Tema

LLMs

2972documentos

222últimos 7 días

GAAC: Lossless KV Compression for Bandwidth-Efficient LLM Inference on Personal Devices

Weight-only, lossy quantization substantially reduces the storage and memory requirements for serving large language models (LLMs), but the growing key-value (KV) cache remains a bottleneck for long-context inference on personal devices. We present Group-Adaptive Additive Codec (GAAC), a lossless KV compression and serving system that exploits exponent redundancy to reduce storage and accelerate attention without introducing KV approximation. GAAC encodes BF16 exponents using calibrated per-channel base-plus-offset codes over contiguous ranges. GAAC adapts code widths across small groups while preserving KV values bit-for-bit. A fused attention kernel reconstructs KV tiles on the fly, avoiding full-cache decompression. Block-wise compression and fixed-capacity slots support incremental cache updates and paged serving. Across five tested models, GAAC achieves 1.405-1.407lossless BF16 KV compression. We also study compression performance on low-precision tensors FP8 E4M3 caches ( 1.09-1.14) and E5M2 caches (1.25-1.30). On NVIDIA DGX Spark, GAAC accelerates decode attention at a 64K-token context by 1.19-1.27and end-to-end W4A16 decode steps by 1.04-1.15. In serving, it reduces resident KV memory by approximately 26% and increases aggregate throughput by 13-29% under comparable KV-memory budgets. On Apple M4, GAAC achieves 1.02-1.25attention speedups, demonstrating portability beyond NVIDIA GPUs.

arXiv AI Papershace 16 días

Is it my preference? Role Grounding Attack against Agent's Memory via Social Interaction

LLM-based agents now share experiences on social platforms and remember new preferences to enhance their performance. However, existing memory systems contain a vulnerability in role-to-principal attribution that attackers can exploit. We identify Role Grounding Misattribution, in which an attacker-designed interaction causes the victim agent to incorrectly record a preference from the attacker as a preference of the victim's user. Specifically, we are the first to formalize Role Grounding Attack (RGA), an attack model that exploits this vulnerability through nested conversation structures. In the memory formation stage, an experience-sharing post quoting a conversation between the attacker agent and its user is published, which explicitly identifies its publisher. The inner conversation expresses preferences from the user of the attacker agent. When the victim agent remembers preferences from this nested structure, its memory loses the context binding and grounds the attacker's preference to the victim's user instead. In the memory retrieval stage, the misattributed preference persists and influences the victim agent's decisions when it encounters relevant scenarios. We instantiate RGA as tool selection manipulation and privacy leakage. Tool selection manipulation causes a tool-call preference to be misattributed to the victim's user, steering the victim toward an attacker-controlled tool. Privacy leakage extends this by misattributing both a tool-call preference and a task-requirement preference to the victim's user. The victim selects the attacker's tool, then includes sensitive information in tool arguments to satisfy the task requirement. We evaluate RGA across five mainstream memory systems and six language models. Tool manipulation and privacy leakage both achieve high attack success rates. These results call for stronger role grounding.

arXiv AI Papershace 16 días

Two Birds with One Stone: Skill-Augmented Generation via Unified Retrieval and Utilization

Skills, as reusable external knowledge, have demonstrated substantial potential to enhance the ability of Large Language Models (LLMs) to solve complex tasks. However, existing skill-augmented generation methods typically rely on separate models for skill retrieval and skill utilization, resulting in disconnected optimization of these two processes. To address this gap, we propose Skill-Uni, a unified reinforcement learning framework that jointly optimizes skill retrieval and utilization. Specifically, we adopt a generative skill retrieval paradigm that enables a single model to both learn how to retrieve relevant skills and leverage them for task solving. To effectively train this unified framework, we introduce a two-stage training strategy consisting of warm-up instruction tuning and GRPO, allowing the model to progressively improve its skill retrieval and utilization capabilities. Extensive experiments on five skill-related benchmarks demonstrate the effectiveness of Skill-Uni, yielding average performance gains of 66.59% and 34.82% over the base Qwen3-4B and LLaMA3.1-8B models, respectively. Besides, Skill-Uni is compatible with various skill retrieval engines, demonstrating its flexibility and generalizability. Additional in-depth analyses also provide insights into the effectiveness of Skill-Uni.

