Search: “triton-inference-server”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference

BeaconKV introduces training-free KV cache compression using beacon queries, cutting long-reasoning inference memory up to 5.8x while preserving accuracy.

The paper shows recency-based KV cache compression assumptions fail in long-horizon reasoning because Thought Revisiting Tokens (TRT) re-attend to distant context such as early task-solving plans. TRT queries cluster into a small number of similarity groups, which BeaconKV exploits by maintaining compact beacon query representatives to anticipate revisited KV pairs without storing full query history. The training-free method achieves up to 5.8x memory reduction and over 4.3x throughput improvement across four open-source large reasoning models while nearly preserving full cache accuracy.

Hugging Face daily papers·Sep 3, 20:00 UTC · 13d agoAI research1

33

harshatheg/Qwen-2.5-1B-RLCD — new model trending #30 on Hugging Face

A community MLX inference engine evaluates constrained JSON schema fields in parallel on Apple Silicon, reporting 5.6-7.0x latency speedups with guaranteed schema validity.

The repository harshatheg/Qwen-2.5-1B-RLCD appeared at #30 on Hugging Face trending, but its content describes Parallel Constrained Decoding, an MLX-based inference engine for structured extraction and classification on Apple Silicon Macs. Benchmarked with mlx-community/Qwen2.5-1.5B-Instruct-4bit on an M4 Max, it reports 5.6x-7.0x latency reductions (e.g., 1,900 ms to 270 ms for a 28-field support triage task) with 100% syntactic validity and calibrated field-level probabilities. The engine prefills a single KV-cache, broadcasts it across all schema fields, and slices logits to valid candidate tokens for enum fields with up to 255 choices.

Hugging Face trending models·Sep 16, 01:35 UTC · 1d agoAI tools & infra

15

PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents

PARSER uses parallel reader subagents and an RL-trained lead agent for long-context QA, beating baselines and cutting latency up to 11x.

The PARSER paper decouples reading from reasoning: frozen subagents each read one document chunk in parallel while an RL-optimized lead agent iteratively broadcasts queries and aggregates evidence in scatter-gather rounds. On multi-hop QA with 7K to 896K token contexts, a 4B-backbone PARSER beats the strongest sequential memory baseline by 5.7 points on average and 12.0 points at 896K tokens, and a 9B version surpasses DeepSeek-V4-Pro by 6.3 points. Controlled experiments show robustness to evidence position, order, and distance perturbations, with inference latency reduced by up to 11x.

Hugging Face daily papers·Sep 5, 20:00 UTC · 11d agoAI research1

52

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

An 8.9B-parameter latent-space language model using next-concept prediction matches OLMo-3-7B pretraining loss with only 51.3% of the training tokens.

NCP-ArchPreview augments next-token prediction with Next Concept Prediction over a product-quantized concept vocabulary built from hidden states, trained jointly end-to-end. The 8.9B model was trained on 5.73T tokens from the Dolma-3 dataset, the largest latent-space language model demonstration to date. It consumes 51.3% of the tokens to reach OLMo-3-7B's final pretraining loss and outperforms it by 2.45 points on the downstream macro-average, including a 5.99-point GSM8K gain. The learned latent space also enables lightweight domain adaptation via a 17M-parameter VQ module and improves speculative drafting accepted length by 4.17%.

Hugging Face daily papers·Sep 8, 20:00 UTC · 8d agoAI research1

52

RetroThinker: Enabling Retrospective Thinking in Speech LLMs

RetroThinker is a post-training framework letting the Moshi speech LLM self-correct reasoning mid-stream, adding 11% GSM8K accuracy at similar latency.

Researchers introduce RetroThinker, a multi-stage post-training framework that equips the Moshi speech LLM to verify and forward-correct chain-of-thought steps during streaming inference. It combines supervised fine-tuning on curated retrospective thinking data with length-based direct preference optimization (DPO). On GSM8K it achieves an 11% absolute accuracy gain over non-retrospective baselines at comparable latency.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:41 UTC · 6d agoAI research1

28

Large Language Models Develop Belief State Geometry In-Context

Probing six open-source LLMs on HMM-generated data shows belief states are linearly decodable (R² 0.83–0.99), suggesting in-context learning approximates Bayesian prediction.

