight matrix with a Newton-Schulz iteration, which performs steepest descent under the spectral norm. Almost all the evidence for it comes from Transformer models, and its behavior on state-space models is largely unreported. We compare Muon with AdamW on Mamba-2 130M under a controlled protocol that varies only which weight groups are trained with Muon. The benefit is localized. Muon on the output projection alone beats Muon on the input projection or on both. The advantage is mainly one of token efficiency. It holds on two corpora and two token budgets, and persists when training continues well past the compute-optimal point. Conditioning does not explain the gain. Muon lowers the condition number of whichever projection it trains, but the better-conditioned input projection is not the one that helps.
Chain-of-thought (CoT) reasoning can improve performance on difficult video questions but often wastes decoding tokens on simple ones. We study whether a video multimodal large language model can adapt its reasoning effort to each question. We propose AdaThinkV, an adaptive framework for video reasoning that learns whether to reason explicitly without offline difficulty labels, manually tuned confidence thresholds, or an external router. During reinforcement learning, AdaThinkV samples matched rollouts in explicit reasoning and direct answering modes for each prompt. ThinkGain estimates the prompt-level utility of explicit reasoning by balancing its accuracy gain against additional response length, providing supervision for both conditional response generation and autonomous mode selection. For difficult prompts, limited rollout exploration can yield groups in which every response is unsuccessful and accuracy rewards show little variation, providing insufficient signal for learning. We therefore introduce Variance Recovery Policy Optimization (VRPO), which retains and progressively expands these groups to recover informative signals from prompts that are difficult yet solvable. At inference, AdaThinkV selects a response mode and generates the response in a single autoregressive sequence. Across a unified suite of video reasoning evaluations, AdaThinkV achieves a mean accuracy of 40.79 with an average of 257.20 output tokens, outperforming the strongest evaluated adaptive baseline by 2.98 points while using 22.7% fewer tokens. Project page: https://trilarflagz.github.io/AdaThinkV/
LLM-based multi-agent systems (LLM-MAS) solve complex tasks through specialized collaboration, but inter-agent dependencies can propagate hallucinated or malicious outputs into system-level failures. Auditor agents mitigate these risks, yet existing strategies face an efficiency dilemma: end-only auditing reviews long trajectories and final outputs, potentially weakening audit effectiveness and enlarging rollback scope, while auditing every agent each round improves detection and localization at high token cost. How can guard performance be preserved while minimizing token cost? To address this problem, we propose BRA-Audit, a budget-aware runtime auditing framework that models MAS execution as a dynamic dependency graph and formulates audit scheduling as audit-point placement under a fixed audit-call budget to minimize cumulative unchecked exposure. Its greedy scheduler prioritizes influential and long-unaudited regions, while trusted audit points enable localized recovery. Across structured coordination, complex reasoning, and open-ended tasks, BRA-Audit restores performance close to the clean setting, remains competitive with heavy guard methods and reduces end-to-end token consumption by \(17.2\%\)--\(40.6\%\).
Wenxin Tang, Jingyu Xiao, Zhenyu Liu +6cs.CV cs.SE
Rendering source code as images offers a promising way to reduce the input costs of Multimodal Large Language Models (MLLMs). Adjusting image resolution can trade visual token cost against content fidelity. However, resolution scaling alone overlooks two sources of inefficiency: blank regions created by line breaks and indentation, and code regions irrelevant to the current instruction. Moreover, the best compression setting varies across inputs, tasks, and models, limiting fixed-ratio strategies. We propose CodeShrink, an adaptive visual compression framework with three components. Blank-Free Rendering replaces whitespace-dependent layouts with compact layouts and explicit structural markers, removing layout-induced tokens. Adaptive Compression Configuration uses a lightweight agent trained with reinforcement learning to predict a per-input setting that balances token efficiency and readability. Dominant Token Selection jointly analyzes the instruction and code image to prune task-irrelevant visual tokens during inference. We evaluate CodeShrink on code question answering, clone detection, and code completion. CodeShrink reduces visual token use by up to 71.2\% while matching or exceeding uncompressed text-only inputs, and consistently outperforms text-based and visual compression baselines across all three tasks. These results show that combining layout compaction, adaptive configuration, and instruction-aware pruning can make multimodal code understanding more efficient. Our code is available at https://github.com/vinsontang1/CodeShrink.
