Strategy/projects/files/sdr/sdr_manual_eval_2026_03_08.md
+

sdr_manual_eval_2026_03_08

SDR Manual Evaluation — 2026-03-08

Run date: 2026-03-08 12:20 MSK
Model: llama-3.1-8b-instant (Groq free tier)
Config: .env.benchmark (reduced token limits, SearXNG local search)
Queries run: 3/10 (sci_001, sci_002, sci_003)
Evaluator: Feanor agent (manual scoring)

Evaluation Rubric (1–5 scale)

Criterion Description
Coverage Are all major relevant methods/concepts mentioned?
Accuracy Are facts, numbers, and attributions correct?
Coherence Is the report logically structured and readable?
Depth Are mechanisms explained? Math/algorithms present?
Citations Are real, verifiable sources provided?
Completeness Does the report fully answer the question?

Results

sci_001: “State-of-the-art methods for operator splitting in ML optimization”

Report size: 5641 chars

Criterion Score Notes
Coverage 1/5 Missed ALL real methods: ADMM, ISTA/FISTA, Douglas-Rachford, proximal gradient, PDHG. Instead invented “ADO”, “PDOF”, “PDV”, “DAOE” — none of these exist.
Accuracy 1/5 Complete hallucination. All 4 “methods” are fabricated. Comparison tables are meaningless.
Coherence 2/5 Structural flow is logical (intro → overview → comparison → recs), but hollow.
Depth 1/5 No math, no convergence rates, no proximal operator definition, no connection to ML tasks.
Citations 1/5 “[Один]”, “[Два]”, “[Три]”, “[Четыре]” — Russian words for numbers, not real citations.
Completeness 1/5 Critically incomplete. A domain expert would reject this immediately.
Total 7/30 23%

Critical failure: Model hallucinated entire methods instead of recalling real literature. Likely caused by poor web search quality + insufficient model knowledge (8b parameter model on specialized domain).


sci_002: “How does FlashAttention improve transformer training efficiency?”

Report size: 2988 chars

Criterion Score Notes
Coverage 2/5 Mentions FlashAttention exists and is faster. Misses: IO-aware computation, SRAM blocking, O(N²)→O(N) memory, FlashAttention 2/3 versions.
Accuracy 2/5 “Уменьшение точности на 2%” — factually wrong. FlashAttention is mathematically equivalent to standard attention (same output). “4x faster” — no benchmark source.
Coherence 3/5 Best structured of the three. Readable.
Depth 1/5 Doesn’t explain WHY it’s faster (memory bandwidth bound → SRAM tiling). Core mechanism missing.
Citations 1/5 “[1] FlashAttention paper”, “[2] Paper о влиянии…” — placeholder text, not real DOI/arxiv refs.
Completeness 2/5 Superficially touches the topic, key technical insight absent.
Total 11/30 37%

Key error: “reduces accuracy by 2%” is a fabrication — FlashAttention is numerically equivalent to standard attention up to floating-point precision. This is a dangerous hallucination.


sci_003: “Current evidence on neural operators for solving PDEs”

Report size: 1914 chars

Criterion Score Notes
Coverage 2/5 Calls them “нейтральные операторы” (neutral operators?) — likely translation artifact or confusion. No mention of FNO, DeepONet, WNO, U-NO, PINN.
Accuracy 1/5 Vague comparative table with “высокий/низкий” — no actual results, no benchmarks (Navier-Stokes, Darcy flow, Burgers equation).
Coherence 2/5 Short (1914 chars) but somewhat structured.
Depth 1/5 No description of what neural operators actually are, how FNO works, operator learning paradigm.
Citations 1/5 No citations at all.
Completeness 1/5 1914 characters is catastrophically short for “current evidence” on an active research area.
Total 8/30 27%

Critical issue: Report is ~4x shorter than sci_001, which itself was insufficient. A proper literature review of neural operators would need 20-50K chars minimum.


Summary Statistics

Query Coverage Accuracy Coherence Depth Citations Completeness Total %
sci_001 (Operator Splitting) 1 1 2 1 1 1 7 23%
sci_002 (FlashAttention) 2 2 3 1 1 2 11 37%
sci_003 (Neural Operators PDE) 2 1 2 1 1 1 8 27%
Average 1.67 1.33 2.33 1.0 1.0 1.33 8.7 29%

Key Findings

What Failed

  1. Hallucination of methods (sci_001): Model invented 4 non-existent methods instead of real ADMM/ISTA/FISTA. Root cause: llama-3.1-8b doesn’t have reliable knowledge of specialized ML optimization literature.

  2. No real citations (all): All 3 reports provide placeholder-style references. The web search is not being used to find actual papers with DOI/arxiv IDs.

  3. Report size too small: sci_003 = 1914 chars; sci_002 = 2988 chars. Compare to OpenAI Deep Research (typically 10K-50K chars per report).

  4. Depth = 0 across all: No mathematical formulations, no convergence bounds, no algorithmic descriptions. The model is pattern-matching on domain vocabulary without substance.

  5. Factual errors (sci_002): “FlashAttention reduces accuracy by 2%” is factually incorrect and dangerous.

Root Causes

  • Model too small: llama-3.1-8b is insufficient for scientific synthesis. Even llama-3.1-70b struggles with specialized topics.
  • Search not working: Reports don’t cite real sources from web search. Either search is failing or model is ignoring search results.
  • Token limits too restrictive: RESEARCH_MAX_TOKENS=4000, FINAL_REPORT_MAX_TOKENS=6000 may be forcing truncation of retrieved content.
  • No citation extraction: System has no mechanism to extract and verify DOI/arxiv identifiers from retrieved pages.

Recommendations for SDR Improvement

Priority Action Expected Impact
P1 Use claude-haiku-4-5 as research model (instead of llama-3.1-8b) Major quality improvement
P1 Debug web search — verify results are being used in synthesis Fix citation failure
P2 Increase FINAL_REPORT_MAX_TOKENS to 20000+ Longer, more complete reports
P2 Add citation extraction module (extract arxiv/DOI from retrieved pages) Fix citations criterion
P3 Add minimum report length check (reject < 5000 chars) Completeness floor
P3 Add self-consistency check: model lists methods → verify against retrieved content Reduce hallucinations

Comparison Target

For context, this is what SOTA looks like on DeepResearch Bench:
- Gemini 2.5 Pro: 48.88/100 on PhD-level questions, 111 citations/task
- SDR 8b-instant: ~29% on our 6-criterion rubric, ~0-4 citations/task

Conclusion: Current SDR with llama-3.1-8b-instant is not viable for scientific research synthesis. Switching to claude-haiku or fixing the search pipeline should be the first step before any further benchmarking.

Choose icon