Researchers prompted six open-source LLMs with data from 40 hidden Markov models selected for non-trivial belief structure and probed residual-stream activations for belief states (posteriors over hidden states). Belief states were linearly decodable with peak R² values of 0.83–0.99 across HMM/LLM combinations, spanning early to late layers. Patching and steering the probe-identified subspace preserved downstream prediction quality while control interventions degraded performance substantially, establishing functional relevance. The results provide representation-level evidence that in-context learning approximates optimal Bayesian prediction over a context-inferred generative model.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:17 UTC · 1d agoAI research1

45

Long-Lived Characters, Local Inference: Incremental Memory Maintenance for Game NPCs

Researchers present incremental KV-cache memory maintenance for long-lived game NPCs running locally on a quantized Qwen hybrid model.

The paper studies incremental memory maintenance for long-lived game NPCs deployed locally with a quantized Qwen hybrid recurrent-attention language model. The runtime removes superseded attention KV entries, computes replacement records at the true sequence tail, and preserves the continuing recurrent state and unchanged KV. Experiments across eight scripted maintenance rounds show true-tail updates preserve current-state and historical bindings, while slot-preserving alternatives repeat a double-subtraction error.

arXiv cs.AI / cs.LG / cs.CL·Sep 16, 17:09 UTC · 14h agoAI research

25

The Self-Expanding Stolen Inference Supply Chain: An AI Agent Harvesting and Re-Serving LLM Access, (Fri, Sep 11th)

An autonomous coding agent harvested LLM API access from poorly secured gateways and aggregated stolen inference capacity behind a self-hosted gateway

A SANS researcher observed a semi-autonomous coding agent finding weakly secured LLM resale gateways via FOFA queries, creating trial accounts with temporary emails and CAPTCHA solving, and exploiting weak authorization such as client-supplied group_id fields. The agent validated stolen keys using factorial code-logic tests, then loaded roughly 379 upstream endpoints into a self-hosted New-API gateway, disabling 341 fake or dead channels. Five model names including claude-opus-5 and gpt-5.6-sol were served via round-robin and failover, forming a partially self-expanding inference supply chain resembling an evolution of LLMjacking.

SANS Internet Storm Center·Sep 11, 14:40 UTC · 5d agoThreat actor in the wild

68

TPMSpy: Validation of Measured Boot Systems by Low-Level Tracing of TPM Usage

Researchers present TPMSpy, a platform-agnostic method validating TPM Measured Boot via low-level tracing, finding inconsistent Linux systemd measurements that break remote attestation and LUKS decryption.

An arXiv paper (2609.05011) introduces TPMSpy, a method that analyzes virtualized system–TPM interactions to independently reconstruct and validate TPM Event Logs without relying on the quoting mechanism, applicable to open and closed systems and demonstrated on Linux and Windows. A longitudinal analysis of Linux systems running systemd versions 245–258 (2020–2025) found wide divergence in Measured Boot usage, undocumented behavioral changes, and no common usage pattern. The authors report inconsistent measurement of user-space systemd services, which prevents reliable remote attestation and LUKS disk decryption on affected systems.

arXiv cs.CR·Sep 4, 11:23 UTC · 12d agoResearch

40

TokenRhythm/NeoHorse-1-4B — new model trending #30 on Hugging Face

TokenRhythm releases NeoHorse-1-4B, an Apache-2.0 agentic fine-tune of Qwen3.5-4B claiming +5.93 benchmark macro-average gain.

NeoHorse-1-4B is a roughly 4B-parameter text-only causal language model post-trained by TokenRhythm from Qwen/Qwen3.5-4B for agent harnesses, tool use, coding, and instruction following. It applies routing-guided curriculum SFT and routing-guided on-policy distillation over execution trajectories as an early prototype toward recursive self-improvement (RSI). The release reports a 64.87 macro average across ten benchmarks versus 58.94 for Qwen3.5-4B (+5.93) and is distributed under Apache-2.0, trending #30 on Hugging Face.

