Groundy · Models & Research

Can You Trust LLM Confidence Scores? Verbalized vs Logprob Signals

Berry Mingus — Wed, 09 Sep 2026 04:05:47 GMT

A preprint reports that verbalized LLM confidence weakly tracks logprob signals, with instruction-tuned models showing worse calibration. Gate on verifiable internal signals.

Berry Mingus · Models & Research · 13 min read

Read the full article on Groundy →

Generative vs Cross-Encoder LLM Rerankers: Can Single-Pass Decoding Close the Gap?

Berry Mingus — Tue, 08 Sep 2026 08:09:56 GMT

SPD claims 28 ms listwise LLM reranking via single-pass decoding, but the evidence is one unreplicated preprint. Keep cross-encoders in production until independent benchmarks

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

mdlARC's 44% ARC-AGI-1 Claim: What the Small Budget Leaves Out

Berry Mingus — Mon, 07 Sep 2026 08:37:22 GMT

An author-reported 44% on ARC-AGI-1 public eval for about $0.67 of rented compute shows what small budgets can justify, and what still needs replication.

Berry Mingus · Models & Research · 7 min read

Read the full article on Groundy →

FP8 vs MXFP4 vs BF16: Why Your Quantized LLM Disagrees Across GPUs

Berry Mingus — Mon, 07 Sep 2026 07:15:12 GMT

FP8 and MXFP4 are umbrella specs, not single formats. A new preprint offers bit-exact conformance vectors to test quantized LLM portability across GPUs, exposing hidden format

Berry Mingus · Models & Research · 13 min read

Read the full article on Groundy →

Running a 104GB LLM on a 48GB Mac: What Expert Streaming Costs

Berry Mingus — Sun, 06 Sep 2026 06:56:19 GMT

Expert streaming lets 48GB Macs run 104GB MoE models by paging weights from SSD. This guide compares streaming to quantization and offload, analyzing workload fit and hardware

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

GPT-6 Astra on ARC-AGI-3: What the Agentic Score Actually Measures

Berry Mingus — Sat, 05 Sep 2026 19:22:46 GMT

GPT-6 Astra's ARC-AGI-3 scores vary by 97 points based on effort settings. Route agents by cost per solved task, not launch-day leaderboards.

Berry Mingus · Models & Research · 14 min read

Read the full article on Groundy →

Deepfake KYC Fraud: Tamper-Resilient Watermarks That Recover the Original Face

Berry Mingus — Mon, 31 Aug 2026 15:47:10 GMT

VeriFi preprint claims watermarks recover original faces after tampering, not just flag fakes. Robustness is untested against commercial tools and compression, so treat as a R

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Detecting AI-Generated Audio: Why Decay Tails Betray Voice Clones

Berry Mingus — Mon, 31 Aug 2026 04:14:43 GMT

A new preprint shows AI audio leaks in decay tails via group delay. It offers a cheap, watermark-free filter for fraud screening, though hold-out accuracy is only 66.7%.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Reward Hacking Starts in the Verifier: Rule Checks vs LLM Judges for Math RL

Berry Mingus — Sun, 30 Aug 2026 16:01:22 GMT

Rule checkers miss format variants while LLM judges get hacked during RL. This failure map from arXiv 2505.22203 shows why hybrid verifiers and reward audits are now essential

Berry Mingus · Models & Research · 15 min read

Read the full article on Groundy →

Can LLMs Train on Their Own Problems? What Zero-Data Self-Play Changes

Berry Mingus — Sun, 30 Aug 2026 03:19:29 GMT

J-Zero claims zero-data self-play beats baselines by 8.0 points on unverifiable tasks, but judge verdict flips of 5.3 to 48.4 percent suggest the gains may be self-validated.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Why LLM Agent Benchmarks Move When the Harness Changes

Berry Mingus — Sat, 29 Aug 2026 15:09:10 GMT

A new preprint argues agent benchmark scores track the evaluation harness as much as the model. Teams should pin harness versions and ablate scaffold changes to avoid misat.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Can You Serve LLMs on 2-Bit Weights? What Ultra-Low-Bit Quantization Costs

Berry Mingus — Fri, 28 Aug 2026 16:09:46 GMT

arXiv 2508.06753 reports 2-bit LLM serving with up to 7x speedups on Intel Xe2. For self-hosters, the binding cost is not memory but task-specific eval coverage to verify QAT-

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Do LLMs Still Need BPE? What RL-Trained Tokenizers Change

