Qwen2.5-7B-Instruct

2026-10-28

RESEARCH · 1 SOURCE · arXiv cs.AI

When Retrieval Is Not Enough: obsolete retrieved facts can degrade agent accuracy

arXiv:2610.36308v1 analyzes the single-hop FactConsolidation task in MemoryAgentBench and shows that retrieving the newest statement is not sufficient for correct answers. With Qwen2.5-7B-Instruct on a 6K-history setup, BM25 accuracy falls from 45% to 32% when the retrieved evidence budget increases from 64 to 512 tokens despite a gold audit finding the newest statement in each retrieved set; substituting an unrelated statement with an obsolete competing statement drops accuracy from 100% to 44.6% for Qwen2.5 and from 100% to 87.8% for Qwen3-14B (all 50 harmful paired flips were classified as stale), and a 32K-history replication shows BM25-64 outperforming BM25-512 (69% vs 58%).

7.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

Study finds PRM-pruned fragment grafting (PPFG) produces no measurable benefit on tested reasoning LMs

The paper isolates PRM-Pruned Fragment Grafting (PPFG) as a minimal inference-time intervention and tests it on Qwen2.5-7B-Instruct with Math-Shepherd across MATH500 (and replicates across three base LMs, six benchmarks, a second PRM, and gate sweeps). PPFG — in both stagnation- and random-targeting variants — was statistically indistinguishable from an independent parallel-CoT baseline, and analyses attribute inertness to mis-targeted firing (only ~14% of injections hit genuinely struggling chains) and plateau states a graft cannot rescue; the authors also provide an equivalence-testing template for null results.

6.0

2026-09-24

RESEARCH · 1 SOURCE · arXiv cs.AI

Minimum Risk Training boosts Qwen2.5-7B-Instruct for ODIN outage-to-CIM IEC 61968-3 XML generation

An arXiv preprint applies Minimum Risk Training (MRT) to convert heterogeneous power outage reports from the Outage Data Initiative Nationwide (ODIN) into standardized XML compliant with CIM IEC 61968-3. Applying MRT to Qwen2.5-7B-Instruct raised overall accuracy on this structured-generation task from 16.20% to 68.95%.

6.0

2026-09-17

RESEARCH · 1 SOURCE · arXiv cs.AI

Inference Engineering Pareto Atlas (arXiv:2609.17863v1) maps cost/quality/latency trade-offs

The paper releases a Pareto atlas that maps cost, quality, and latency trade-offs for LLM inference by measuring 54 configurations of Qwen2.5-7B-Instruct on vLLM 0.12 across L4, A100, and H100 GPUs and calibrating a simulator that reproduces anchored measurements with <1.5% drift. Key findings: 18 of 36 calibrated configurations reach the Pareto frontier; combinations of optimizations outperform many single methods; AWQ 4bit cuts per-token latency to 0.34× but loses ~5.9% GSM8K accuracy; FP8 weights retain ~99.4% accuracy at 0.61–0.65× latency; a naive FP8 KV cache preserves throughput but answers none of 200 GSM8K questions correctly; H100 is best for tight latency while A100 is best for throughput and cost (~$0.106 per million tokens).

7.0