Extractive prompt compression promises to cut LLM inference costs by removing low-information tokens, and learned compressors such as LLMLingua-2 report strong results on English benchmarks. Most other languages already pay a token premium: the same content costs 1.3-1.8x more tokens than in English. We ask whether compression closes or widens this gap. Using fully parallel data in ten languages spanning five scripts, with controls budget-matched in the target model's tokenizer, we audit four learned compressors against four deterministic baselines, on eleven target models from ten vendors (over 250,000 evaluation calls). Three of the compressors are trained with English supervision (LLMLingua-2 XLM-R/mBERT; Kompress-v2 from the production Headroom stack); the fourth, XProvence, is trained multilingually. First, the transfer gap is real, replicates across target models and compressor backbones, and is strongly rate-dependent: at a 0.33 keep-rate English retains 57-62% of normalized context utilization while Lithuanian retains 10-24% and Chinese essentially none, despite Chinese having the smallest token premium. Second, the gap tracks compression supervision data, not architecture. All three English-trained compressors show it, deterministic methods show no comparable gap, and the multilingually trained XProvence v1 shows none. Its v2 release, retrained on translated data, empties 92% of Chinese contexts at its aggressive threshold without any warning. Third, in a harder long-context setting, aggressive learned compression drives compressed contexts to or below no-context utility in three of five non-English languages. A translate-then-compress pipeline matches or beats native compression at roughly half the token cost in three of five tested languages. We release all code, compressions, and model outputs. Safe compression budgets are much smaller outside English.
Zixuan Wu, Carolyn Jane Anderson, Arjun Guhacs.SE cs.CL
Although coding agents are now very effective in a variety of programming languages, this paper first shows that the cost (in tokens) can very significantly by programming language. We evaluate five recent models on programming problems in Python, Java, Rust, and OCaml. We carefully control for problem difficulty, and show that there can be stark variation in token consumption that is consistent across models. To understand why, we analyze both the structure and content of agent trajectories. First, we re-execute every intermediate solution and abstract each trajectory as a sequence of test-outcome vectors, then label the work between successive solutions. This reveals agents repeatedly producing noncompiling solutions in unfamiliar languages and revising solutions that already pass. Second, we analyze trajectory text, finding that agents plan solutions in code comments, distrust the provided tests in favor of inputs they invent, and sidestep unfamiliar target languages by prototyping in Python. Our results show that by-language token efficiency is a metric that should be considered when benchmarking and developing multilingual agents, and, for the tokenmaxxer, a guide to the most expensive language to work in.
Production AI agents' failures are less often due to an inability to reason well and more often because they cannot manage what is in their reasoning context: conversation histories, large prompts, large tool definitions, and ballooning tool outputs. Agents drown in their own accumulating history while paying a token cost that grows every turn, producing missing recalls within and across conversations. The incumbent response treats this as a storage-and-retrieval problem. We argue that framing is too narrow. Actively managing what an agent holds in mind is a lifecycle, not merely a store: it spans deciding what to remember, extracting and structuring it, choosing the right store per data type, consolidating and forgetting while preserving provenance, deciding what is relevant now, anticipating what is needed next, and compacting context to a budget without losing what matters. In serious production this operates not over a single user but across an organizational scope hierarchy. We name this discipline Agentic Context Management (ACM) and decompose it into five primitives: architecting, ingesting, scoping, anticipating, and compacting & consolidation. We then make the economic case: naive context accumulation grows token cost quadratically in conversation length, crude summarization buys linear cost at the price of an accuracy cliff, and only validated compaction achieves linear cost with preserved fidelity. We describe a reference implementation, Maximem Synap, that realizes the five primitives as a multi-tenant service and reports 92% on LongMemEval and 93.2% on LoCoMo under the configuration detailed in Section 6. We close with dimensions existing benchmarks do not yet capture, latency, token efficiency, and context-rot resistance, and the frontier of decision-level and organization-level context the category points toward.
Pruning long context for coding agents has been a vital technology for efficient context management. While existing context pruning methods such as SWE-Pruner realize this by attaching a separate code classifier, we find the agent itself encodes internal representations indicating the relevance of code context when reading tool output. Based on this finding, we propose SWE-Pruner Pro, which prunes tool outputs directly inside the agent. Concretely, a small head turns the agent's own internal representations into a keep-or-prune label for each line, with a length-aware embedding keyed to each tool output's line count. Across two open-weight backbones and four multi-turn benchmarks, SWE-Pruner Pro saves up to 39% of prompt and completion tokens while preserving task quality, with bounded inference overhead. Notably, on MiMo-V2-Flash SWE-Pruner Pro additionally raises the SWE-Bench Verified resolve rate by +3.8% and the long-context Oolong accuracy by +2.2 points.
Generated tokens are a direct driver of the cost, latency, and energy of generative AI (GAI) code editing. We show the format of feedback is a lever on all three. We compare two deliveries of the same requested changes: a holistic prompt (control) versus the structured, line-anchored export of FileMark (treatment). FileMark is a VSCodium extension for inline comments on any file. In a paired experiment line anchoring cut generated tokens by 22% (Claude Opus) and 58% (Claude Sonnet), reaching 24%-80% on files of 100 lines or more, with four of seven models generating significantly fewer tokens after multiple-testing correction. Correctness rose where models had headroom: +2.0 points pooled and +5 to +7 points for three of five local models. An exploratory experiment in which the harness, not the GAI model, applies function-level patches shows the correctness benefit grows further when the edit-application burden is lifted: local-model correctness on 100+ line files roughly triples under anchoring. Line-anchored feedback reduces what stronger models spend and improves what weaker models get right.