Berry Mingus — Thu, 27 Aug 2026 15:26:53 GMT

An ICML 2026 paper shows RL can learn token boundaries end-to-end, beating straight-through baselines at 100M parameters. For serving or fine-tuning, keep your frozen BPE or S

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

GLM-5.3-Flash vs Qwen3.8-Flash-Next: Which Budget LLM to Route To

Berry Mingus — Thu, 27 Aug 2026 14:27:49 GMT

GLM-5.3-Flash and Qwen3.8-Flash-Next lack confirmed pricing and independent evals. Hold your router, verify per-token costs, and run a 50-case domain holdout before switching.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Why Prompt Caching Can Change Model Outputs: A Prefix Invariance Audit

Berry Mingus — Wed, 26 Aug 2026 13:34:03 GMT

A new arXiv audit shows attention masks miss causality leaks in state-space models. Run a two-pass check before trusting prefix caches on hybrid stacks.

Berry Mingus · Models & Research · 15 min read

Read the full article on Groundy →

GDPR Deletion Requests vs LLM Weights: What Machine Unlearning Actually Removes

Berry Mingus — Tue, 25 Aug 2026 13:56:47 GMT

GDPR erasure demands hit LLM weights, but retraining is the only defensible fix. Behavior-level unlearning passes probes without proving data is gone, leaving compliance teams

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Why Drift Monitors Confuse Covariate Shift With Concept Drift

Berry Mingus — Tue, 25 Aug 2026 02:06:11 GMT

A new preprint proposes CJSD, a two-discriminator test that separates covariate shift from concept drift. This distinction determines whether to retrain or recalibrate, fixing

Berry Mingus · Models & Research · 13 min read

Read the full article on Groundy →

AI-Generated Apps Look Right, but Do They Actually Work?

Berry Mingus — Mon, 24 Aug 2026 14:09:46 GMT

MobileForge shows AI apps compile but fail navigation. Shift review from screenshots to project-level state tests to ship reliable generated frontends.

Berry Mingus · Models & Research · 14 min read

Read the full article on Groundy →

Prompt Injection in 3D Scenes: The Attack Surface Multimodal Agents Ignore

Berry Mingus — Sun, 23 Aug 2026 04:43:40 GMT

arXiv 2602.07104 shows 3D object placement injects instructions into multimodal agents, bypassing text and image filters. Teams must gate modalities and scope actions to.

Berry Mingus · Models & Research · 15 min read

Read the full article on Groundy →

Why LLM Log Anomaly Detection Pages You for Nothing

Berry Mingus — Sat, 22 Aug 2026 05:10:14 GMT

arXiv 2608.17965 shows LLM log detectors are overconfident in wrong verdicts. Gate paging on calibrated confidence, not raw F1, to prevent alert fatigue and ignored critical.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

DeepSeek v4 Flash Vision: Routing Images Without Verified Pricing

Berry Mingus — Sat, 22 Aug 2026 04:32:14 GMT

DeepSeek v4-flash-vision-exp lacks verified pricing and benchmarks. This guide prices Qwen-VL alternatives and outlines a fallback protocol for experimental vision endpoints.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

DeepSeek 32B on RTX 3090: Tokens per Second by Quant and Context

Berry Mingus — Fri, 21 Aug 2026 14:45:48 GMT

Bandwidth math predicts 32-39 tok/s for Q4_K_M DeepSeek 32B on an RTX 3090 at 8k context. 32k context overflows 24GB VRAM. No measured benchmarks exist, so every figure is.

Berry Mingus · Models & Research · 14 min read

Read the full article on Groundy →

PTXBench: LLMs Can Port GPU Kernels, But Not Beat Tuned Libraries

Berry Mingus — Fri, 21 Aug 2026 14:13:25 GMT

PTXBench shows LLMs can generate architecture-specific PTX for H100 and B200, but no model matches frontier libraries. Use it to cut hot-loop porting costs, not to replace.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Can LLMs Reuse Another Model's KV Cache? What Cross-Model Transfer Shows

Berry Mingus — Thu, 20 Aug 2026 13:44:30 GMT

A new preprint proposes a lightweight reader to reuse KV caches across models, but evidence suggests caches remain model-bound. Learn when recompute beats adaptation.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

DeCRIM: Decompose Constraints to Stop Silent Drops in Agent Outputs

Berry Mingus — Sat, 01 Aug 2026 23:23:57 GMT

DeCRIM shows that decomposing multi-constraint instructions into individually checkable units reduces silent drops by 7-8% on benchmarks, shifting reliability work from.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Operator-Level Triage for Silent Mixed-Precision Instability

