Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #8553

MuSiQue

Related event timeline, sources and context from the news index.

EVENT TIMELINE

4

RESEARCH · 1 SOURCE · arXiv cs.AI

Modeling LLM reasoning dynamics with T/E/S structural representations (arXiv:2611.23749v1)

This arXiv paper introduces T/E/S (text, local events, spans of activity) to represent temporal structure in language-model reasoning. Across reasoning and information-seeking tasks, adding events and spans improves next-outcome prediction in 70 of 76 evaluation groups for a Transformer and 74 of 76 for Qwen with LoRA; span duration and preceding events associate with transitions and retaining event history across spans helps prediction. The authors formalize budget-allocation counting and action-conditional continuation queries, prove counting hardness under explicit encodings, derive conditions for decision certification, and show in finite-process experiments that bounds which cancel shared continuation mass recover optimal actions while reducing mean inference CPU time by 77.7%. A compact structural observer also lowers Brier loss from 0.0836 to 0.0756 on 71 held-out HotpotQA and MuSiQue questions, but an annotation audit finds substantial errors in that observer, separating the computational findings from semantic validation of its features.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Answer configuration determines whether LLMs follow their reasoning trace or current evidence

The arXiv preprint (2610.20805v1) studies how keeping a model's reasoning trace at answer time interacts with verified evidence by comparing two answer configurations: continuation of the model's own thinking versus a separate answer call. Across gpt-oss, Gemma4 and Qwen3 on HotpotQA and MuSiQue, using a separate answer call raises evidence-following by 29.9–60.6 points versus continuation; removing the trace lowers accuracy similarly under both configurations, but continuation disproportionately repeats stale conclusions while a separate call follows current evidence far more often. The paper also finds that the final line of a trace carries most of its value, gold evidence helps most before the trace is written, distractors weaken correction under a separate call, and renewed reasoning (as opposed to a verification instruction) typically follows current evidence in conflicts, leading the authors to recommend reporting both trace benefit and evidence-following alongside answer configuration when evaluating traces.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Study shows rationales mainly affect verifier judgments, not answer accuracy

The paper introduces a message-intervention diagnostic that holds evidence and candidate answers constant while varying only the rationale passed from a reasoner to a verifier. On 400 examples across MuSiQue, HotpotQA and 2WikiMultiHopQA using DeepSeek as generator and verifier, faithful rationales add almost no answer accuracy versus no rationale, but corrupted rationales substantially change verifier support judgments (10–22% under a blind verifier prompt, 34–55% with explicit rationale-checking), while final answers change less (2–30%); human audits reveal instances of model overtrust.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Paper shows RLVR (with GRPO) works on Qwen3.5-0.8B without distillation

The arXiv preprint applies Reinforcement Learning with Verifiable Rewards (RLVR) using Group Relative Policy Optimization (GRPO) and an interleaved Wikipedia search tool to train Qwen3.5-0.8B on MuSiQue and a seven-benchmark QA suite. The best run reaches 0.352 average exact-match (vs 0.092 untrained), a 3.8× improvement with no distillation, and the authors find reward shape matters—sparse exact-match rewards perform worst for this model scale.

6.0