Hugging Face trending models·Sep 5, 08:21 UTC · 11d agoModel release1

28

Agora: Git as Shared Memory for Collective AutoResearch

Agora records multi-agent research as an append-only Git DAG; 13 LLM workers ran nearly 12 days on a weight-transfer problem.

Agora stores every result, hypothesis, and verification as an immutable commit in a Git-stored DAG, with a derived index exposing the frontier and verification status of claims. In a nearly 12-day run, 13 language-model workers with no assigned tasks or central planner published 1,703 contributions on initializing a frozen 119.6M-parameter attention-SSM hybrid from 141 donor models. They improved the evaluator from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M, with 165 independent reproductions posted and none failing.

Hugging Face daily papers·Sep 15, 20:00 UTC · 1d agoAI research

45

rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inference

rMuscle, a caching-based inference framework for vision-language-action models, achieves 1.29-1.42x speedups on RTX 4090 and Jetson Thor while preserving success rates.

rMuscle is a real-time inference framework for Vision-Language-Action (VLA) models that exploits cross-execution similarity in repetitive robot tasks via a dual-phase muscle-memory cache. The Context Cache reuses visual-token outputs to reduce computation, while the Action Cache reuses neuron activation patterns to reduce weight accesses, with online recomputation and sliding-window retrieval keeping overhead low. It achieves 1.29-1.42x speedups on RTX 4090 and Jetson Thor across LIBERO, RoboTwin, and real-world manipulation tasks while maintaining original success rates.

arXiv cs.AI / cs.LG / cs.CL·Sep 16, 17:34 UTC · 14h agoAI tools & infra

32

Maverick: Private and Verifiable LLM Inference Made Practical via Matrix-Vector Multiplication Delegation

Maverick protocol delivers private and verifiable LLM inference via matrix-vector multiplication delegation, achieving up to 45x throughput gains over local inference on Qwen3-4B.

Maverick introduces an information-theoretically sound protocol for delegating matrix-vector multiplication with transparent preprocessing, efficient batch verification, and virtually no server overhead, combined with LPN-based pseudorandom masking for input privacy. It addresses privacy and correctness concerns when users delegate open-weight LLM inference to third-party providers. An end-to-end prototype evaluated on Qwen3-4B achieved throughput gains over local inference of up to 45x with precomputed privacy masks and 44x for verification-only workloads, with a CPU server using up to 128 threads.

arXiv cs.CR·Sep 9, 14:53 UTC · 7d agoResearch1

40

Echo: Learning-based Matching Decompilation using Trusted Back Translation

Researchers introduce Echo, a matching decompiler using trusted back-translation that roughly doubles exact-match rates and outperforms GPT-5.6 and Codex on Mirai.

Echo performs matching decompilation by using compilation as trusted feedback for iterative search: a domain-specific model generates candidate code and compilation configurations, which are recompiled, compared at assembly level, and repaired via rule-based rewriting, neural refinement, and reasoning-based refinement. On function-level benchmarks, Echo produces 2.43x more exact matches than the strongest baseline and the highest structural similarity to ground truth. On the Mirai malware binary, it matches 2.75x and 7.4x as many functions as GPT-5.6 and Codex, respectively.

arXiv cs.CR·Sep 16, 14:14 UTC · 17h agoResearch

25

Domain-Specific Hallucination Detection in Large Language Models

A multi-signal pipeline detects LLM hallucinations, reaching F1 0.915 on HaluEval and cutting Qwen2.5-0.5B hallucination rates from 85.5% to 37.7% via DPO.

The paper presents a hallucination detection pipeline combining fine-tuned DeBERTa-v3 classification, Monte Carlo Dropout uncertainty, and temperature-scaled calibration. It achieves F1 0.915 and AUROC 0.977 on general-domain HaluEval tasks, with MC Dropout inference raising accuracy to 93.2%. Applying DPO to a Qwen2.5-0.5B generator reduces its hallucination rate from 85.5% to 37.7%, while cross-domain evaluation shows poor general-domain transfer to SciFact (F1 0.52) and PubMedBERT fine-tuning as the strongest adaptation (F1 0.63).

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:45 UTC · 6d agoAI research1

32

SpecGuard: Inference-Time Backdoor Detection For Free

SpecGuard detects backdoored LLM behavior at inference time using speculative decoding acceptance rates, adding no extra model computation.