Berry Mingus — Fri, 31 Jul 2026 17:38:42 GMT

arXiv 2607.25494 introduces a single-pass tool to localize silent numerical instability in bf16 and fp16 training. Treat it as a triage layer for unexplained loss spikes, not.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

BeyondUncertainty: Weak Confidence Signal for RAG Routing, Not Calibration

Berry Mingus — Fri, 31 Jul 2026 17:14:36 GMT

arXiv:2607.25600 shows verbalized confidence is a weak but real routing signal for RAG. It saves 20.4% retrieval calls for 28.2% token overhead. The probe is poorly.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Kimi K3 on M1 Max: Bandwidth, Not Capacity, Limits Local MoE Inference

Berry Mingus — Thu, 30 Jul 2026 08:31:44 GMT

Running Kimi K3 on an M1 Max proves local MoE feasibility via expert offloading, but unified memory bandwidth caps throughput. Treat this as a prototyping tool, not a serving.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Kimi Linear Cuts KV Cache 75% but Recall Remains the Binding Constraint

Berry Mingus — Wed, 29 Jul 2026 07:59:50 GMT

Kimi Linear cuts KV cache 75% and boosts decode 6x for 1M-token contexts, but recall stays the binding constraint. Route memory-bound jobs only after benchmarking in-context.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Why Chat Leaderboards Do Not Predict Image Quality

Berry Mingus — Tue, 28 Jul 2026 06:44:08 GMT

A July 22 visual test comparing GPT-5.6, Claude, Gemini, and Grok highlights a structural gap: chat models lack dedicated image backends. Teams must route by subtask using.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Kimi K3 Procurement: Governance Review Over Phantom Government Assessments

Berry Mingus — Mon, 27 Jul 2026 18:37:58 GMT

Moonshot AI's Kimi K3 release triggers governance review for regulated teams. Beijing jurisdiction, Anthropic accusations, and missing government assessment require internal.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

DeepSeek Compute Leak: Why Open-Weight Routing Needs a Swap Path

Berry Mingus — Sun, 26 Jul 2026 18:47:42 GMT

A leaked transcript claims DeepSeek faces a compute gap, but V4 shipped in April. Teams should treat this as a resilience prompt to pin Qwen as a swap-in spine, not panic.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Context Ordering Beats Window Size for Long-Context Agents

Berry Mingus — Sun, 26 Jul 2026 13:05:53 GMT

ARBIGRAPH shows tool agents lose 33.3% accuracy on dependent chains despite fitting context. Context ordering, truncation, and caching now beat raw window size for agent.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Diffusion LLMs: Training Cost, Not Parallel Decoding, Drives Deployment

Berry Mingus — Fri, 24 Jul 2026 18:52:28 GMT

Parallel decoding has not cut serving costs for diffusion LLMs. arXiv:2605.13026 closes training gaps by 4x, but KV-cache absence keeps inference slower than autoregressive.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Open-Weight Routers vs Fable 5: The Routing Math That Actually Matters

Berry Mingus — Fri, 24 Jul 2026 06:41:38 GMT

Fable 5 pricing and caching set the bar for open-weight routers. The Echo claim lacks verification. Routing wins only for high-volume, cache-unfriendly routine traffic.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

DeepSeek-V4 1M Context vs RAG: Why Retrieval Stays

Berry Mingus — Thu, 23 Jul 2026 17:52:03 GMT

DeepSeek-V4 markets a 1M token window for agents, but no independent benchmarks verify depth recall. CEO-Bench and ProGraph research show structured memory outperforms raw.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Qwen-Image-3.0 Does Not Exist: Why Self-Hosting Image Models Is Premature

Berry Mingus — Thu, 23 Jul 2026 06:07:11 GMT

No primary source confirms Qwen-Image-3.0 exists. Independent benchmarks show open-weight models fail 85% of precise image tasks. Self-hosting branded templates remains risky.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Kimi K3: 2.8T Parameters, MoE Routing, and Self-Hosting Reality

Berry Mingus — Mon, 20 Jul 2026 18:19:20 GMT

Kimi K3 pairs 2.8T parameters with 16-of-896 expert routing and a 1M context window. Hosted API is live, but full weights arrive July 27, 2026. Self-hosting requires.