arXiv AI Papershace 16 días

From Natural Language to Real Schematics: An Automated Workflow for Power Electronics Schematic Synthesis

Designing power electronics converters traditionally requires intensive human expertise across topology sizing, commercial component selection, circuit simulation verification, and schematic drafting. While Large Language Models (LLMs) show promise in electronic design automation (EDA), applying standalone LLMs directly to board-level power electronics schematic synthesis suffers from severe engineering hallucinations, including inaccurate analytical sizing, fabricated or mismatched component part numbers, violated voltage derating margins, and broken topological connectivity. In this paper, we propose a Bounded Dual-Loop Automated Schematic Synthesis Framework that translates natural language specifications into verified, editable KiCad schematics and manufacturing deliverables. Our workflow integrates six stages across three functional columns: (1) requirement parsing and knowledge-driven power stage synthesis covering non-isolated (Buck, Boost) and isolated/resonant (Flyback, DAB, half-bridge LLC) topologies; (2) contract- and evidence-coordinated component selection grounded in distributor APIs and OEM datasheets, coupled with closed-loop ngspice transient verification and stress feedback; and (3) modular control hardware integration and multi-format deliverable generation featuring native KiCad netlist round-trip verification. End-to-end case studies on a 24-36V to 12V,2A Buck converter (29 positions, 83 verified terminals) and an isolated 48V to 24V, 0.5A, Dual-Active Bridge (DAB) converter (87 positions, 341 verified terminals), alongside comparative evaluations against a Vanilla LLM baseline, demonstrate that our framework improves component verification coverage from 22.2 to 98.6, nominal value consistency from 0 to 97.7, and SPICE simulation pass rates from 0 to 100.

arXiv AI Papershace 16 días

Component Translation and Judge-Labelled Refusal: A Multilingual Safety Evaluation

Whole-prompt translation evaluates a request in different languages but does not compare which component carries a language. We introduce a paired reversal: translate the complete request component or the complete operation component into the same target language, holding behaviour, English frame and translated-span count fixed. Component identity and non-English text quantity change together. Assigning English, Chinese and Spanish independently to the two components yields nine configurations on 190 of 300 sampled harmful behaviours retained after consensus annotation and automatic gating. Refusal scored by language-model judges is the primary outcome; unsafe compliance is reported separately. On Gemini 2.5 Flash and DeepSeek, three refusal swaps are about 5.3 percentage points and the fourth is -0.5; none survives the pre-specified six-test Holm correction. In an exploratory extension, Mistral Small 4 has a 13.7-point Spanish swap (95% interval [8.4, 18.9]), which survives the twelve-test Holm sensitivity correction under the primary judge, and 12.1 points under a second judge. GPT-5.4-mini estimates are within 2 points of zero. Component assignment supplies an additional evaluation dimension. Evaluations should record component languages and translated text quantity, comparing matched behaviours at equal translated-span count while interpreting refusal and unsafe compliance separately.