Researchers propose SpecGuard, an inference-time backdoor detector that repurposes draft-token acceptance rates from speculative decoding as a detection signal at zero added model-computation cost. When a trigger shifts the target model toward attacker-controlled behavior, the clean draft model's acceptance rate changes, exposing the backdoor; the paper formalizes when this signal appears and shows suppressing it weakens the backdoor. Experiments across diverse backdoor types and model families show reliable detection, including stealthy cases invisible to input-level filters. Speculative decoding is positioned as a free, always-on monitor for frequently updated deployed models.

arXiv cs.CR·Sep 10, 16:51 UTC · 6d agoAI safety & security 2 sources2

45

CodeTD: Topology of Attention Detects Hallucinations in Code LLMs

CodeTD detects hallucinations in code LLMs before execution by analyzing topological patterns of attention maps, outperforming recent baselines.

CodeTD applies topological data analysis (TDA) to code LLM attention maps to quantify prompt-generation mismatch as a pre-execution correctness signal. Experiments cover HumanEval, MBPP, BigCodeBench, and MultiPL-E across 5 programming languages and 10 code LLMs up to 34B parameters. The method outperforms recent baselines and transfers between coding benchmarks, helping catch code that fails the task or embeds security vulnerabilities.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:26 UTC · 9d agoAI research1

24

Atlas: Efficient Verifiable Semantic Search

Atlas delivers zero-knowledge proofs for HNSW semantic search, verifying RAG retrieval in under a second on SIFT1M and 2.0 seconds at 100M vectors.

Atlas lets a search provider prove that a query was answered correctly against a committed HNSW index without revealing the index, addressing provider deviations like truncation or bias. It combines offline preprocessing, a fixed-size-state restructuring of HNSW with a correctness proof, and timestep-tagged batching of per-step arguments. The system proves queries in under a second on SIFT1M and 2.0 seconds at 100 million vectors while preserving plaintext HNSW recall, and proven retrieval maintains end-to-end RAG answer quality at lower cost than prior verifiable retrieval systems.

arXiv cs.CR·Sep 10, 17:30 UTC · 6d agoResearch4

40

A Three-Layer Caching Architecture for Low-Latency LLM Web Search on Commodity CPU Hardware

OreoLook's three-layer Redis caching architecture cuts redundant LLM calls and embedding work for CPU-hosted web-search answer generation.

The paper describes a three-layer caching architecture for OreoLook (formerly lixSearch), an open-source LLM answer engine: a Redis session context window with Huffman-compressed disk overflow, a semantic query cache matching rephrasings via embedding cosine similarity, and a URL embedding cache deduplicating embedding computations. Deployed on a single 8-vCPU Intel Cascade Lake server with 30 Hypercorn workers across three containerized replicas, it achieved an 89.3% aggregate Redis keyspace hit rate, 0.1 ms read latency, and 1.38 MB memory overhead. An LRU eviction daemon migrates idle sessions to disk and rehydrates them for resumption hours or days later.

Hugging Face daily papers·Aug 11, 20:00 UTC · Aug 11, 2026AI tools & infra2

34

Shuffling is Not Enough: Breaking Permutation-Based Model Confidentiality in Hybrid FHE Inference

Attack breaks permutation-based model confidentiality in hybrid FHE inference, recovering all ResNet-20 linear layers exactly with d+1 queries per layer.

The paper shows output-permutation plus noise fails to protect model confidentiality in hybrid FHE inference: d+1 admissible queries recover an exact permutation-invariant summary of a d-input linear layer, and shuffle-model DP amplification premises cannot hold under correctness-bounded noise. The authors recovered all linear layers of a Safhire-style ResNet-20 end-to-end from TFHE transcripts with zero error, using 5,712 total queries. Exact per-layer recovery was also confirmed on pretrained ImageNet-scale CNNs and ViT-B/16. Leaked layer spectra enable model fingerprinting, lineage attribution, and improved logit-based extraction, while suppressing them destroys inference utility.

arXiv cs.CR·Sep 11, 14:37 UTC · 5d agoResearch1

45

Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments

Researchers extend the SwiftSage dual-process agent with adaptive memory and self-reflection modules, improving scores in interactive environments.