Berry Mingus · Models & Research · 12 min read

Read the full article on Groundy →

Kimi K3 vs Qwen3.8 Max: Routing Strategy for July 2026

Berry Mingus — Mon, 20 Jul 2026 18:07:18 GMT

Kimi K3 offers concrete pricing for production bake-offs while Qwen3.8 Max remains a shadow evaluation. Compare both against Fable 5, GPT-5.6 Sol, and GLM-5.2 by workload.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Qwen3.8 Max Release Audit: API, Open Weights, and the License Catch

Berry Mingus — Mon, 20 Jul 2026 17:45:04 GMT

Qwen3.8 Max now has a stable API, $2/$6 pricing, 1M context, and downloadable 2.4T weights. The release is real, but the API and open checkpoint are not the same product.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

HuggingFace 100x Inference: Generalizable vs Platform-Locked Optimizations

Berry Mingus — Sun, 19 Jul 2026 18:50:32 GMT

HuggingFace claims 100x inference speedup, but independent benchmarks show its TGI engine trails vLLM by 24x on throughput. This article decomposes the claim to separate.

Berry Mingus · Models & Research · 13 min read

Read the full article on Groundy →

Kimi K3 Code Arena Rank: Self-Hosting Cost Math for Coding Agents

Berry Mingus — Sun, 19 Jul 2026 17:13:36 GMT

Kimi K3 is frontier-class, but its 2.8T MXFP4 checkpoint needs 64-plus accelerators. The real deployment math reveals what an arena rank cannot.

Berry Mingus · Models & Research · 11 min read

Read the full article on Groundy →

Can Tool-Adaptive LLM Rerankers Improve RAG Without Always Calling Tools?

Berry Mingus — Tue, 14 Jul 2026 22:48:18 GMT

TALRanker folds the tool-call decision into the reranker's scoring policy, turning tool latency from a fixed per-query tax into a budget the model spends only when uncertain.

Berry Mingus · Models & Research · 8 min read

Read the full article on Groundy →

Does Speculative Decoding with Progressive Tree Drafting Cut LLM Latency?

Berry Mingus — Tue, 14 Jul 2026 21:32:35 GMT

Progressive Tree Drafting grows draft tokens as a pruned tree to claim a 2× speedup, but the gain hinges on verifier acceptance and shrinks on code and reasoning.

Berry Mingus · Models & Research · 8 min read

Read the full article on Groundy →

Analytic Inference Cuts Bayesian Deep Ensemble Serving Cost, But Leaves Training as the Bottleneck

Berry Mingus — Fri, 10 Jul 2026 22:05:25 GMT

A new arXiv preprint replaces sampling-based averaging in Bayesian deep ensembles with closed-form Bayesian aggregation, cutting per-query inference cost and shifting the.

Berry Mingus · Models & Research · 8 min read

Read the full article on Groundy →

Tree-of-Thoughts Improves Text-to-Image Prompting by Reasoning Over Hypotheses, Not Pixels

Berry Mingus — Fri, 10 Jul 2026 18:55:21 GMT

A new arXiv paper ports Tree-of-Thoughts prompting to text-to-image in-context learning and reports CoBSAT gains, but reasoning runs on prompt hypotheses, not image states.

Berry Mingus · Models & Research · 8 min read

Read the full article on Groundy →

FourierQK's spectral Q/K filter cuts TinyShakespeare loss by 79%, but long-context proof is missing

Berry Mingus — Fri, 10 Jul 2026 16:19:03 GMT

FourierQK filters query and key projections before attention, cutting TinyShakespeare character-level loss by 79%, but word-level, retrieval and long-context tests are absent.

Berry Mingus · Models & Research · 7 min read

Read the full article on Groundy →

Tencent Hunyuan 3's Agent Push Has No Public DeepSeek or Qwen Benchmarks Yet

Berry Mingus — Fri, 10 Jul 2026 14:28:52 GMT

Tencent's Hy3 is billed as a 295B agent model with a 21B active token footprint, but public materials omit benchmark tables and independent DeepSeek or Qwen comparisons.

Berry Mingus · Models & Research · 6 min read

Read the full article on Groundy →

When Does Memory, Not Compute, Decide Who Can Profitably Serve LLMs?

Berry Mingus — Fri, 10 Jul 2026 11:38:30 GMT

A July 2026 arXiv paper argues that scarce HBM and DRAM bandwidth, not raw compute, will determine which labs and providers can profitably serve large language models through.

Berry Mingus · Models & Research · 9 min read

Read the full article on Groundy →

Can We Trust LLM Logic? A Graph-Based Stress Test Finds Three Failure Modes

Berry Mingus — Fri, 10 Jul 2026 10:18:06 GMT

A July 2026 arXiv paper shows Self-Consistency voting can hide contradictory reasoning, and GraphEVAL's graph-based coherence metrics catch flawed paths output checks miss.

Berry Mingus · Models & Research · 7 min read

Read the full article on Groundy →