arXiv AI Papershace 16 días

DecompGym: Benchmarking LLM Agents on Designing Decomposition Algorithms

Benders decomposition and column generation solve large optimization problems by passing cuts, columns, and dual values between a master problem and subproblems, and are hard to implement correctly. Large language models now automate parts of algorithm design, which points toward agents that build decomposition solvers for new problems. As a first step toward such agents, we study how well current coding agents design and implement these two algorithms, and how they fail. We introduce DecompGym, 81 instances from nine problem families (five Benders, four column generation), in which an agent writes a solver for a prescribed method. Each solver must meet three criteria: execution within a 60-second time limit; numerical feasibility together with a method audit, in which an LLM judge reads the solver's code; and an objective within 1% of the stored optimal objective. We evaluate 12 GPT, Claude, and Gemini models running in coding-agent tools. Over up to ten attempts, we compare independent regeneration with reflection, which first diagnoses the failure. On the first attempt, two models solve every instance and seven solve more than 90%. Most Benders failures (48 of 62) are missing or suboptimal solutions. The most common column-generation failure is a method-audit rejection. Output checks alone would accept 93 of the 108 attempts that fail the method audit. The most common rejection, an optimalif, occurs 96 times, 92 of them in one model.Reflection turns 66% of retries into successes, compared with 39% for regeneration, and it repairs method-audit failures, for which the judge names the defecective gaps. Agents for decomposition therefore need checks on how a solver reaches its answer and feedback that locates the error. The benchmark is available at https://github.com/SIKAI-C/DecompGym.

arXiv AI Papershace 16 días

Decompose, Don't Expand: Capability-Independent AI Systems by Construction

A recurring design choice in building AI systems is whether, when a task is too hard for a model, to expand the model (make it larger, better-prompted, fine-tuned), or to decompose the task (extract the parts that admit a deterministic solution to a host, leaving the model only the irreducibly fuzzy residue). We make three claims. First, that many tasks commonly delegated whole to a model are in fact deterministic tasks with a small generative core. Second, and our central empirical result, that decomposing such a task yields a measurable and desirable property we call parity by construction: on the extracted sub-tasks the model-size performance curve becomes flat, because no model performs them at all. We measure this across nine models spanning a 494M-parameter local checkpoint to a frontier Opus-class model, a thirty-fold parameter range on the local family alone: once the host resolves the context, operation selection scores identically for all nine models and selector resolution for eight of nine, while the one task that does not decompose, labelling, remains sloped across the same range. Because the task leaves the model, a stock checkpoint suffices: no task-specific training and no fine-tuning. Third, that this is a general design discipline, not a domain trick: we show the same move recurring across physics setup, motion retargeting, and object-model editing, and in a Web-CLI tool family (same author, so not an independent replication). We give a method for locating the cut, the construction that makes a cut exist (a deterministic surface and a closed grammar over it), and a way to make the residue safe (resolve the deterministic context; verify the generative output), position the work against neuro-symbolic AI, tool-augmented language models, and generate-and-verify approaches, and are explicit about what is measured (one domain) versus illustrative (the rest).

arXiv AI Papershace 16 días

Answer-Only Disclosure: Private-Context LLM Assistance over Web APIs with a Leakage Ledger

Web assistants increasingly answer questions about a user's private documents by pairing a small on-device model with a cloud LLM behind a Web API. In collaborative decoding, the document stays on the client, and the cloud receives only a public view that the user approves, such as the question alone, but the two models exchange messages while the answer is generated. We show that the message transcript can disclose private content that neither the public view nor the answer contains. In a planted-secret audit, speculative decoding with private drafts did not include the secret code in its answers, yet sent the code to the cloud in rejected drafts. We present ADLER, a decoding protocol for streaming LLM APIs in which every message to the cloud depends only on the public view and the answer the user sees, regardless of how the client combines the two models or how the cloud behaves. A client-side leakage ledger then bounds how much the answer itself reveals relative to a public reference process built from the cloud's replies. In the audit, no ADLER configuration exposed the secret beyond the answer. On GSM8K math problems with numerical values masked in the cloud's input, ADLER with an unlimited leakage budget raises the exact-match accuracy of a 0.5B client from 0.41 to 0.83 and of a 1.5B client from 0.67 to 0.85 under lenient answer parsing, close to the 0.92 achieved by uploading the full problem. It also exceeds a zero-leakage program baseline (0.69) by 0.14 and 0.16, respectively. In this baseline, the cloud writes a program that the client runs offline using the private numbers. The ledger records the disclosure charge associated with this gain; budgets of 16 nats or less leave accuracy near the cloud-only baseline on GSM8K.