The work adds an Adaptive Memory Module (AMM) for salience-gated episodic storage and trigger-driven retrieval, and a Self-Reflection Module (SRM) for bounded execution-time validation and corrective intervention, to the SwiftSage agent. Controlled ablations on ScienceWorld across four configurations show the full system achieves the best mean final score (64.62), success rate (43.17%), and successful-step efficiency (19.33 steps). SRM is the strongest standalone contributor, suggesting execution-time control is the dominant bottleneck while episodic memory helps once the runtime loop is stable.

arXiv cs.AI / cs.LG / cs.CL·Sep 16, 17:50 UTC · 14h agoAI research

25

TasmScan: Continuation-Aware Taint Analysis for TVM Bytecode with Savelist Abstraction

TasmScan introduces source-free taint analysis for TON smart-contract bytecode, detecting 95.3% of defects with 96.8% precision and 17x speedup.

TasmScan is the first bytecode-level static analysis framework for the TON Virtual Machine, enabling cross-continuation data flow reasoning without source code by modeling savelist semantics through forward register analysis with formal over-approximation guarantees. It lifts bytecode into a typed intermediate representation (TASIR) and performs path-sensitive taint analysis. On a 208-contract benchmark with human-confirmed ground truth it detects 95.3% of defects across five classes at 96.8% precision, and resolves 294,546 dynamic continuation targets with 100% precision across 2,921 registry contracts. It achieves a 17x median speedup over symbolic-execution baselines.

arXiv cs.CR·Sep 15, 11:01 UTC · 1d agoResearch1

28

CoRA-NAS: Coarse Ranking and Anchor-Residual Refinement for Neural Architecture Search

CoRA-NAS combines zero-cost proxy ranking with low-cost learning-curve refinement, achieving the best worst-space Spearman correlation across NAS benchmarks.

The paper proposes CoRA-NAS, a two-stage neural architecture search framework pairing a static ranking prior (CoRA-Rank) with learning-curve refinement (CoRA-Refine) that extrapolates early validation curves for sampled anchors and propagates residual corrections with an ExtraTrees model at about 1% of full training cost. Across NAS-Bench-201, NAS-Bench-101, TransNAS-Bench-101, and NATS-SSS it achieves mean Spearman correlations of 0.946, 0.715, 0.786, and 0.894 respectively, with the best worst-space correlation of 0.715 among compared methods. On NAS-Bench-201/CIFAR-100 its selected architecture reaches 73.32% accuracy versus a 73.37% ground-truth best.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:49 UTC · 6d agoAI research

28

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

A*-Thought-V2 compresses redundant chain-of-thought steps into latent tokens guided by hidden-state geometry, improving accuracy up to 2.6% while halving response length.

A*-Thought-V2 models chain-of-thought as a hidden-state trajectory projected into a 3D PCA space and compresses steps whose transitions deviate from the question-to-solution direction into continuous latent tokens, keeping aligned steps explicit. Training uses stepwise embedding forcing and label forcing with soft multi-modal vocabulary supervision. On Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks it improves average accuracy by up to 2.6%, cuts response length by up to half, and raises Accuracy per Computation Unit 2.29x while reducing preprocessing and training time by 94.6% and up to 80.3%.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:56 UTC · 9d agoAI research

36

Nous Research Adds One-Click Local Model Setup to Hermes Desktop

Nous Research's Hermes Desktop now offers one-click local model setup that reads hardware, picks a fitting quantization, downloads weights, and configures llama.cpp automatically.

Hermes Desktop, the MIT-licensed build of the open-source Hermes Agent, now sets up local models in one click: it reads the machine's hardware, selects a model that fits, downloads weights, and configures the inference runtime. It manages a pinned llama.cpp build with CUDA, Metal, Vulkan, HIP, and CPU backends, and shows green/amber/red memory-fit verdicts per catalog model before download. Quantization floors at 4-bit, and recommended models guarantee at least a 64K context window protected by ordered RAM offload (expert weights first, never the attention cache). It runs on macOS 12+, Windows 10/11, and Linux with no account required for local models.

