Imagine you're in a packed stadium trying to hear one friend shouting from the far side. The problem isn't really that your friend is far away — it's that ten thousand closer people are also talking, and your ears naturally weight nearby sound more heavily. That's the Proximity Trap this paper identifies in long-context LLMs. The model's attention budget gets consumed not by distance per se, but by the sheer mass of irrelevant context that happens to be closer to the query. The committed claim: existing long-context retrieval work has been solving the wrong problem. The field has focused on making models attend across longer distances, but LYRA argues the real bottleneck is cumulative interference from proximal background text. This is a reframing paper — it doesn't just propose a fix, it redefines the failure mode. The authors introduce a t-distributed directional matching mechanism that reshapes the retrieval distribution to have heavier tails, pushing attention mass toward genuinely relevant evidence regardless of position, while preserving the positional encoding structure the model has already learned. Architecturally, LYRA sits in the attention-redistribution family — not replacing the attention mechanism but reshaping its output distribution post-hoc. The t-distribution choice is deliberate: it has heavier tails than the Gaussian-like distributions attention typically produces, which means rare-but-relevant tokens in distant positions get more probability mass. This is analogous to robust statistics choosing t-distributions over Gaussians to resist outlier collapse. The method is applied as an alignment layer, not a fundamental architecture swap, which means it's model-agnostic in principle. The experimental ladder is solid but crowded. They evaluate on LongBench-v2, RULER, and LongBench — all established community benchmarks — and report consistent improvements across context lengths and task categories. The paper doesn't name a single dominant SOTA to beat; instead it frames improvements as consistent gains over baseline models (the same LLMs without LYRA). This is typical of the 'plug-in improvement' paper genre: the win is measured as delta-over-self rather than delta-over-best-in-class. The authors also introduce ProxBench, a new multi-level benchmark specifically designed to measure distant evidence utilization under increasing proximal interference — this is the most original contribution, as it operationalizes the Proximity Trap hypothesis into something measurable. Integrity is reasonable but not ironclad. The benchmarks used (LongBench-v2, RULER) are community-standard, which is good. ProxBench is new and self-authored, which means it hasn't been stress-tested by independent groups yet. The paper is 18 pages, suggesting reasonable experimental detail. Code and project page are provided. The core risk: the Proximity Trap framing is elegant but the experiments may not fully disentangle distance effects from proximal interference effects — separating these two correlated variables is genuinely hard. The milestone question is about scaling: LYRA works on current long-context benchmarks (up to ~128K tokens based on standard LongBench-v2 ranges), but the real test is whether the Proximity Trap gets worse super-linearly as context grows to 1M+ tokens, which is where the field is heading. If proximal interference scales worse than distance, LYRA's framing becomes more important over time, not less. The concrete next number: demonstrating gains at 500K-1M token contexts where the proximal noise budget is 5-10× larger. The obvious experiment not run: applying LYRA to retrieval-augmented generation (RAG) pipelines, where the proximal-vs-distant dynamic is even more explicit (retrieved chunks vs. conversation history). This is almost certainly being saved for a follow-up paper — the framing maps perfectly to RAG, and the authors would know this. The second missing piece is a head-to-head against other attention-redistribution methods (landmark attention, StreamingLLM's sink tokens) on ProxBench specifically.