Hacker News (AI)hace 16 días

Survival LLM: AI assistant and emergency library that runs on your own computer

en▲ 11

arXiv AI Papershace 16 días

GEPA-Curator: Curated Learning for Self Improving Agents

Large Language Models (LLMs) are increasingly used in multi-turn agentic workloads that require sustained interaction with complex environments. Existing reflective prompt optimizers such as GEPA learn from failed rollouts, but agentic failures can also come from prompt weaknesses, harness limitations, stochastic execution, or noisy judges. We argue that such failures may not provide useful learning signal, and may even be harmful; and thus it is essential to curate which tasks are provided to the prompt optimizer for better performance. We introduce GEPA-Curator, a framework that augments the prompt optimizer to select which agent experiences should shape prompt optimization. In particular, GEPA-Curator consists of two components: task curation, which filters out tasks that are likely to provide noisy or unreliable learning signals, and prior success retrieval, which selectively retrieves past successful trajectories for a task and provides them as additional context to the prompt optimizer. We evaluate GEPA-Curator on GoBrowse, ALFWorld, AppWorld, and τ-bench, using both Gemini 3.1 Pro and Claude Opus 5 as optimizer models. GEPA-Curator matches or exceeds the mean performance of both baselines, with gains of up to 20.7% over GEPA and 21.2% over ACE.

arXiv AI Papershace 16 días

Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Scientific discovery often begins with an unexpected observation that prompts a new research question. In a world with altered physical laws, can large language models (LLMs) notice unexpected behavior, such as heavier objects falling more slowly, and discover the law through experiments without being told what to investigate? To evaluate this ability, we introduce LongChang, an automated framework that creates simulated worlds with altered physics, where one or more laws differ from ordinary physics. The evaluated LLMs act as explorers, manipulating objects and building tools to investigate their own questions without research tasks or hints about these laws. Separate LLM agents generate and review scenes, review tool designs, and evaluate reports for correctness and experimental support using the altered physical laws and exploration records. These records guide scene evolution for further exploration, allowing the full cycle to run without human intervention. Among 12 LLMs, GPT-6-Astra achieves a discovery rate of 16.7%, while the other 11 score 0%. The exploration records reveal uninvestigated anomalies and only partial identification of the altered physical laws.

arXiv AI Papershace 16 días

Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis

We explore how to build performant large language model (LLM) tool-use agents that complete complex, multi-step tasks. Specifically, we study how data composition in synthetically generated tool-use environments affects the training of such agents. To do so, we propose the SynthAPI pipeline, which explicitly controls the diversity of generated environments along the axes of tools, tasks, and the databases on which the tools operate. By systematically varying SynthAPI's generation controls in 458,752 tasks across 849 trained models of four different sizes, we derive a scaling law that predicts downstream tool-use agent generalization based on the scale and composition of the synthesized environments. These predictable relationships reveal several lessons for scaling tool-use agent data. As data budgets grow, scaling tool and database diversity becomes more effective than increasing the number of tasks, and increasing data complexity yields larger gains. Scaling according to these findings yields agents that outperform models trained on data from prior work in large-scale tool-use agent and synthetic environment generation across four evaluation benchmarks.

arXiv AI Papershace 16 días

MuonGS: Pretraining LLMs with Relaxed Weight-Norm Control

Weight decay and explicit weight-norm constraints improve optimizer performance in language model pretraining, yet a unified understanding of why such control helps and how much is needed remains incomplete. We analyze norm control in Muon over the entire training trajectory using the accumulated effective learning rate, $C_T=_{t