MarkTechPost·Sep 5, 19:12 UTC · 11d agoAI tools & infra1

25

Drift-Constrained Optimization: Only Direction Matters in Fine-Tuning Instruct Models

Drift-Constrained Optimization reformulates fine-tuning as update-direction selection, letting Qwen3 models improve target tasks within a behavioral drift budget.

The paper specifies a behavioral drift budget before optimization and shows that update direction is the remaining degree of freedom, reformulating fine-tuning as a direction-selection problem. In a stringent QA-only setting where instruct models must still generate multi-step reasoning at inference, a coarse layer-selective probe reverses the failure of QA-only fine-tuning. Across Qwen3-8B and Qwen3-14B, these directions substantially improve scientific reasoning and multilingual translation, matching or outperforming dedicated translation systems over 100+ languages and giving stronger initialization for reinforcement learning.

Hugging Face daily papers·Sep 11, 20:00 UTC · 5d agoAI research

48

Register Tokens for Bounded-State Reasoning in Diffusion Language Models

Register tokens let diffusion language models like LLaDA and Dream carry reasoning state across cleared chunks, gaining up to 19.5 points on code.

Researchers propose register tokens: dedicated fixed-position tokens whose continuous hidden states are trained to carry reasoning progress across generation chunks in masked diffusion language models. After decoding and clearing a chunk, the model continues from the prompt and the carried register state instead of retaining earlier text. On LLaDA and Dream, registers outperform discrete-text carry on every benchmark, with gains up to 8.5 points on math and 19.5 points on code. Registers are especially effective for bounded code generation and can be further refined with reinforcement learning on long-horizon reasoning tasks.

Hugging Face daily papers·Sep 13, 20:00 UTC · 3d agoAI research

38

Architecting memory and storage in the AI era

Analysis argues AI inference shifts data-center bottlenecks to memory and storage, urging balanced compute, memory, storage, and network architecture over raw compute.

MIT Technology Review, citing Tirias Research principal analyst Jim McGregor, argues that AI inference and agentic workloads make data movement the key constraint, elevating memory and storage from background hardware to strategic assets. The piece says RAG and real-time inference require continuous data retrieval and caching that legacy infrastructure cannot support. It frames infrastructure planning as a business decision balancing performance, efficiency, cost, and scalability in healthcare, finance, and customer-facing AI.

MIT Technology Review · AI·Sep 4, 18:39 UTC · 12d agoAI industry

14

Beyond Training: A Feasibility Taxonomy for Inference-Time AI Governance

A study maps twenty inference-time AI governance mechanisms, finding commercial readiness only against cooperative deployers and no adequate defense versus state-level adversaries.

The paper develops a feasibility taxonomy of twenty inference-time AI governance mechanisms across monitoring, verification, and enforcement, each rated on a four-point readiness scale against a four-vendor evidence base. Fifteen of the twenty mechanisms have commercial technical substrates in production today, though governance-grade assurance and adversarial robustness vary substantially. Stress testing shows readiness holds only against a cooperative deployer and low-to-medium-capability user: no mechanism rates adequate against a high-capability state-level deployer, and fine-tuning removes model-internal enforcement components. A second-rater reliability check on readiness ratings returned a quadratic-weighted Cohen's kappa of 0.74.

arXiv cs.CR·Sep 9, 12:37 UTC · 7d agoAI policy

28

How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus

Reproduction study finds Orthrus speculative-decoding trajectories match the reference model in only ~45% of cases under BF16, but 100% under FP32.

Researchers independently reproduced Orthrus, a hybrid autoregressive-diffusion architecture claiming lossless speculative decoding via intra-model consensus, testing exact trajectory matching on 1,190 prompts across 12 domains. Under BF16, exact matching occurred in only 45% of cases for the authors' checkpoint and 43% for an independently trained model, with matching probability strongly tied to reference-model response-conditional perplexity. Despite trajectory divergence, downstream lm-eval-harness benchmarks showed no systematic degradation, while FP32 evaluation yielded exact matching on all prompts.

Hugging Face daily papers·Sep 13, 20:00 UTC · 3d agoAI research1

25