arXiv AI Papershace 16 días

Multi-State Linear Attention with Generalized Second-Order Routing

Self-attention plays a key role in long-sequence modeling, but its quadratic time complexity fundamentally limits the scaling of Large Language Models. Recently, Multi-State Linear Attention (MSLA) mechanisms have been proposed to make sequence modeling more efficient by splitting the context into segments and compressing each in its own memory state. MSLA aggregates the information extracted from each memory state using only the mean alignment between a query and each segment's keys, ignoring how the alignment is distributed within the segment. In order to model the relevance of a segment to a query more comprehensively and incorporate the variance of the alignment in the aggregation, we propose MERLIN (MomEnt Routing for multi-state LINear attention), a method that tracks the first and second moments of each segment's keys. Using the moments of keys, a query can recover both the mean and the variance of its alignment with each segment to perform a more content-aware aggregation of per-state readouts. Experiments on multi-state variants of Gated DeltaNet-1.3B and Mamba-2-780M show that MERLIN consistently outperforms existing aggregation methods on various long-context tasks. In particular, MERLIN outperforms baselines on RULER needle-in-a-haystack and long-context QA (+8.3% for Gated DeltaNet-1.3B and +7.0% for Mamba-2-780M) and diverse subtasks in LongBench (+13.5% and +12.0%). It even outperforms baselines that consume 4x more memory. MERLIN provides a simple and general routing mechanism for improving long-context capabilities of MSLA models.

arXiv AI Papershace 16 días

Metacognitive Decoding: Efficient Inference for Diffusion Language Agents

Diffusion language models offer a promising path to efficient agents through parallel token generation. However, efficient interaction requires adapting both where computation is spent and how much text is generated. We introduce a metacognitive controller that guides diffusion generation according to the needs of the current response and the agent's interaction with the environment. It learns when to finalize tokens and how far generation should extend, reducing overall computation within responses. Our controller, a lightweight plug-in for the diffusion backbone, is trained on agent trajectories to balance correctness and generation cost. This approach connects the agent's generation process to the interaction the agent serves. Experiments across five benchmarks demonstrate that our approach reduces inference FLOPs by 20.6% on average compared with the most efficient baseline, while maintaining competitive task accuracy. Ablation studies further validate the effectiveness of our approach.

arXiv AI Papershace 16 días

What Improves During Pretraining? Tracking Knowledge Acquisition with Point-in-Time Models

Language models pretrained on web corpora collected over multiple years learn both general and domain-specific knowledge. How these different kinds of knowledge are learned as pretraining progresses remains unclear. Recent point-in-time models trained on web snapshots from different years provide a way to probe how knowledge prevalence varies over time. Using these models, we propose a temporal variability score (TVS), defined as the coefficient of variation of an item's negative log-likelihood across models. A higher score suggests that the knowledge may be less consistently represented across years of web text. We evaluate six benchmarks covering diverse topics using six models from multiple open-source model families. We find that low- and high-TVS items are learned similarly early in training, but high-TVS items improve faster later on. At the end of the pretraining run, high-TVS items outperform low-TVS items in 34 of 36 model-benchmark settings, with an average advantage of 7.3 points of chance-normalized accuracy. This pattern is most robust on knowledge-intensive tasks, even when comparing items of similar difficulty. Our findings reveal differences in how knowledge is learned over the course of pretraining, with temporal variability providing a diagnostic of these learning dynamics beyond aggregate benchmark scores.

arXiv AI Papershace 16 días

Rare, Relevant, and Still Wrong: How Memory Fails LLM Group Decisions

When an LLM group's discussion is compressed into a short memory for a final decision maker, what should the memory keep? Two natural heuristics keep the rarest claims or the claims relevant to the decision criteria. On hidden-profile tasks both look attractive, because how widely a fact is held is confounded with whether it helps. We introduce CrossProfile, a diagnostic benchmark of 60 templates with automatically verified answer keys, in which the ownership of supporting, decoy, and irrelevant facts is crossed while their text is held fixed (480 tasks). Across Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models, both heuristics fail in predictable places. Selecting by relevance readmits decoys, because decoys are on topic: even a perfect relevance filter loses 23 to 31 points to rarest-first selection in the classic hidden-profile configuration. Rarest-first selection is no better than random selection on average and wins only where the rare facts happen to be the useful ones. In both cases accuracy follows the balance of evidence the memory keeps, not the policy that kept it. The decision model adds a third failure, a form of social proof: it gives extra weight to facts attributed to many agents. Removing the attribution tags, or reassigning them at random with line order fixed, raises accuracy in the classic configuration by 13 to 24 points. In natural discussions the evidence is lost before any memory acts: the complete transcript yields lower accuracy than no notes at all.

arXiv AI Papershace 16 días

VideoLVR: Latent Visual Reasoning over Grounded Video Evidence

Chain-of-thought reasoning has substantially advanced video understanding in multimodal large language models (MLLMs). However, language-based reasoning traces do not provide a reliable mechanism for models to revisit relevant visual evidence, leading to visual forgetting and impairing their ability to reason over video content. Existing approaches maintain visual grounding during reasoning in two main ways: they either refer to evidence using textual coordinates, which specify symbolic locations rather than the visual content itself, or reinsert cropped regions or frames into the multimodal context, preserving visual details at the cost of substantial context expansion. To address these limitations, we introduce VideoLVR, which constructs latent visual evidence from local regions at relevant video moments, thereby providing visual support for intermediate reasoning steps. To supervise latent visual evidence generation, we introduce a visual token prediction objective that formulates visual token reconstruction as prediction over a vocabulary derived from the input video, thereby unifying visual evidence reconstruction and text generation within the same token prediction framework. This shared prediction framework enables MLLMs to reason over interleaved text tokens and latent visual states while integrating evidence from different moments. We further extend VideoLVR to reinforcement learning by treating sampled visual tokens as actions alongside text tokens and jointly optimizing the policy using final-answer rewards. By efficiently revisiting task-relevant visual information during reasoning, VideoLVR improves video understanding without substantially expanding the multimodal context. Experiments on multiple video benchmarks demonstrate its effectiveness, especially on tasks requiring fine-grained visual evidence.

arXiv AI Papershace 16 días

Beyond Answer Changes: Matched Same-Class Controls for Local Audio Evidence

For audio-language models, an answer change after a local sound-event edit can reflect a response to the event's class, sensitivity to waveform rewriting, or both. We compare matched same-class and different-class replacements of a queried event. Both use the same interval, level estimate, donor fitting, and crossfade, while the same-class replacement preserves the correct answer. Across 600 synthetic soundscapes and three models from two architecture families, different-class replacements produce 49.7 to 60.9 percentage points more answer changes than same-class replacements. Same-class accuracy remains within 5.2 percentage points of original accuracy. Among trials with correct original and same-class answers, soundscape-weighted selective accuracy for the inserted class is 73.7 to 81.9%, averaging eligible trials within each soundscape and then weighting eligible soundscapes equally. This protocol tests how model answers respond to local changes in event class.

arXiv AI Papershace 16 días

TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed Information

Many real-world decisions require teams to integrate information distributed across participants. Yet AI agents are still primarily trained and benchmarked in isolation. We introduce TeamGym, an environment for evaluating and training language model-based agents on team decision-making problems constructed from real-world legal, finance, medical, and software engineering corpora. TeamGym introduces a method to identify a minimal set of facts, each element of which is necessary and all of which are collectively sufficient for a correct final decision, which we call pivotal information. This enables TeamGym to both control the distribution of pivotal information and track how well agents identify and integrate pivotal information when making team decisions. Examining 8 language models on 12 source corpora, we find that current AI agents often fail to effectively share information: Distributing pivotal information across participants lowers success rate by 22.9 percentage points on average relative to giving the same information to a single model (a communication gap). We also document that participants do not effectively retrieve decision-relevant information on their own: A team's success rate is lowered by a further 19.9 points when participants must identify and extract pivotal information from source material rather than receiving it directly (a retrieval gap). To help remedy these deficits, we use TeamGym to train policies through reinforcement learning on team decision loss. Across two policy-training experiments, we find that training improves final-decision quality and induces emergent coordination in participation, speaker selection, and communication, even though the reward evaluates only the final decision.